The latest articles on DEV Community by Daming Wu (@wudaming00). https://dev.to/wudaming00 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%2F4078588%2F53c9602e-c5a3-41eb-bcaa-7cc0394bdce2.png https://dev.to/wudaming00 en Daming Wu Sat, 15 Aug 2026 07:01:38 +0000 https://dev.to/wudaming00/cpu-only-local-speech-to-text-for-driving-coding-agents-parakeet-tdt-paraformer-over-whisper-5d61 https://dev.to/wudaming00/cpu-only-local-speech-to-text-for-driving-coding-agents-parakeet-tdt-paraformer-over-whisper-5d61 <p>I dictate prompts to coding agents (Claude Code, Cursor) all day. Typing paragraph-long prompts was my bottleneck, and I wanted the speech side fully local - push-to-talk, CPU-only, nothing leaving the machine. The model choices surprised me, so here's the stack.</p> <h2> What I ended up with </h2> <p>Built in Rust on <a href="https://github.com/k2-fsa/sherpa-onnx" rel="noopener noreferrer">sherpa-onnx</a>:</p> <ul> <li> <strong>NVIDIA Parakeet TDT v3</strong> for European languages: one model covers 25 languages with automatic detection, and TDT decoding is noticeably lighter on CPU than Whisper-class encoder-decoder models in my use. It also emits its own punctuation.</li> <li> <strong>Paraformer (zh-en bilingual)</strong> for Mandarin: code-switching ("??? function ?? async") is where general multilingual models fall apart, and a dedicated bilingual model handles it far better. Punctuation comes from a separate local CT-Transformer pass.</li> <li>Models download on demand (a few hundred MB each); everything runs offline afterwards.</li> </ul> <h2> Why not Whisper </h2> <p>Nothing wrong with Whisper - but for live push-to-talk on CPU, the latency/quality point of TDT worked better for me, and Paraformer beat everything I tried on bilingual Mandarin. I deliberately don't publish latency numbers (too hardware-dependent to be honest about), so treat this as one person's experience, not a benchmark.</p> <h2> The part nobody talks about: insertion </h2> <p>Text <em>insertion</em> is harder than recognition. Getting clean text into whatever field has focus - terminal, editor, browser - across Windows and macOS is its own rabbit hole. The worst one: Windows silently stops delivering low-level keyboard hooks (WH_KEYBOARD_LL) when a Chromium-based window is focused. It's on record against Tauri and CEF too, and nothing you write inside the hook can fix it - the procedure is never entered. The fix is architectural: the focused window answers its own keydown events, and the global hook only owns the unfocused case. That one cost me a week.</p> <p>Acronym handling matters more than WER for this use case: "m c p" has to come out as MCP, and a user-editable replacement dictionary catches project jargon the model can't know.</p> <h2> Disclosure </h2> <p>This shipped as a small paid app - VocalCode, USD 4.99 one-time with a 30-day free trial, at <a href="https://vocalcode.app" rel="noopener noreferrer">vocalcode.app</a> - so I'm affiliated. The model discussion is the point of this post, though: if you're running TDT-family models for live dictation, or you've gotten good zh-en code-switching out of anything else local, I'd genuinely like to compare notes in the comments.</p> showdev rust ai productivity