<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Ofer Fryman</title>
    <description>The latest articles on DEV Community by Ofer Fryman (@ofer_fryman_69213ebc646a7).</description>
    <link>https://dev.to/ofer_fryman_69213ebc646a7</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4245990%2F4733470f-6ca1-4a9a-9e8b-f6de7abcfe39.png</url>
      <title>DEV Community: Ofer Fryman</title>
      <link>https://dev.to/ofer_fryman_69213ebc646a7</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/ofer_fryman_69213ebc646a7"/>
    <language>en</language>
    <item>
      <title>How to Build an Offline, On-Device Voice-to-Voice AI Agent in React Native</title>
      <dc:creator>Ofer Fryman</dc:creator>
      <pubDate>Sun, 15 Nov 2026 14:33:10 +0000</pubDate>
      <link>https://dev.to/ofer_fryman_69213ebc646a7/how-to-build-an-offline-on-device-voice-to-voice-ai-agent-in-react-native-2a69</link>
      <guid>https://dev.to/ofer_fryman_69213ebc646a7/how-to-build-an-offline-on-device-voice-to-voice-ai-agent-in-react-native-2a69</guid>
      <description>&lt;p&gt;We’ve spent the past year helping teams like Volkswagen, Verbali and LunaFit ship voice agents that run entirely on the phone. This is what we learned building a full voice-to-voice agent in React Native, from the wake word to the spoken reply.&lt;/p&gt;

&lt;p&gt;I’m the founder of DaVoice, and I drafted this guide with AI help, which fits the subject. Our example app ships how-to files (&lt;a href="https://github.com/frymanofer/ReactNative_DaVoice/blob/main/example/englishHowTo.md" rel="noopener noreferrer"&gt;&lt;code&gt;englishHowTo.md&lt;/code&gt;&lt;/a&gt; and &lt;a href="https://github.com/frymanofer/ReactNative_DaVoice/blob/main/docs/AI_AGENT_INTEGRATION_GUIDE.md" rel="noopener noreferrer"&gt;&lt;code&gt;AI_AGENT_INTEGRATION_GUIDE.md&lt;/code&gt;&lt;/a&gt;) written for AI coding agents. Hand them to Claude Code, Cursor or Copilot along with your codebase, and the agent can add the wake word, speech-to-text, local LLM and on-device voice to your app step by step, in the order that works. I checked every detail here against the repo.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Short answer:&lt;/strong&gt; To build an on-device Voice-to-Voice AI Agent in React Native, chain five stages that all run on the phone: a &lt;strong&gt;wake word&lt;/strong&gt; detector, optional &lt;strong&gt;speaker verification&lt;/strong&gt;, &lt;strong&gt;speech-to-text (STT)&lt;/strong&gt;, a &lt;strong&gt;small local LLM that calls your app's tools&lt;/strong&gt;, and &lt;strong&gt;on-device text-to-speech (TTS)&lt;/strong&gt;. Stream the LLM's reply into TTS sentence by sentence, and hand the microphone carefully between stages. This tutorial does exactly that with the open-source &lt;a href="https://github.com/frymanofer/ReactNative_DaVoice" rel="noopener noreferrer"&gt;DaVoice React Native example app&lt;/a&gt; and three npm packages from the &lt;a href="https://davoice.io/voice-agent-sdk" rel="noopener noreferrer"&gt;DaVoice Voice AI Agent SDK&lt;/a&gt;: &lt;code&gt;react-native-wakeword&lt;/code&gt;, &lt;code&gt;react-native-davoice&lt;/code&gt; and &lt;code&gt;react-native-davoice-llm&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;The result is an agent that listens for "Hey Coach", understands "open the rear right window", runs a real function in your app, and answers out loud: "I opened the rear right window." No cloud LLM, no per-token bill, and no audio leaving the device during the voice loop.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;What "token-free" means here.&lt;/strong&gt; A local LLM still processes tokens. "Token-free" means you pay &lt;strong&gt;no cloud LLM inference token fees&lt;/strong&gt;, because inference runs on the user's phone. DaVoice SDKs are licensed separately; evaluation and non-commercial proofs of concept are free, and production use needs a commercial license.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h2&gt;
  
  
  Used in production
&lt;/h2&gt;

&lt;p&gt;The same on-device voice stack already runs in shipped products:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Verbali (MaTalk AI, an AAC app for non-verbal children)&lt;/strong&gt; moved its wake word, speech-to-text and text-to-speech fully on-device with DaVoice, so no audio leaves the tablet and voice keeps working without internet. In &lt;a href="https://www.verbali.io/news#davoice-on-device" rel="noopener noreferrer"&gt;Verbali's own write-up&lt;/a&gt;, DaVoice reached 97.65% wake word accuracy with zero false positives observed on their test setup, versus about 75% for the other engines they evaluated. In their words: "We have no affiliation with DaVoice beyond being a happy customer."&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Volkswagen&lt;/strong&gt; chose DaVoice for OTTO, its generative AI car assistant launched in Brazil (&lt;a href="https://davoice.io/use-case/automotive" rel="noopener noreferrer"&gt;customer story&lt;/a&gt; · &lt;a href="https://www.youtube.com/watch?v=VGfJDX3VfVg" rel="noopener noreferrer"&gt;video&lt;/a&gt;).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;LunaFit&lt;/strong&gt; built its entire voice-to-voice AI agent on DaVoice, including speaker-aware speech-to-text that hears its user in a noisy gym (&lt;a href="https://www.youtube.com/watch?v=uYpaCXAvjew" rel="noopener noreferrer"&gt;video&lt;/a&gt;). The interruption pattern in section 8 comes from LunaFit's workout mode.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/VGfJDX3VfVg" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;

&lt;p&gt;  &lt;iframe src="https://www.youtube.com/embed/uYpaCXAvjew" width="710" height="399"&gt;
  &lt;/iframe&gt;
&lt;/p&gt;




&lt;h2&gt;
  
  
  Table of contents
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;What you will build&lt;/li&gt;
&lt;li&gt;Architecture: wake word → STT → local LLM → TTS&lt;/li&gt;
&lt;li&gt;What is really offline (and what is not)&lt;/li&gt;
&lt;li&gt;Prerequisites&lt;/li&gt;
&lt;li&gt;Run the DaVoice example app&lt;/li&gt;
&lt;li&gt;Implementation, stage by stage&lt;/li&gt;
&lt;li&gt;Add the agent to your existing React Native app&lt;/li&gt;
&lt;li&gt;Advanced: speaker verification, voice cloning, interruptions&lt;/li&gt;
&lt;li&gt;Testing and troubleshooting&lt;/li&gt;
&lt;li&gt;FAQ&lt;/li&gt;
&lt;li&gt;Resources&lt;/li&gt;
&lt;/ol&gt;




&lt;h2&gt;
  
  
  1. What you will build
&lt;/h2&gt;

&lt;p&gt;A hands-free assistant inside a React Native app (iOS and Android) that:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;listens continuously for a custom wake word ("Hey Coach" in the example);&lt;/li&gt;
&lt;li&gt;optionally checks that the speaker is the enrolled owner;&lt;/li&gt;
&lt;li&gt;transcribes the request in real time;&lt;/li&gt;
&lt;li&gt;lets a &lt;strong&gt;0.6B-parameter on-device model&lt;/strong&gt; decide whether to call one of your app's tools, ask a clarifying question, decline, or hand the request to an (optional) remote model;&lt;/li&gt;
&lt;li&gt;speaks the result with an on-device neural voice, which can be a cloned voice.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The example ships two demos you can try immediately:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Demo&lt;/th&gt;
&lt;th&gt;Model (downloaded on first use)&lt;/th&gt;
&lt;th&gt;Tools&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Car demo&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;DaVoice Automotive Agent 0.6B (~380 MB GGUF)&lt;/td&gt;
&lt;td&gt;22 simulated car tools: windows, doors, locks, climate, navigation, radio&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;strong&gt;Calories demo&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;DaVoice Personal Agent 0.6B (~380 MB GGUF)&lt;/td&gt;
&lt;td&gt;calorie tracker, personal memory, web search&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Honest status from the repo:&lt;/strong&gt; the automotive model is labeled a &lt;strong&gt;test model, still in training&lt;/strong&gt;. On DaVoice's frozen automotive test set it gets about 82% of Brazilian Portuguese commands exactly right (tool + arguments) and picks the right tool about 90% of the time. Every car action is &lt;strong&gt;simulated&lt;/strong&gt; and checked by app-side guards. Treat it as a preview of the voice pipeline, not final accuracy.&lt;/p&gt;
&lt;/blockquote&gt;




&lt;h2&gt;
  
  
  2. Architecture: how a Voice-to-Voice AI Agent works on-device
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;                 ┌────────────────────────── on the phone ──────────────────────────┐
  microphone ──► │ Wake word  ──►  Speaker      ──►  Speech-to-  ──►  Local LLM    │
                 │ "Hey Coach"     verification      text (STT)       Agent        │
                 │ (react-native-  (optional,        (react-native-   (react-native-│
                 │  wakeword)       owner only)       davoice)         davoice-llm) │
                 │                                                       │   ▲      │
                 │                                          tool call ──►│   │result│
                 │                                                  App tools (JS)  │
                 │                                                       │          │
  speaker  ◄──── │ Text-to-speech (react-native-davoice) ◄── sentence-by-sentence ─┘ │
                 └──────────────────────────────────────────────────────────────────┘
                     optional, needs network: remote model hand-over, web search, MCP
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The loop, step by step:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Wake word&lt;/strong&gt; runs continuously at low power. When it fires, detection is &lt;strong&gt;paused (not stopped)&lt;/strong&gt; and the mic is handed to STT.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Speaker verification&lt;/strong&gt; (optional) tags every STT result with &lt;code&gt;speakerId&lt;/code&gt; / &lt;code&gt;speakerIds&lt;/code&gt;, and can restrict the wake word to the enrolled voice.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;STT&lt;/strong&gt; streams partial results. The end of the user's turn is a &lt;strong&gt;silence timeout&lt;/strong&gt; (2,000 ms in the example).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The local agent&lt;/strong&gt; (&lt;code&gt;Agent.turn&lt;/code&gt;) gets the transcript plus history. The fine-tuned model decides: call a tool, ask, decline, or hand over.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Tool calling:&lt;/strong&gt; your JavaScript handler runs and returns &lt;code&gt;toolReply(result, sentence)&lt;/code&gt;. The app, not the model, says what actually happened, and the turn ends without a second model round.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;TTS&lt;/strong&gt; speaks each sentence as soon as it is complete. STT is paused while the agent speaks, then the mic reopens.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Why a pipeline instead of one speech-to-speech model? Each stage can be tuned, tested and replaced on its own; tools can be gated by speaker identity; and a 0.6B action model stays small enough for a phone.&lt;/p&gt;




&lt;h2&gt;
  
  
  3. What is really offline (and what is not)
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Stage&lt;/th&gt;
&lt;th&gt;Runs on device&lt;/th&gt;
&lt;th&gt;Notes&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Wake word&lt;/td&gt;
&lt;td&gt;Yes&lt;/td&gt;
&lt;td&gt;model shipped in the app bundle&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Speaker verification&lt;/td&gt;
&lt;td&gt;Yes&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;sv.dm&lt;/code&gt; shipped in the app bundle&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Speech-to-text&lt;/td&gt;
&lt;td&gt;Yes, for installed on-device languages&lt;/td&gt;
&lt;td&gt;DaVoice's guide notes that languages without an installed on-device model fall back to cloud recognition&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Local LLM + tool calling&lt;/td&gt;
&lt;td&gt;Yes, after a one-time model download&lt;/td&gt;
&lt;td&gt;GGUF downloaded on first use, checked by sha256&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Text-to-speech + voice cloning&lt;/td&gt;
&lt;td&gt;Yes&lt;/td&gt;
&lt;td&gt;English and Brazilian Portuguese voices are bundled; other languages download once&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Remote model hand-over&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;No (optional, off by default)&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;LOCAL_LLM_LAYER2 = 'none'&lt;/code&gt; in the example&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Web search / MCP tools&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;No (optional)&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;Wikipedia by default in the calories demo&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;"Full AI Chat" mode (Gemini)&lt;/td&gt;
&lt;td&gt;&lt;strong&gt;No (optional cloud demo)&lt;/strong&gt;&lt;/td&gt;
&lt;td&gt;needs your own Gemini key, not part of the offline pipeline&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;So: once models are downloaded, the core loop (wake word → STT → local agent → app tools → TTS) works in airplane mode for supported languages. Anything labeled optional above is &lt;strong&gt;cloud&lt;/strong&gt; and is not part of the offline pipeline.&lt;/p&gt;




&lt;h2&gt;
  
  
  4. Prerequisites
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Real devices.&lt;/strong&gt; Wake word, speaker verification, STT and the LLM need a real microphone and real memory. Use simulators for UI only.

&lt;ul&gt;
&lt;li&gt;iPhone: iPhone 11 or newer per DaVoice's guide (Metal GPU on A14+; CPU on A13).&lt;/li&gt;
&lt;li&gt;Android: 64-bit ARM (&lt;code&gt;arm64-v8a&lt;/code&gt;), &lt;code&gt;minSdkVersion 29&lt;/code&gt;.&lt;/li&gt;
&lt;/ul&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;React Native CLI app.&lt;/strong&gt; The example uses React Native &lt;strong&gt;0.82.0&lt;/strong&gt; and React 19.1.1. (An Expo version lives in the same repo under &lt;code&gt;expo/&lt;/code&gt;; it needs a development build, not Expo Go.)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Node 18+&lt;/strong&gt;, Xcode + CocoaPods, Android Studio.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Git LFS&lt;/strong&gt;, because model files are stored with LFS.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;A DaVoice license key&lt;/strong&gt; for your own app (the example contains a time-limited demo key; don't copy it). Contact &lt;code&gt;info@davoice.io&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Disk/network for the first run:&lt;/strong&gt; the car demo downloads about 780 MB (model + pre-warmed state) once.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Versions used by the example (&lt;code&gt;example/package.json&lt;/code&gt;) - use these together, since all three DaVoice packages ship native binaries:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="nl"&gt;"react-native"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"0.82.0"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="nl"&gt;"react-native-davoice"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"1.0.140"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="nl"&gt;"react-native-davoice-llm"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"1.0.13"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="nl"&gt;"react-native-fs"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"^2.20.0"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="nl"&gt;"react-native-wakeword"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"1.1.171"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;






&lt;h2&gt;
  
  
  5. Run the DaVoice example app
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Want to hear it before you build?&lt;/strong&gt; The &lt;a href="https://davoice.io/playground" rel="noopener noreferrer"&gt;DaVoice Playground&lt;/a&gt; runs in your browser: &lt;a href="https://davoice.io/playground-app/#wakeword" rel="noopener noreferrer"&gt;try the wake word&lt;/a&gt; and &lt;a href="https://davoice.io/playground-app/#tts" rel="noopener noreferrer"&gt;listen to the text-to-speech voices&lt;/a&gt; without installing anything.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;git clone https://github.com/frymanofer/ReactNative_DaVoice.git
&lt;span class="nb"&gt;cd&lt;/span&gt; ReactNative_DaVoice
git lfs pull                 &lt;span class="c"&gt;# model files are LFS objects&lt;/span&gt;
&lt;span class="nb"&gt;cd&lt;/span&gt; example
npm install                  &lt;span class="c"&gt;# the project uses npm / package-lock.json&lt;/span&gt;
&lt;span class="nb"&gt;cd&lt;/span&gt; ios &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; pod install &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="nb"&gt;cd&lt;/span&gt; ..
npx react-native run-android &lt;span class="c"&gt;# or: npx react-native run-ios&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Useful extras from the repo's how-to:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;adb reverse tcp:8081 tcp:8081                 &lt;span class="c"&gt;# Android debug build over USB (Metro)&lt;/span&gt;
adb logcat &lt;span class="o"&gt;|&lt;/span&gt; grep ReactNativeJS               &lt;span class="c"&gt;# JS logs on Android&lt;/span&gt;
&lt;span class="nb"&gt;cd&lt;/span&gt; android &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; ./gradlew :app:assembleRelease  &lt;span class="c"&gt;# Android release build&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;First run walkthrough:&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Pick a language (English and Brazilian Portuguese are built in; others download once).&lt;/li&gt;
&lt;li&gt;Accept the microphone permission.&lt;/li&gt;
&lt;li&gt;Pick a voice, or clone a new one from a recording.&lt;/li&gt;
&lt;li&gt;Speaker verification: create a new voice signature (5 samples), reuse a saved one, or skip.&lt;/li&gt;
&lt;li&gt;Say &lt;strong&gt;"Hey Coach"&lt;/strong&gt;, then pick &lt;strong&gt;Local AI · Car demo&lt;/strong&gt;.&lt;/li&gt;
&lt;li&gt;Wait for the loading screen (download → load into memory → warm up). The car demo ships a pre-warmed prompt state, so after the download it is ready in about a second.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Then try: "open the rear right window", "set the temperature to 22 degrees", "lock the car", "how much range do I have?". Incomplete requests should trigger a question: "open a window" → "Which window?".&lt;/p&gt;

&lt;p&gt;A general question like "tell me a fun fact about Brazil" returns &lt;strong&gt;"I need to send this to a remote Model…"&lt;/strong&gt;. That is expected: no remote model is configured, and the 0.6B model is trained to act, not to chat.&lt;/p&gt;




&lt;h2&gt;
  
  
  6. Implementation, stage by stage
&lt;/h2&gt;

&lt;p&gt;The example's voice logic is split into copyable modules under &lt;a href="https://github.com/frymanofer/ReactNative_DaVoice/tree/main/example/src" rel="noopener noreferrer"&gt;&lt;code&gt;example/src/&lt;/code&gt;&lt;/a&gt;. &lt;code&gt;App.tsx&lt;/code&gt; is the orchestration layer. The snippets below are the minimal APIs documented in the repo, plus excerpts from the real modules.&lt;/p&gt;

&lt;h3&gt;
  
  
  6.1 Install and native setup
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npm install react-native-wakeword react-native-davoice react-native-davoice-llm react-native-fs
&lt;span class="nb"&gt;cd&lt;/span&gt; ios &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; pod install &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="nb"&gt;cd&lt;/span&gt; ..
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Android&lt;/strong&gt; - each DaVoice package ships its native library as a local Maven repo. From &lt;code&gt;example/android/build.gradle&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight gradle"&gt;&lt;code&gt;&lt;span class="c1"&gt;// minSdkVersion = 29&lt;/span&gt;
&lt;span class="n"&gt;allprojects&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="n"&gt;repositories&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="n"&gt;maven&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;url&lt;/span&gt; &lt;span class="s2"&gt;"${project("&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="n"&gt;react&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="k"&gt;native&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;wakeword&lt;/span&gt;&lt;span class="s2"&gt;").projectDir}/libs"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="n"&gt;maven&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;url&lt;/span&gt; &lt;span class="s2"&gt;"${project("&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="n"&gt;react&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="k"&gt;native&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;davoice&lt;/span&gt;&lt;span class="s2"&gt;").projectDir}/libs"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="n"&gt;maven&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;url&lt;/span&gt; &lt;span class="s2"&gt;"${project("&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="n"&gt;react&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="k"&gt;native&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;davoice&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;llm&lt;/span&gt;&lt;span class="s2"&gt;").projectDir}/libs"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Add &lt;code&gt;RECORD_AUDIO&lt;/code&gt; and &lt;code&gt;INTERNET&lt;/code&gt; (for the one-time model download) to &lt;code&gt;AndroidManifest.xml&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;iOS&lt;/strong&gt; - in &lt;code&gt;Info.plist&lt;/code&gt;: &lt;code&gt;NSMicrophoneUsageDescription&lt;/code&gt;, &lt;code&gt;NSSpeechRecognitionUsageDescription&lt;/code&gt;, and &lt;code&gt;UIBackgroundModes&lt;/code&gt; → &lt;code&gt;audio&lt;/code&gt; if the wake word should listen in the background. For the LLM, add the entitlement the example uses:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight xml"&gt;&lt;code&gt;&lt;span class="nt"&gt;&amp;lt;key&amp;gt;&lt;/span&gt;com.apple.developer.kernel.increased-memory-limit&lt;span class="nt"&gt;&amp;lt;/key&amp;gt;&lt;/span&gt;
&lt;span class="nt"&gt;&amp;lt;true/&amp;gt;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Metro&lt;/strong&gt; - let &lt;code&gt;require()&lt;/code&gt; resolve model files (&lt;code&gt;metro.config.js&lt;/code&gt;):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="n"&gt;resolver&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="n"&gt;assetExts&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;defaultConfig&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;resolver&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;assetExts&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s1"&gt;'dm'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s1"&gt;'onnx'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s1"&gt;'wav'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s1"&gt;'mp3'&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Model files loaded by name&lt;/strong&gt; (wake word &lt;code&gt;hey_coach_model_28_22012026b.dm&lt;/code&gt;, speaker verification &lt;code&gt;sv.dm&lt;/code&gt;) go into &lt;code&gt;android/app/src/main/assets/&lt;/code&gt; and into the Xcode target's &lt;em&gt;Copy Bundle Resources&lt;/em&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  6.2 Wake word (&lt;code&gt;react-native-wakeword&lt;/code&gt;)
&lt;/h3&gt;

&lt;p&gt;From the repo's integration guide:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;createKeyWordRNBridgeInstance&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;setWakewordAudioRoutingConfig&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="s1"&gt;'react-native-wakeword'&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;setWakewordAudioRoutingConfig&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;defaultAudioRoutingConfig&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="c1"&gt;// iOS routing, from src/wakeword/index.ts&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;inst&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;createKeyWordRNBridgeInstance&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'main'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="c1"&gt;// One or more models: names, thresholds, buffer counts, ms between callbacks.&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;inst&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;createInstanceMulti&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s1"&gt;'your_phrase_model.dm'&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mf"&gt;0.999&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;3&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1000&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;sub&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;inst&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;onKeywordDetectionEvent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;phrase&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;string&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="nf"&gt;onWakeWord&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;phrase&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;inst&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setKeywordDetectionLicense&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="no"&gt;DAVOICE_LICENSE_KEY&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="cm"&gt;/* show license error */&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;inst&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;startKeywordDetection&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;0.999&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;In the callback, &lt;strong&gt;pause&lt;/strong&gt; detection before giving the mic to STT, and un-pause afterwards:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;inst&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;pauseDetection&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;   &lt;span class="c1"&gt;// false: keep the mic session&lt;/span&gt;
&lt;span class="c1"&gt;// ... STT / LLM / TTS ...&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;inst&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;unPauseDetection&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Tuning knobs: &lt;code&gt;threshold&lt;/code&gt; (0-1) and &lt;code&gt;bufferCnt&lt;/code&gt; trade false triggers against misses (the example uses &lt;code&gt;0.999&lt;/code&gt; / &lt;code&gt;3&lt;/code&gt;); &lt;code&gt;msBetweenCallbacks&lt;/code&gt; debounces repeats. Custom wake phrases are separate models you get from DaVoice. To see how detection behaves before you write code, test it in the &lt;a href="https://davoice.io/playground-app/#wakeword" rel="noopener noreferrer"&gt;wake word playground&lt;/a&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  6.3 Speech-to-text and text-to-speech (&lt;code&gt;react-native-davoice&lt;/code&gt;)
&lt;/h3&gt;

&lt;p&gt;STT and TTS share one &lt;code&gt;Speech&lt;/code&gt; object because they share one audio session:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="s1"&gt;'react-native-davoice/speech'&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setAudioRoutingConfig&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;defaultAudioRoutingConfig&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="c1"&gt;// iOS; same object as the wake word&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setLicense&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="no"&gt;DAVOICE_LICENSE_KEY&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="cm"&gt;/* show license error */&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;initAll&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;locale&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'en-US'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;require&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'./assets/models/model_ex2_rich_kate.dm'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;pauseSpeechRecognition&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="c1"&gt;// start paused; un-pause when it's the user's turn&lt;/span&gt;

&lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;onSpeechPartialResults&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;e&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="cm"&gt;/* e.value[0]: text so far */&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;onSpeechResults&lt;/span&gt;        &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;e&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="cm"&gt;/* e.value[0]: final text; e.speakerId when verification is on */&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;onSpeechError&lt;/span&gt;          &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;e&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="cm"&gt;/* log; most codes are recoverable */&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;unPauseSpeechRecognition&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="c1"&gt;// listen&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;pauseSpeechRecognition&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;     &lt;span class="c1"&gt;// stop listening (e.g. before TTS)&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;speak&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'I opened the window.'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mf"&gt;1.0&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="c1"&gt;// text, speaker index, speed&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;To pick a voice, listen to them in the &lt;a href="https://davoice.io/playground-app/#tts" rel="noopener noreferrer"&gt;TTS playground&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;&lt;code&gt;initAll&lt;/code&gt; can take many seconds the first time, so the example wraps it in a timeout and keeps the wake word running if it stalls (&lt;code&gt;initializeWakewordBootstrap&lt;/code&gt; in &lt;code&gt;src/wakeword/index.ts&lt;/code&gt;):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;withTimeout&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
  &lt;span class="nf"&gt;initializeSpeechLibrary&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;enrollmentJsonPath&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="kc"&gt;null&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;speechInitTimeoutMs&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="s1"&gt;'Speech.initAll'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  6.4 The local LLM and the Agent (&lt;code&gt;react-native-davoice-llm&lt;/code&gt;)
&lt;/h3&gt;

&lt;p&gt;The package has two layers: an &lt;strong&gt;engine&lt;/strong&gt; (&lt;code&gt;setLicense&lt;/code&gt;, &lt;code&gt;loadModel&lt;/code&gt;, &lt;code&gt;generateStream&lt;/code&gt;, &lt;code&gt;prefill&lt;/code&gt;, &lt;code&gt;cancel&lt;/code&gt;, &lt;code&gt;saveState&lt;/code&gt;, &lt;code&gt;loadState&lt;/code&gt;) and an &lt;strong&gt;&lt;code&gt;Agent&lt;/code&gt;&lt;/strong&gt; that runs a whole assistant turn - prompt assembly, tool calls, guards, memory and streaming - configured by an &lt;strong&gt;agent pack&lt;/strong&gt; of JSON files instead of code. DaVoice's &lt;a href="https://davoice.io/voice-agent-guide/on-device-llm" rel="noopener noreferrer"&gt;on-device LLM guide&lt;/a&gt; covers model choice in more depth.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Models.&lt;/strong&gt; The GGUF files are downloaded at runtime and checked by sha256 - never bundled in the binary. From &lt;code&gt;src/local_llm/constants.ts&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="n"&gt;id&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="no"&gt;LOCAL_LLM_DAVOICE_AUTOMOTIVE_R4B&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;displayName&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'DaVoice Automotive Agent 0.6B'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;filename&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'davoice-automotive-qwen3-0.6b-r4b-it100-q4_0.gguf'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;url&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'https://pub-27f94b1602c647ab95a601086cf52820.r2.dev/agent-llm/auto5_r4b_it100-q4_0.gguf'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;sizeBytes&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;381_565_920&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;sha256&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'1c078e961c5f4e16ae0f667d213fa06de53f3f4bc92c9248427bdb37237bd205'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;contextSize&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;8192&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;chatTemplate&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="no"&gt;CHAT_TEMPLATE_PRESETS&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;qwen3&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;template&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="c1"&gt;// ...&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The picker also lists general models (Qwen2.5 1.5B/3B with &lt;code&gt;chatml&lt;/code&gt;, Gemma 3n E2B/E4B with &lt;code&gt;gemma&lt;/code&gt;). They are larger and slower; the 0.6B agents are the defaults.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Minimal agent setup&lt;/strong&gt; (from the repo's &lt;code&gt;englishHowTo.md&lt;/code&gt;):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;setLicense&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;loadModel&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nc"&gt;Agent&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;toolReply&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="s1"&gt;'react-native-davoice-llm'&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;setLicense&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="no"&gt;LICENSE&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;loadModel&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;pathToGguf&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;contextSize&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;8192&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;agent&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Agent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;create&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="n"&gt;pack&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;require&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'./agent_automotive/agent.json'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;skills&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;databasePath&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sb"&gt;`${RNFS.DocumentDirectoryPath}/davoice_agent_car.db`&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;chatTemplate&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'qwen3'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;onRequest&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;configure&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;intentRouter&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;webSearchTool&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;timeTool&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;guards&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="cm"&gt;/* see provider.ts */&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setTools&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;myTools&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;   &lt;span class="c1"&gt;// a handler returns toolReply(result, 'I opened the window.') to say it itself&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;r&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;turn&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;history&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;ev&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;ev&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="kd"&gt;type&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="s1"&gt;'text'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="nf"&gt;speak&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;ev&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="c1"&gt;// agent.cancel() stops a running turn&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The example's real configuration for the fine-tuned models (&lt;code&gt;src/aichat/provider.ts&lt;/code&gt;, condensed - the original reads &lt;code&gt;temperature&lt;/code&gt; and memory options from the model entry) shows the settings that matter:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;configure&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="n"&gt;chatTemplate&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;chatTemplate&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;remote&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="n"&gt;handsOver&lt;/span&gt; &lt;span class="o"&gt;?&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;
    &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="no"&gt;LOCAL_LLM_LAYER2&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="s1"&gt;'device'&lt;/span&gt; &lt;span class="o"&gt;?&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;enabled&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;answer&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'layer2'&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;enabled&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;timeoutMs&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;15000&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;intentRouter&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;usesRouter&lt;/span&gt; &lt;span class="o"&gt;!==&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;   &lt;span class="c1"&gt;// false: the model decides with every tool&lt;/span&gt;
  &lt;span class="n"&gt;webSearchTool&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;modelTools&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;timeTool&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;modelTools&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;memory&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;profile&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'small'&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;sampling&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;temperature&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;               &lt;span class="c1"&gt;// greedy for tool calling&lt;/span&gt;
  &lt;span class="n"&gt;guards&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nf"&gt;callGuardSettings&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;callGuards&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="s1"&gt;'off'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;historyAsRun&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;historyAsRun&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;historyAsRun&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;guidedToolCalls&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;guidedToolCalls&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;promptOrder&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;promptOrder&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="s1"&gt;'trained'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Create the agent once&lt;/strong&gt; and answer its requests (&lt;code&gt;src/agent/LocalAgent.ts&lt;/code&gt;). &lt;code&gt;onRequest&lt;/code&gt; is where the app decides about remote hand-over and sensitive tools:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="n"&gt;onRequest&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="k"&gt;async&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="kd"&gt;type&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="s1"&gt;'remote'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nf"&gt;answerRemoteRequest&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt; &lt;span class="k"&gt;as&lt;/span&gt; &lt;span class="nc"&gt;AgentRemoteRequest&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="kd"&gt;type&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="s1"&gt;'authorizeTool'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;authorized&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nf"&gt;isToolAuthorized&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt; &lt;span class="k"&gt;as&lt;/span&gt; &lt;span class="nc"&gt;AgentApprovalRequest&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="kd"&gt;type&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="s1"&gt;'confirmTool'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;confirmed&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;askToConfirm&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;request&lt;/span&gt; &lt;span class="k"&gt;as&lt;/span&gt; &lt;span class="nc"&gt;AgentApprovalRequest&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;throw&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nb"&gt;Error&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sb"&gt;`The app does not answer "${request.type}" requests.`&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  6.5 Tool calling: define tools in MCP shape, let the app confirm
&lt;/h3&gt;

&lt;p&gt;Tools are plain JSON in MCP shape (&lt;code&gt;name&lt;/code&gt;, &lt;code&gt;description&lt;/code&gt;, &lt;code&gt;inputSchema&lt;/code&gt;, optional &lt;code&gt;execution&lt;/code&gt;). From &lt;code&gt;src/car/carToolDefinitions.h3.json&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"set_window"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"description"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Opens or closes windows, fully, halfway or just a crack."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nl"&gt;"inputSchema"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"object"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"properties"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nl"&gt;"window"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"string"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="nl"&gt;"enum"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"all"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"front_left"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"front_right"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"rear_left"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"rear_right"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
                 &lt;span class="s2"&gt;"driver"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"passenger"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"front"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"rear"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"left"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"right"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="nl"&gt;"description"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Which window: ... Ask if the user did not say which"&lt;/span&gt;
      &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
      &lt;span class="nl"&gt;"position"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"string"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="nl"&gt;"enum"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"open"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"close"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"half"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"crack"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="nl"&gt;"description"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"open (lower it), close (raise it), half, crack (a little)"&lt;/span&gt;
      &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="nl"&gt;"required"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="s2"&gt;"window"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="s2"&gt;"position"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A sensitive tool declares an authorization scope, and the agent core &lt;strong&gt;never runs it&lt;/strong&gt; without a "yes" from the app:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight json"&gt;&lt;code&gt;&lt;span class="nl"&gt;"execution"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nl"&gt;"authorization"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nl"&gt;"scope"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"door_locks"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nl"&gt;"message"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"Unlock the car?"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The JS handler does the work and returns the sentence to speak. This is the pattern in &lt;code&gt;src/car/carTools.ts&lt;/code&gt; (trimmed):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;toolReply&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="s1"&gt;'react-native-davoice-llm'&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;toTool&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;t&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;any&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;ToolDefinition&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;t&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;description&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;t&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;description&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;inputSchema&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;t&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;inputSchema&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;t&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;execution&lt;/span&gt; &lt;span class="o"&gt;?&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;execution&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;t&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;execution&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="n"&gt;handler&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;rawArgs&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nc"&gt;Args&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;args&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;rawArgs&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
      &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;run&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="no"&gt;HANDLERS&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;t&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
      &lt;span class="k"&gt;try&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="n"&gt;run&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;throw&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;Refused&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;"Sorry, this car doesn't have that function."&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
        &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;summary&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;result&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;say&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;run&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;args&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
        &lt;span class="c1"&gt;// The app says what it did and the turn ends there: no second model round.&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;say&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nf"&gt;toolReply&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;ok&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;say&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;ok&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
      &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;catch&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;error&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
        &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;reason&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;error&lt;/span&gt; &lt;span class="k"&gt;instanceof&lt;/span&gt; &lt;span class="nb"&gt;Error&lt;/span&gt; &lt;span class="o"&gt;?&lt;/span&gt; &lt;span class="n"&gt;error&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;message&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;String&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;error&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;error&lt;/span&gt; &lt;span class="k"&gt;instanceof&lt;/span&gt; &lt;span class="nc"&gt;Refused&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nf"&gt;toolReply&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;ok&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;refused&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;reason&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;reason&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
        &lt;span class="k"&gt;throw&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nb"&gt;Error&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sb"&gt;`Not done: ${reason}`&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
      &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Why &lt;code&gt;toolReply&lt;/code&gt;? Two reasons from the repo's field testing: the spoken confirmation is built from the &lt;strong&gt;real result&lt;/strong&gt; ("The front right door is already locked."), so the model can't announce something that didn't happen; and skipping the second model round saves seconds on a phone CPU. More patterns are in the &lt;a href="https://davoice.io/voice-agent-guide/tool-calling" rel="noopener noreferrer"&gt;tool calling guide&lt;/a&gt;.&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;strong&gt;Tip for small models:&lt;/strong&gt; a 0.6B model depends heavily on tool descriptions. Keep them short, say explicitly "absolute vs. relative" and which side is which, and keep &lt;strong&gt;exactly&lt;/strong&gt; the wording the model was trained on (the car model expects &lt;code&gt;carToolDefinitions.h3.json&lt;/code&gt;).&lt;/p&gt;
&lt;/blockquote&gt;

&lt;h3&gt;
  
  
  6.6 The voice loop: one turn, streamed into TTS
&lt;/h3&gt;

&lt;p&gt;This is the order of calls that keeps mic hand-offs stable (from &lt;code&gt;docs/AI_AGENT_INTEGRATION_GUIDE.md&lt;/code&gt;):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;startup:
  mic permission
  setWakewordAudioRoutingConfig + Speech.setAudioRoutingConfig   (iOS, same object)
  create wake-word instance, attach listener, license
  startKeywordDetection
  pauseDetection(false)
  Speech.setLicense, Speech.initAll (with timeout), Speech.pauseSpeechRecognition
  unPauseDetection                    -&amp;gt; listening for the wake word

wake word detected:
  pauseDetection(false)
  Speech.unPauseSpeechRecognition(-1) -&amp;gt; user talks; partials -&amp;gt; agent.prefill
  silence timeout:
    Speech.pauseSpeechRecognition
    agent.turn                        -&amp;gt; fillers and sentences go to Speech.speak as they arrive
  all sentences spoken:
    Speech.unPauseSpeechRecognition(-1)  (follow-up turn)
    or unPauseDetection                  (back to the wake word)
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The turn itself (&lt;code&gt;src/aichat/provider.ts&lt;/code&gt;):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;turn&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
  &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;userText&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;history&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nf"&gt;toAgentHistory&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;history&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;profile&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;longTermFacts&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;localPassages&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;date&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nf"&gt;currentDateText&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;event&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;event&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="kd"&gt;type&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="s1"&gt;'text'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="nf"&gt;onDelta&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;event&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="n"&gt;onEvent&lt;/span&gt;&lt;span class="o"&gt;?&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;event&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="nf"&gt;onDelta&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;''&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;toolMessages&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;result&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;toolMessages&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;webSearch&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;result&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;webSearch&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Keep &lt;code&gt;result.toolMessages&lt;/code&gt; in the history so the next turn knows what ran ("no, the other one at the back" depends on it).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Speak sentence by sentence, without awaiting each one.&lt;/strong&gt; The example buffers streamed text and cuts complete sentences (&lt;code&gt;splitCompletedSentences&lt;/code&gt; in &lt;code&gt;src/aichat/index.ts&lt;/code&gt;), carefully not splitting "DaVoice.io" or "12.5":&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// A period ends a sentence only before a space (or a closing quote): not inside "DaVoice.io", "12.5" or "e.g".&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;index&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="s1"&gt;'.'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;next&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;index&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;next&lt;/span&gt; &lt;span class="o"&gt;===&lt;/span&gt; &lt;span class="kc"&gt;undefined&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="o"&gt;!/&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;\&lt;span class="n"&gt;s&lt;/span&gt;&lt;span class="s2"&gt;"')\]]/.test(next)) continue;
}
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Each complete sentence is handed to &lt;code&gt;Speech.speak&lt;/code&gt; immediately; the native library queues them, so there are no gaps. The reply is finished when every queued sentence has resolved - not on the first &lt;code&gt;onFinishedSpeaking&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Latency tricks used by the example:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Warm up&lt;/strong&gt; when the chat screen opens: download, load, and process the system prompt (~3,100 tokens for the car demo).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Save and restore the processed prompt&lt;/strong&gt; with &lt;code&gt;saveState&lt;/code&gt; / &lt;code&gt;loadState&lt;/code&gt;. Measured on a Galaxy S20 (CPU): restore 0.7-0.9 s instead of a 30-40 s first warm-up. The car model ships a pre-warmed state (restore 0.23 s, first answer 0.34 s).&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Prefill while the user talks:&lt;/strong&gt; call &lt;code&gt;agent.prefill(history, partialText)&lt;/code&gt; on partial transcripts so only the last words remain when the silence timeout fires.&lt;/li&gt;
&lt;li&gt;Measured decision time on a Galaxy S20, CPU only: about &lt;strong&gt;1.3-2 s per command&lt;/strong&gt;; iPhones with Metal are faster.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  7. Add the agent to your existing React Native app
&lt;/h2&gt;

&lt;p&gt;The repo contains two documents written specifically for this: &lt;a href="https://github.com/frymanofer/ReactNative_DaVoice/blob/main/example/englishHowTo.md" rel="noopener noreferrer"&gt;&lt;code&gt;example/englishHowTo.md&lt;/code&gt;&lt;/a&gt; and &lt;a href="https://github.com/frymanofer/ReactNative_DaVoice/blob/main/docs/AI_AGENT_INTEGRATION_GUIDE.md" rel="noopener noreferrer"&gt;&lt;code&gt;docs/AI_AGENT_INTEGRATION_GUIDE.md&lt;/code&gt;&lt;/a&gt;. They're written so you can hand them to an AI coding agent (Claude Code, Cursor, Copilot) along with your codebase. The core rule: &lt;strong&gt;copy the module from &lt;code&gt;src/&lt;/code&gt; plus its assets, install the package, use your own license.&lt;/strong&gt;&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;You want&lt;/th&gt;
&lt;th&gt;Copy&lt;/th&gt;
&lt;th&gt;Assets / setup&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Wake word&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;src/wakeword/index.ts&lt;/code&gt; (&lt;code&gt;instanceConfigs&lt;/code&gt;, &lt;code&gt;initializeWakewordBootstrap&lt;/code&gt;, &lt;code&gt;attachKeywordListenerOnce&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;wake word &lt;code&gt;.dm&lt;/code&gt; in native bundle&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Speaker verification&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;src/speaker_verification/&lt;/code&gt; (&lt;code&gt;onboarding.ts&lt;/code&gt;, &lt;code&gt;verification.ts&lt;/code&gt;, &lt;code&gt;currentSpeaker.ts&lt;/code&gt;), thresholds in &lt;code&gt;src/appflow.ts&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;sv.dm&lt;/code&gt; in native bundle&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;STT + TTS&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;src/initialization/&lt;/code&gt;, &lt;code&gt;src/stt/&lt;/code&gt;, &lt;code&gt;src/tts/&lt;/code&gt; (+ &lt;code&gt;voiceClone.ts&lt;/code&gt;)&lt;/td&gt;
&lt;td&gt;TTS model + voice-cloning encoder per language&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Local agent + tools&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;src/local_llm/&lt;/code&gt;, &lt;code&gt;src/agent/LocalAgent.ts&lt;/code&gt;, &lt;code&gt;src/agent/AgentPack.ts&lt;/code&gt;, &lt;code&gt;src/aichat/provider.ts&lt;/code&gt;, &lt;code&gt;src/aichat/fragments.ts&lt;/code&gt;, &lt;code&gt;src/aichat/remote.ts&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;agent pack JSON (&lt;code&gt;agent/&lt;/code&gt; or &lt;code&gt;agent_automotive/&lt;/code&gt;), your tool JSON&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;&lt;strong&gt;A safe integration order:&lt;/strong&gt;&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Install only the packages you need and add the Android Maven lines / iOS plist entries. Build.&lt;/li&gt;
&lt;li&gt;Wake word: ship your &lt;code&gt;.dm&lt;/code&gt;, copy &lt;code&gt;defaultAudioRoutingConfig&lt;/code&gt; and &lt;code&gt;initializeWakewordBootstrap&lt;/code&gt;. Build and test on a device.&lt;/li&gt;
&lt;li&gt;STT/TTS: &lt;code&gt;Speech.initAll&lt;/code&gt; with a timeout; register handlers from &lt;code&gt;src/stt/index.ts&lt;/code&gt; (keep &lt;code&gt;mergeSmartKeepPunct&lt;/code&gt; - Android restarts partials per segment). Build and test.&lt;/li&gt;
&lt;li&gt;Agent: copy &lt;code&gt;src/local_llm/&lt;/code&gt; for download/sha256/load-once, &lt;code&gt;getLocalAgent()&lt;/code&gt; for a single agent, and &lt;code&gt;provider.ts&lt;/code&gt; for configure + warm-up + streaming.&lt;/li&gt;
&lt;li&gt;Replace the car tools with your own tool JSON and handlers that return &lt;code&gt;toolReply(...)&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Edit the agent pack (persona, fixed replies, messages) - behavior changes without code changes.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;&lt;strong&gt;Rules worth following from the guide:&lt;/strong&gt;&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Never copy the example's demo license key. Put a placeholder like &lt;code&gt;DAVOICE_LICENSE_KEY&lt;/code&gt; in a git-ignored &lt;code&gt;local.config.ts&lt;/code&gt; (template: &lt;code&gt;local.config.example.ts&lt;/code&gt;).&lt;/li&gt;
&lt;li&gt;Keep &lt;strong&gt;one&lt;/strong&gt; wake-word instance, &lt;strong&gt;one&lt;/strong&gt; listener and &lt;strong&gt;one&lt;/strong&gt; agent.&lt;/li&gt;
&lt;li&gt;Keep the &lt;strong&gt;order of native calls&lt;/strong&gt; the same as the example; that order is what makes audio hand-offs stable.&lt;/li&gt;
&lt;li&gt;Don't &lt;code&gt;require()&lt;/code&gt; a &lt;code&gt;.dm&lt;/code&gt; you don't use - every &lt;code&gt;require()&lt;/code&gt; ships the file.&lt;/li&gt;
&lt;li&gt;Don't give a &lt;code&gt;.ts&lt;/code&gt; module and a &lt;code&gt;.json&lt;/code&gt; file the same base name; Metro resolves the &lt;code&gt;.json&lt;/code&gt; first.&lt;/li&gt;
&lt;li&gt;Change one feature at a time and build after each.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  8. Advanced capabilities
&lt;/h2&gt;

&lt;h3&gt;
  
  
  Speaker verification: an owner-only agent
&lt;/h3&gt;

&lt;p&gt;Enroll once (5 samples in the example), save the enrollment JSON, then gate both the wake word and STT live:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;import&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;createSpeakerVerificationMicController&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;from&lt;/span&gt; &lt;span class="s1"&gt;'react-native-wakeword'&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;ctrl&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;createSpeakerVerificationMicController&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'svMic1'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;ctrl&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;create&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nb"&gt;JSON&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;stringify&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;modelPath&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'sv.dm'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;options&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;decisionThreshold&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mf"&gt;0.35&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;ctrl&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;beginOnboarding&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'davoice'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="c1"&gt;// 5x: await ctrl.getNextEmbeddingFromMic(); -&amp;gt; progress events; onboardingDone carries enrollmentJson&lt;/span&gt;
&lt;span class="c1"&gt;// save the JSON (react-native-fs), destroy the controller, then enroll:&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;wakewordInstance&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;addSpeaker&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'owner'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;enrollmentJsonPath&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;addSpeaker&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'owner'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;enrollmentJsonPath&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="c1"&gt;// turn the filter off again:&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;wakewordInstance&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setSpeakers&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setSpeakers&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Each STT result then carries &lt;code&gt;speakerIds&lt;/code&gt; / &lt;code&gt;speakerId&lt;/code&gt;. The example uses it to authorize the &lt;code&gt;door_locks&lt;/code&gt; scope (&lt;code&gt;src/speaker_verification/currentSpeaker.ts&lt;/code&gt;):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;ownerMayAct&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;allowed&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;boolean&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="n"&gt;reason&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kt"&gt;string&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="n"&gt;gateOn&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;allowed&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;reason&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'speaker verification is off (demo allows it)'&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;ownerId&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="n"&gt;heard&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;has&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;ownerId&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;allowed&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;reason&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sb"&gt;`owner "${ownerId}" verified`&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;allowed&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;reason&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;heard&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;size&lt;/span&gt; &lt;span class="o"&gt;?&lt;/span&gt; &lt;span class="sb"&gt;`heard ${[...heard].join(', ')}, not the owner`&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'no verified speaker in this request'&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;In a car, that means a passenger or the radio can't unlock the doors. The enrollment JSON is personal voice data: keep it in the app's documents folder, never in logs. See also the &lt;a href="https://davoice.io/voice-agent-guide/speaker-verification" rel="noopener noreferrer"&gt;owner-only agent guide&lt;/a&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  Voice cloning
&lt;/h3&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setExternalVoicesPath&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;clonedVoicesPath&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;result&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;createVoiceClone&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
  &lt;span class="n"&gt;audioPath&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="sb"&gt;`${clonedVoicesPath()}/${name.trim()}.emd`&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="nf"&gt;getVoiceEmbedder&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;   &lt;span class="c1"&gt;// e.g. require('./assets/models/embedder_en.dm')&lt;/span&gt;
&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="c1"&gt;// every .emd in the folder becomes a voice: Speech.changeVoice(name)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If the result has &lt;code&gt;warnings&lt;/code&gt;, ask the user to record again. Switching voices inside one model is instant (&lt;code&gt;changeVoice&lt;/code&gt;); switching to another model file needs &lt;code&gt;Speech.destroyAll()&lt;/code&gt; and &lt;code&gt;initAll&lt;/code&gt; again.&lt;/p&gt;

&lt;h3&gt;
  
  
  Turn-taking and interruptions (barge-in)
&lt;/h3&gt;

&lt;p&gt;The example is &lt;strong&gt;turn-based by design&lt;/strong&gt;: STT is paused while the agent thinks and speaks, so the agent never transcribes its own voice. If you want "Hey Coach" to interrupt the reply, the repo documents the change (based on a production app):&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Keep the wake word listening while TTS plays: call &lt;code&gt;unPauseDetection()&lt;/code&gt; right after &lt;code&gt;pauseSpeechRecognition()&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Turn echo cancellation on: replace &lt;code&gt;WakewordAEC: { regular: false, duringTTS: false }&lt;/code&gt; with &lt;code&gt;AEC: { mode: 'always' }&lt;/code&gt; in &lt;code&gt;defaultAudioRoutingConfig&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;In the keyword callback, if the agent is speaking, stop everything and open a new turn:
&lt;/li&gt;
&lt;/ol&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;aiChatInFlightRef&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;current&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;             &lt;span class="c1"&gt;// the AI is thinking or speaking&lt;/span&gt;
  &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;interruptInFlightRef&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;current&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;  &lt;span class="c1"&gt;// the same "Hey Coach" detected twice&lt;/span&gt;
  &lt;span class="n"&gt;interruptInFlightRef&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;current&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;try&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="n"&gt;aiChatRequestIdRef&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;current&lt;/span&gt; &lt;span class="o"&gt;+=&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;         &lt;span class="c1"&gt;// late text and events of the old turn are ignored&lt;/span&gt;
    &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;getLocalAgent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;cancel&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;        &lt;span class="c1"&gt;// stops the LLM turn&lt;/span&gt;
    &lt;span class="n"&gt;aiChatSpeechQueueRef&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;current&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;       &lt;span class="c1"&gt;// drops the sentences not spoken yet&lt;/span&gt;
    &lt;span class="n"&gt;ignoreNextFinishedSpeakingRef&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;current&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="kc"&gt;true&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;stopSpeaking&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;             &lt;span class="c1"&gt;// silences the voice now&lt;/span&gt;
    &lt;span class="nf"&gt;resetSpeechTranscriptState&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;            &lt;span class="c1"&gt;// discards whatever STT caught meanwhile&lt;/span&gt;
    &lt;span class="n"&gt;aiChatInFlightRef&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;current&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;myInstanceRef&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;current&lt;/span&gt;&lt;span class="o"&gt;?&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;pauseDetection&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nc"&gt;Speech&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;unPauseSpeechRecognition&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="c1"&gt;// new user turn; preFetchMs 0&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="k"&gt;finally&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="n"&gt;interruptInFlightRef&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;current&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Open STT with &lt;code&gt;preFetchMs: 0&lt;/code&gt;, otherwise it replays the previous ~2 s of audio and transcribes "Hey Coach" as the new message. Combined with speaker verification, only the owner can interrupt. Details: &lt;a href="https://davoice.io/voice-agent-guide/turn-taking" rel="noopener noreferrer"&gt;turn-taking and interruptions guide&lt;/a&gt;.&lt;/p&gt;

&lt;h3&gt;
  
  
  Optional cloud layer (clearly not offline)
&lt;/h3&gt;

&lt;p&gt;Set &lt;code&gt;LOCAL_LLM_LAYER2&lt;/code&gt; in &lt;code&gt;src/local_llm/constants.ts&lt;/code&gt; to &lt;code&gt;'cloud'&lt;/code&gt; to hand general-knowledge questions to Gemini (&lt;code&gt;src/aichat/remote.ts&lt;/code&gt;), or implement &lt;code&gt;answerRemoteRequest&lt;/code&gt; for your own backend. This is an &lt;strong&gt;optional cloud integration&lt;/strong&gt; and incurs that provider's token costs. Keep keys out of the client in production (use a backend proxy).&lt;/p&gt;




&lt;h2&gt;
  
  
  9. Testing and troubleshooting
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Run the app tests:&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;npx jest __tests__/carTools.test.ts __tests__/fragments.test.ts
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Read the conversation log&lt;/strong&gt; (user words, model output, guard decisions, tool calls, timings):&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;adb shell run-as com.exampleapp cat files/davoice_chat.log   &lt;span class="c"&gt;# debug build&lt;/span&gt;
adb logcat &lt;span class="o"&gt;|&lt;/span&gt; grep DavoiceChat                                  &lt;span class="c"&gt;# release build&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;For every command check:&lt;/strong&gt; the right tool, exact arguments, a spoken sentence that matches what happened, a question when info is missing, a decline when unsupported, and &lt;strong&gt;no action&lt;/strong&gt; on small talk or speech noise.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Symptom&lt;/th&gt;
&lt;th&gt;Likely cause&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Gradle: &lt;code&gt;Could not find com.davoice:...&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;Missing &lt;code&gt;maven { url ".../libs" }&lt;/code&gt; line for that package&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Model "not found" at runtime&lt;/td&gt;
&lt;td&gt;File not in the bundle/assets under that exact name, or it's a Git LFS pointer (run &lt;code&gt;git lfs pull&lt;/code&gt;)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;
&lt;code&gt;setLicense&lt;/code&gt; returns &lt;code&gt;false&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;Expired or wrong key for this bundle id / package name&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;The app transcribes its own voice&lt;/td&gt;
&lt;td&gt;STT not paused during TTS, or AEC off&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Wake word fires twice&lt;/td&gt;
&lt;td&gt;Listener attached twice; use &lt;code&gt;attachKeywordListenerOnce&lt;/code&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;STT dead after the wake word&lt;/td&gt;
&lt;td&gt;Detection was stopped instead of paused, or STT never un-paused&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Gaps between spoken sentences&lt;/td&gt;
&lt;td&gt;Awaiting each &lt;code&gt;speak&lt;/code&gt;; enqueue without awaiting&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;iOS app killed while loading the LLM&lt;/td&gt;
&lt;td&gt;Missing increased-memory-limit entitlement&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;First reply very slow&lt;/td&gt;
&lt;td&gt;No warm-up/prefill, or no saved prompt state&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;"I need to send this to a remote Model"&lt;/td&gt;
&lt;td&gt;Expected: a general question with &lt;code&gt;LOCAL_LLM_LAYER2 = 'none'&lt;/code&gt;
&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;"Sorry, I didn't catch that" on short phrases&lt;/td&gt;
&lt;td&gt;Speech-noise filter (&lt;code&gt;src/aichat/fragments.ts&lt;/code&gt;) - expected for filler like "The."&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;




&lt;h2&gt;
  
  
  10. FAQ
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;What is a Voice-to-Voice AI Agent?&lt;/strong&gt;&lt;br&gt;
An assistant you talk to and that talks back, while taking actions through tools. Here it's a pipeline - wake word, STT, a local LLM with tool calling, TTS - not a single speech-to-speech model.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can it run fully offline?&lt;/strong&gt;&lt;br&gt;
The core loop can, once models are downloaded and for languages with an on-device STT model. Remote model hand-over, web search, MCP tools and the Gemini demo mode need a network and are optional.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Is it really token-free?&lt;/strong&gt;&lt;br&gt;
There are no cloud LLM inference token charges, because the model runs on the phone. The local model still processes tokens, and DaVoice SDKs require a commercial license for production.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How big is the model and how fast is it?&lt;/strong&gt;&lt;br&gt;
The DaVoice Agent 0.6B GGUF is about 380 MB, downloaded once. On a Galaxy S20 (CPU only) the decision takes about 1.3-2 s per command; iPhones with A14+ use the GPU.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can I use my own LLM?&lt;/strong&gt;&lt;br&gt;
Yes - the engine runs GGUF models with several chat templates (&lt;code&gt;qwen3&lt;/code&gt;, &lt;code&gt;chatml&lt;/code&gt;, &lt;code&gt;gemma&lt;/code&gt;, &lt;code&gt;llama3&lt;/code&gt;, &lt;code&gt;phi3&lt;/code&gt;). Larger general models are slower on phones.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Does it work with Expo?&lt;/strong&gt;&lt;br&gt;
Yes, with a development build or EAS Build (not Expo Go). See &lt;code&gt;expo/&lt;/code&gt;, &lt;code&gt;docs/EXPO_AI_AGENT_INTEGRATION_GUIDE.md&lt;/code&gt; and the &lt;a href="https://davoice.io/voice-agent-guide/expo" rel="noopener noreferrer"&gt;Expo guide&lt;/a&gt;. A &lt;a href="https://davoice.io/voice-agent-guide/flutter" rel="noopener noreferrer"&gt;Flutter version&lt;/a&gt; exists too.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Can users interrupt the agent?&lt;/strong&gt;&lt;br&gt;
The example is turn-based. Barge-in with the wake word plus echo cancellation is documented in section 8.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;How do I stop the model from doing dangerous things?&lt;/strong&gt;&lt;br&gt;
Declare &lt;code&gt;execution.authorization&lt;/code&gt; or &lt;code&gt;execution.confirmation&lt;/code&gt; on the tool, gate it in &lt;code&gt;onRequest&lt;/code&gt;, enable call guards (&lt;code&gt;callGuards: 'correct'&lt;/code&gt;), and let the app - not the model - confirm results via &lt;code&gt;toolReply&lt;/code&gt;.&lt;/p&gt;




&lt;h2&gt;
  
  
  11. Resources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;Example app (React Native + Expo): &lt;a href="https://github.com/frymanofer/ReactNative_DaVoice" rel="noopener noreferrer"&gt;github.com/frymanofer/ReactNative_DaVoice&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Integration how-to used in this article: &lt;a href="https://github.com/frymanofer/ReactNative_DaVoice/blob/main/example/englishHowTo.md" rel="noopener noreferrer"&gt;&lt;code&gt;example/englishHowTo.md&lt;/code&gt;&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Guide for AI coding agents: &lt;a href="https://github.com/frymanofer/ReactNative_DaVoice/blob/main/docs/AI_AGENT_INTEGRATION_GUIDE.md" rel="noopener noreferrer"&gt;&lt;code&gt;docs/AI_AGENT_INTEGRATION_GUIDE.md&lt;/code&gt;&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Voice-loop and tool code: &lt;a href="https://github.com/frymanofer/ReactNative_DaVoice/blob/main/example/src/aichat/provider.ts" rel="noopener noreferrer"&gt;&lt;code&gt;src/aichat/provider.ts&lt;/code&gt;&lt;/a&gt;, &lt;a href="https://github.com/frymanofer/ReactNative_DaVoice/blob/main/example/src/agent/LocalAgent.ts" rel="noopener noreferrer"&gt;&lt;code&gt;src/agent/LocalAgent.ts&lt;/code&gt;&lt;/a&gt;, &lt;a href="https://github.com/frymanofer/ReactNative_DaVoice/blob/main/example/src/car/carTools.ts" rel="noopener noreferrer"&gt;&lt;code&gt;src/car/carTools.ts&lt;/code&gt;&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;DaVoice Voice Agent Guide: &lt;a href="https://davoice.io/voice-agent-guide" rel="noopener noreferrer"&gt;davoice.io/voice-agent-guide&lt;/a&gt; and the &lt;a href="https://davoice.io/voice-agent-guide/react-native" rel="noopener noreferrer"&gt;React Native tutorial&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://davoice.io/voice-agent-sdk" rel="noopener noreferrer"&gt;Voice AI Agent SDK&lt;/a&gt; · &lt;a href="https://davoice.io/wake-word" rel="noopener noreferrer"&gt;Wake word&lt;/a&gt; · &lt;a href="https://davoice.io/speaker-recognition" rel="noopener noreferrer"&gt;Speaker recognition&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Try it in your browser: &lt;a href="https://davoice.io/playground" rel="noopener noreferrer"&gt;DaVoice Playground&lt;/a&gt; · &lt;a href="https://davoice.io/playground-app/#tts" rel="noopener noreferrer"&gt;Text-to-speech playground&lt;/a&gt; · &lt;a href="https://davoice.io/playground-app/#wakeword" rel="noopener noreferrer"&gt;Wake word playground&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Customer stories: &lt;a href="https://www.verbali.io/news#davoice-on-device" rel="noopener noreferrer"&gt;Verbali on going fully on-device&lt;/a&gt; · &lt;a href="https://davoice.io/use-case/automotive" rel="noopener noreferrer"&gt;Volkswagen OTTO&lt;/a&gt; · &lt;a href="https://www.youtube.com/@DaVoice_AI" rel="noopener noreferrer"&gt;DaVoice on YouTube&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;Licenses and custom wake words: &lt;code&gt;info@davoice.io&lt;/code&gt;
&lt;/li&gt;
&lt;/ul&gt;

</description>
      <category>reactnative</category>
      <category>ai</category>
      <category>mobile</category>
      <category>tutorial</category>
    </item>
  </channel>
</rss>
