<?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: 肖邦强</title>
    <description>The latest articles on DEV Community by 肖邦强 (@zxm9988).</description>
    <link>https://dev.to/zxm9988</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%2F385954%2F5e45c48c-d0c9-47be-bdb2-efc30b623a80.png</url>
      <title>DEV Community: 肖邦强</title>
      <link>https://dev.to/zxm9988</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/zxm9988"/>
    <language>en</language>
    <item>
      <title>I ship audio as plain MP3 files, not an app</title>
      <dc:creator>肖邦强</dc:creator>
      <pubDate>Mon, 16 Nov 2026 13:14:55 +0000</pubDate>
      <link>https://dev.to/zxm9988/i-ship-audio-as-plain-mp3-files-not-an-app-3c2o</link>
      <guid>https://dev.to/zxm9988/i-ship-audio-as-plain-mp3-files-not-an-app-3c2o</guid>
      <description>&lt;p&gt;I ship the output of my language tool as plain MP3 files, not as an app. Every other tool in this space wants to own playback — queue, streak, sync, recap — and I went the other way on purpose. Here is the reasoning, and the parts that actually cost me time.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. The player already exists
&lt;/h2&gt;

&lt;p&gt;A file plays on the phone in your pocket, the car stereo, a 12-year-old MP3 player, a cheap Bluetooth speaker, and &lt;code&gt;mpv&lt;/code&gt; on a laptop. That is the entire compatibility matrix, solved by not writing a player.&lt;/p&gt;

&lt;p&gt;The moment you build a player, you inherit: background audio permissions, media session controls, offline caching, sync conflicts, and a store review process. For a tool whose job is "turn two texts into one audio track", none of that is the interesting part.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Predictable filenames beat a library view
&lt;/h2&gt;

&lt;p&gt;The output for a chapter looks like this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;lesson-03/
  01-ja-en-0001.mp3
  01-ja-en-0002.mp3
  01-ja-en-0003.mp3
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Zero-padded, language-pair in the name, sentence index last. Cheap players sort lexicographically, so &lt;code&gt;1, 10, 11, 2&lt;/code&gt; is what you get without padding. Fixing that one detail removed the most common complaint I had with my own earlier attempts.&lt;/p&gt;

&lt;p&gt;Folder depth matters too. Some car head units only browse two levels. &lt;code&gt;Artist/Album/track&lt;/code&gt; is safe; &lt;code&gt;courses/japanese/season-1/chapter-3/lesson-4/track&lt;/code&gt; is not.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Pauses are the product
&lt;/h2&gt;

&lt;p&gt;Once sentence pairs exist, generating audio is the easy part. What makes a bilingual track unlistenable is silence: too much and you lose the sentence you just heard, too little and the transition feels like a cut.&lt;/p&gt;

&lt;p&gt;I settled on roughly 350 ms between the two languages of a pair and 900 ms between pairs. Those two numbers do more for comprehension than any voice-quality tuning I tried.&lt;/p&gt;

&lt;p&gt;Second lesson: &lt;strong&gt;do not normalize loudness per clip.&lt;/strong&gt; Normalizing each sentence independently makes the two languages pump against each other. Measure the pair, apply one gain to both.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Tagging is where the bugs live
&lt;/h2&gt;

&lt;p&gt;ID3v2.3 vs v2.4, UTF-16 vs Latin-1 for non-ASCII titles, players that ignore &lt;code&gt;track&lt;/code&gt; but honour the filename, players that do the reverse. I ended up writing only &lt;code&gt;title&lt;/code&gt;, &lt;code&gt;track&lt;/code&gt;, &lt;code&gt;album&lt;/code&gt;, and &lt;code&gt;artist&lt;/code&gt;, all in ID3v2.3, and letting filenames carry the real ordering. Boring, and it works on everything I've tested.&lt;/p&gt;

&lt;h2&gt;
  
  
  What this buys you
&lt;/h2&gt;

&lt;p&gt;Because the artifact is a file, the workflow stays inspectable. You can re-cut a chapter at a different pause length six months later without re-aligning anything, or drop the folder on a USB stick and hand it to someone. There is no account state to migrate.&lt;/p&gt;

&lt;p&gt;I wrote the boring details up separately — &lt;a href="https://bilingualmix.com/en/guides/import-to-listening-devices" rel="noopener noreferrer"&gt;file naming, folder depth and bitrate for bilingual MP3s on real devices&lt;/a&gt; — because that guide turned out to be more useful than any feature I shipped.&lt;/p&gt;

&lt;p&gt;The upstream half, pairing a Japanese audiobook with its English edition and getting sentence-level alignment out of two texts that never segment the same way, is here: &lt;a href="https://bilingualmix.com/en/guides/japanese-english-audiobook-listening" rel="noopener noreferrer"&gt;Japanese–English audiobook listening, line by line&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where it breaks
&lt;/h2&gt;

&lt;p&gt;Alignment on arbitrary prose. Dialogue-first matching covers maybe 70% of fiction; the long tail is a manual pass. Graded readers are fine, literary novels are not, and I have not found a general solution that doesn't involve a human reading the chapter.&lt;/p&gt;

&lt;p&gt;If you've built something in this space, I'd rather hear about your segmentation failures than your UI — that's the part I'm still losing to.&lt;/p&gt;

</description>
      <category>languagelearning</category>
      <category>podcast</category>
      <category>webdev</category>
      <category>showdev</category>
    </item>
    <item>
      <title>How I shipped a remote MCP server on Cloudflare Workers (and got it into the Registry)</title>
      <dc:creator>肖邦强</dc:creator>
      <pubDate>Sun, 08 Nov 2026 08:03:05 +0000</pubDate>
      <link>https://dev.to/zxm9988/how-i-shipped-a-remote-mcp-server-on-cloudflare-workers-and-got-it-into-the-registry-2lm7</link>
      <guid>https://dev.to/zxm9988/how-i-shipped-a-remote-mcp-server-on-cloudflare-workers-and-got-it-into-the-registry-2lm7</guid>
      <description>&lt;p&gt;I run &lt;a href="https://bilingualmix.com/en/mcp" rel="noopener noreferrer"&gt;Bilingual Mix&lt;/a&gt;, a small tool that turns parallel text into sentence-by-sentence alternating listening audio (one line in language A, then the same line in language B). I recently exposed it as a &lt;strong&gt;remote MCP server&lt;/strong&gt; so an AI client can build those tracks directly, and then published the endpoint to the &lt;a href="https://registry.modelcontextprotocol.io/" rel="noopener noreferrer"&gt;official MCP Registry&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;This is the checklist I wish I'd had: the five design decisions that actually mattered in production, followed by the exact manifest and commands that got the server listed. No SDK, no long-running process, no API key to hand out.&lt;/p&gt;

&lt;h2&gt;
  
  
  The shape of the server
&lt;/h2&gt;

&lt;p&gt;The endpoint is a single &lt;code&gt;POST https://bilingualmix.com/mcp&lt;/code&gt; — Streamable HTTP, stateless (no &lt;code&gt;Mcp-Session-Id&lt;/code&gt;), running on a Cloudflare Worker. It exposes five tools:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Tool&lt;/th&gt;
&lt;th&gt;What it does&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;create_bilingual_audio&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Sentence pairs → one alternating track, plus a per-sentence timeline&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;align_audio&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Two public audio URLs → transcribed, semantically aligned sentence pairs&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;translate_lines&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Fills in the missing language (members only)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;get_quota&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Sign-in state, minutes used/left, per-run limits&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;list_plans&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;Pricing and allowances&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;I hand-wrote the JSON-RPC dispatch instead of pulling in &lt;code&gt;@modelcontextprotocol/sdk&lt;/code&gt;. The Worker free tier has a 3 MB bundle budget, I only needed four methods (&lt;code&gt;initialize&lt;/code&gt;, &lt;code&gt;tools/list&lt;/code&gt;, &lt;code&gt;tools/call&lt;/code&gt;, &lt;code&gt;resources/list&lt;/code&gt;), and pinning protocol version &lt;code&gt;2025-06-18&lt;/code&gt; while still accepting &lt;code&gt;2024-11-05&lt;/code&gt; clients is easier when a dependency isn't quietly changing the wire format underneath you. The dispatcher is about 80 lines.&lt;/p&gt;

&lt;h2&gt;
  
  
  Five decisions that mattered
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Decision&lt;/th&gt;
&lt;th&gt;Reason&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;Never return audio in a tool result&lt;/td&gt;
&lt;td&gt;One minute of 24 kHz mono WAV is ~2.9 MB; base64 would blow up the model's context&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Count quota in D1, not KV&lt;/td&gt;
&lt;td&gt;KV is eventually consistent — I saw &lt;code&gt;401 → 200 → 401&lt;/code&gt; for the same IP&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Normalize IPs to &lt;code&gt;/64&lt;/code&gt;
&lt;/td&gt;
&lt;td&gt;IPv6 privacy extensions rotate addresses on every request&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Make the free→login seam an HTTP 401&lt;/td&gt;
&lt;td&gt;Clients discover OAuth from the challenge, not from a prose error&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Cap work per call at 30 pairs&lt;/td&gt;
&lt;td&gt;60 pairs pushed requests past the client's 60 s timeout&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;h3&gt;
  
  
  1. Return a URL and a timeline, never a blob
&lt;/h3&gt;

&lt;p&gt;A tool result is text that lands in a model's context. Audio does not belong there. So the tool stores each sentence's raw PCM in KV, returns a &lt;strong&gt;download URL valid for six hours&lt;/strong&gt; plus a timeline of &lt;code&gt;{a_start, a_end, b_start, b_end}&lt;/code&gt; for every pair. The timeline is the genuinely useful part: an agent can subtitle, cut clips, or write exercises from it without ever touching the bytes.&lt;/p&gt;

&lt;p&gt;The download route is a &lt;code&gt;ReadableStream&lt;/code&gt;: it writes the 44-byte WAV header first (the total PCM length was recorded when the job was created), then streams clips out of KV one pair at a time. Building the whole thing in a &lt;code&gt;Uint8Array&lt;/code&gt; first would fail — a 45-minute job is ~240 MB of PCM against a 128 MB Worker memory limit. Peak memory ends up being one clip.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Quotas need a strongly consistent counter
&lt;/h3&gt;

&lt;p&gt;My first version used KV for "runs today / minutes today". KV is eventually consistent across colos, so two consecutive requests from one IP could read different values. The symptom was infuriating: the third call was rejected, the fourth was allowed again.&lt;/p&gt;

&lt;p&gt;The fix is a D1 table with an atomic upsert:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight sql"&gt;&lt;code&gt;&lt;span class="k"&gt;INSERT&lt;/span&gt; &lt;span class="k"&gt;INTO&lt;/span&gt; &lt;span class="n"&gt;run_counters&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="k"&gt;scope&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;subject&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;site_id&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="k"&gt;day&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;runs&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;minutes&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;updated_at&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;VALUES&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="o"&gt;?&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="o"&gt;?&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="o"&gt;?&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nb"&gt;datetime&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'now'&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
&lt;span class="k"&gt;ON&lt;/span&gt; &lt;span class="n"&gt;CONFLICT&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="k"&gt;scope&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;subject&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;site_id&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="k"&gt;day&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;DO&lt;/span&gt; &lt;span class="k"&gt;UPDATE&lt;/span&gt; &lt;span class="k"&gt;SET&lt;/span&gt;
  &lt;span class="n"&gt;runs&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;runs&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;excluded&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;runs&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;minutes&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;minutes&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;excluded&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;minutes&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;updated_at&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nb"&gt;datetime&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s1"&gt;'now'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;scope&lt;/code&gt; is &lt;code&gt;ip&lt;/code&gt; or &lt;code&gt;user&lt;/code&gt;, &lt;code&gt;day&lt;/code&gt; is UTC, and a cron prunes rows older than seven days. The charge happens synchronously in the same call that generated the audio, so the next request sees it. Rule of thumb: &lt;strong&gt;if a user must immediately see their own last write, don't put it in KV.&lt;/strong&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  3. An IPv6 address is not a stable identity
&lt;/h3&gt;

&lt;p&gt;Anonymous quota is keyed by the caller's IP. If you use &lt;code&gt;cf-connecting-ip&lt;/code&gt; verbatim, IPv6 clients with privacy extensions present a fresh address on nearly every request — which means the free tier never runs out. Normalize to the &lt;code&gt;/64&lt;/code&gt; prefix, and unwrap IPv4-mapped addresses (&lt;code&gt;::ffff:192.0.2.1&lt;/code&gt; → &lt;code&gt;192.0.2.1&lt;/code&gt;) so dual-stack clients don't get a surprise third bucket:&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="k"&gt;export&lt;/span&gt; &lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;ipKey&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;raw&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="nx"&gt;ip&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;String&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;raw&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="dl"&gt;""&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;trim&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="nx"&gt;ip&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;unknown&lt;/span&gt;&lt;span class="dl"&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="nx"&gt;ip&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;includes&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;:&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nx"&gt;ip&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;mapped&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="sr"&gt;/^::ffff:&lt;/span&gt;&lt;span class="se"&gt;(\d&lt;/span&gt;&lt;span class="sr"&gt;+&lt;/span&gt;&lt;span class="se"&gt;\.\d&lt;/span&gt;&lt;span class="sr"&gt;+&lt;/span&gt;&lt;span class="se"&gt;\.\d&lt;/span&gt;&lt;span class="sr"&gt;+&lt;/span&gt;&lt;span class="se"&gt;\.\d&lt;/span&gt;&lt;span class="sr"&gt;+&lt;/span&gt;&lt;span class="se"&gt;)&lt;/span&gt;&lt;span class="sr"&gt;$/i&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;exec&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;ip&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="nx"&gt;mapped&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nx"&gt;mapped&lt;/span&gt;&lt;span class="p"&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;// ...expand the address, zero-pad to 8 groups, keep the first 4&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nx"&gt;groups&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;slice&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="mi"&gt;4&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;join&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;:&lt;/span&gt;&lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="dl"&gt;"&lt;/span&gt;&lt;span class="s2"&gt;::/64&lt;/span&gt;&lt;span class="dl"&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;Known limitation: a dual-stack client can still hold one IPv4 bucket and one IPv6 bucket. You can't fully merge those without accounting for shared NAT exits, but &lt;code&gt;/64&lt;/code&gt; normalization makes the v6 side stable.&lt;/p&gt;

&lt;h3&gt;
  
  
  4. Let the 401 start the OAuth flow
&lt;/h3&gt;

&lt;p&gt;Anonymous visitors get 2 runs / 20 minutes a day (≤10 minutes and ≤8 pairs per run). A free account gets 4 runs / 40 minutes; members get 500 minutes a month with ≤45 minutes and ≤30 pairs per run. When the free allowance is spent, the server replies:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight http"&gt;&lt;code&gt;&lt;span class="k"&gt;HTTP&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="m"&gt;1.1&lt;/span&gt; &lt;span class="m"&gt;401&lt;/span&gt;
&lt;span class="na"&gt;WWW-Authenticate&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s"&gt;Bearer resource_metadata="https://bilingualmix.com/.well-known/oauth-protected-resource"&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;MCP clients use that challenge to discover the authorization server, register dynamically, and open a real consent screen. The auth is OAuth 2.1 with authorization code + PKCE (S256 only), one-hour access tokens, and no API key or password ever reaching the agent.&lt;/p&gt;

&lt;p&gt;Two details that saved me debugging time:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;initialize&lt;/code&gt; and &lt;code&gt;tools/list&lt;/code&gt; must never 401.&lt;/strong&gt; Only a &lt;code&gt;tools/call&lt;/code&gt; that actually needs quota may challenge. If you authenticate discovery, half the clients fail before they can show your tools.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Give the agent a free preview.&lt;/strong&gt; &lt;code&gt;create_bilingual_audio&lt;/code&gt; accepts &lt;code&gt;dry_run: true&lt;/code&gt;, which returns the pair list, estimated minutes, and quota cost without generating anything or spending quota. The agent can ask "is this okay?" before committing. That single flag makes cost conversations possible.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  5. Bound each call
&lt;/h3&gt;

&lt;p&gt;Remote MCP calls sit behind client-side timeouts. Sixty sentence pairs pushed requests past the client's 60 s limit; thirty does not. Cap the input, error with a clear "split this into batches" message, and make the limits visible through &lt;code&gt;get_quota&lt;/code&gt; so the agent can plan. Long content is a loop of small calls, not one heroic request.&lt;/p&gt;

&lt;h2&gt;
  
  
  Try it
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;claude mcp add &lt;span class="nt"&gt;--transport&lt;/span&gt; http bilingual-mix https://bilingualmix.com/mcp
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Or add this to your client's MCP config:&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="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"mcpServers"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"bilingual-mix"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"http"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://bilingualmix.com/mcp"&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then just ask: &lt;em&gt;"Here are 12 lines of Japanese with English translations. First do a dry run to show the cost, then generate the audio and give me the link and timeline."&lt;/em&gt; Clients that only accept local stdio servers need an HTTP-to-stdio bridge such as &lt;code&gt;mcp-remote&lt;/code&gt;; nothing on the server side changes. The full tool list, limits, and discovery endpoints are also in &lt;a href="https://bilingualmix.com/llms.txt" rel="noopener noreferrer"&gt;llms.txt&lt;/a&gt;, and the &lt;a href="https://bilingualmix.com/en/guides/make-bilingual-audio" rel="noopener noreferrer"&gt;walkthrough&lt;/a&gt; shows what the audio is for.&lt;/p&gt;

&lt;h2&gt;
  
  
  Publishing to the official MCP Registry
&lt;/h2&gt;

&lt;p&gt;A Registry record is one file: &lt;code&gt;server.json&lt;/code&gt; at the root of a &lt;strong&gt;public&lt;/strong&gt; repository. Directories that mirror the Registry pick it up on their own schedule, so this is a publish-once move rather than a dozen submission forms. Here is the exact file that passed validation for my server (also in the &lt;a href="https://github.com/zxm9988/bilingual-mix" rel="noopener noreferrer"&gt;public repo&lt;/a&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="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"$schema"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://static.modelcontextprotocol.io/schemas/2025-12-11/server.schema.json"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"name"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"io.github.zxm9988/bilingual-mix"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"description"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"Remote MCP server for sentence-by-sentence alternating bilingual listening audio."&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"version"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"1.0.0"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"websiteUrl"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://bilingualmix.com/en/mcp"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"repository"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://github.com/zxm9988/bilingual-mix"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="nl"&gt;"source"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"github"&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;},&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="nl"&gt;"remotes"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"type"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"streamable-http"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;&lt;span class="w"&gt;
      &lt;/span&gt;&lt;span class="nl"&gt;"url"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="s2"&gt;"https://bilingualmix.com/mcp"&lt;/span&gt;&lt;span class="w"&gt;
    &lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
  &lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="w"&gt;
&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Things worth knowing before you run the CLI:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;name&lt;/code&gt; is reverse-DNS. With GitHub authentication it &lt;strong&gt;must&lt;/strong&gt; start with &lt;code&gt;io.github.&amp;lt;your-username&amp;gt;/&lt;/code&gt;; DNS-based auth unlocks domain-based namespaces instead.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;description&lt;/code&gt; is capped at &lt;strong&gt;100 characters&lt;/strong&gt;. My first 175-character version got a &lt;code&gt;422&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;A hosted server needs only &lt;code&gt;remotes&lt;/code&gt;. You do &lt;strong&gt;not&lt;/strong&gt; need a &lt;code&gt;packages&lt;/code&gt; entry or an npm/Docker release if nobody installs your server locally.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Install the publisher (the releases page has builds for every platform; this one is macOS arm64), then validate, log in, and publish:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;curl &lt;span class="nt"&gt;-fsSL&lt;/span&gt; &lt;span class="s2"&gt;"https://github.com/modelcontextprotocol/registry/releases/latest/download/mcp-publisher_darwin_arm64.tar.gz"&lt;/span&gt; | &lt;span class="nb"&gt;tar &lt;/span&gt;xz mcp-publisher

./mcp-publisher validate server.json
./mcp-publisher login github      &lt;span class="c"&gt;# device flow: open the URL, paste the code&lt;/span&gt;
./mcp-publisher publish server.json
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;login github&lt;/code&gt; uses a device code and grants only the &lt;code&gt;io.github.&amp;lt;you&amp;gt;/*&lt;/code&gt; namespace. &lt;strong&gt;Don't run &lt;code&gt;logout&lt;/code&gt; just to check your auth state&lt;/strong&gt; — it clears the credential you just obtained and you'll have to redo the device flow.&lt;/p&gt;

&lt;h3&gt;
  
  
  Verify by exact version, not by search
&lt;/h3&gt;

&lt;p&gt;A successful CLI exit is not proof of a readable record. Search has a delay, and the endpoint that actually returns your metadata is the versioned one:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# exact read-back (name URL-encoded)&lt;/span&gt;
curl &lt;span class="nt"&gt;-s&lt;/span&gt; &lt;span class="s1"&gt;'https://registry.modelcontextprotocol.io/v0.1/servers/io.github.zxm9988%2Fbilingual-mix/versions/1.0.0'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  | jq &lt;span class="s1"&gt;'.server | {name, version, remotes}'&lt;/span&gt;

&lt;span class="c"&gt;# search returns records nested under .server&lt;/span&gt;
curl &lt;span class="nt"&gt;-s&lt;/span&gt; &lt;span class="s1"&gt;'https://registry.modelcontextprotocol.io/v0.1/servers?search=bilingual-mix'&lt;/span&gt; &lt;span class="se"&gt;\&lt;/span&gt;
  | jq &lt;span class="s1"&gt;'.servers[].server.name'&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Three gotchas I hit:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;code&gt;/v0.1/servers/&amp;lt;name&amp;gt;&lt;/code&gt; &lt;strong&gt;without&lt;/strong&gt; &lt;code&gt;/versions/&amp;lt;version&amp;gt;&lt;/code&gt; returns 404. That is not "not published".&lt;/li&gt;
&lt;li&gt;Search results nest the payload under &lt;code&gt;server&lt;/code&gt;, so &lt;code&gt;.servers[].name&lt;/code&gt; prints &lt;code&gt;null&lt;/code&gt; and looks like an empty result.&lt;/li&gt;
&lt;li&gt;Search tokens matter: &lt;code&gt;search=bilingual-mix&lt;/code&gt; finds my record, &lt;code&gt;search=bilingualmix&lt;/code&gt; returns zero. Try the full hyphenated name before concluding anything failed.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Finally, a Registry record is metadata, not a compatibility certification: it doesn't prove every client handles your OAuth or your tool schemas. And when you change published metadata, bump &lt;code&gt;version&lt;/code&gt; rather than re-publishing the same one.&lt;/p&gt;

&lt;h2&gt;
  
  
  Resources
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;a href="https://bilingualmix.com/en/mcp" rel="noopener noreferrer"&gt;Bilingual Mix MCP docs&lt;/a&gt; — tools, quota, client config&lt;/li&gt;
&lt;li&gt;
&lt;a href="https://bilingualmix.com/llms.txt" rel="noopener noreferrer"&gt;llms.txt&lt;/a&gt; — machine-readable summary&lt;/li&gt;
&lt;li&gt;&lt;a href="https://github.com/zxm9988/bilingual-mix" rel="noopener noreferrer"&gt;Public repo with server.json&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://registry.modelcontextprotocol.io/v0.1/servers/io.github.zxm9988%2Fbilingual-mix/versions/1.0.0" rel="noopener noreferrer"&gt;Registry record (v1.0.0)&lt;/a&gt;&lt;/li&gt;
&lt;li&gt;&lt;a href="https://bilingualmix.com/en/guides/make-bilingual-audio" rel="noopener noreferrer"&gt;How the alternating audio is used&lt;/a&gt;&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;If you're building a remote MCP server that returns anything heavier than text, the short version is: stream it from a URL, keep counters strongly consistent, challenge with a proper 401, and give the agent a dry run. Everything else is polish.&lt;/p&gt;

</description>
      <category>mcp</category>
      <category>tutorial</category>
      <category>serverless</category>
      <category>ai</category>
    </item>
    <item>
      <title>Building sentence-level alignment for bilingual audio</title>
      <dc:creator>肖邦强</dc:creator>
      <pubDate>Sat, 24 Oct 2026 03:03:39 +0000</pubDate>
      <link>https://dev.to/zxm9988/building-sentence-level-alignment-for-bilingual-audio-52jn</link>
      <guid>https://dev.to/zxm9988/building-sentence-level-alignment-for-bilingual-audio-52jn</guid>
      <description>&lt;p&gt;Most language-learning apps assume a text-first workflow: paste a paragraph, translate it, then generate audio. A harder and more useful case starts with two real recordings: the same story, lesson, or podcast delivered once in English and once in Mandarin or Japanese. The sentence boundaries do not match. The speaking speeds do not match. One track may have intro music, extra narration, or a different cut.&lt;/p&gt;

&lt;p&gt;The goal for Bilingual Mix is to turn those two files into one alternating MP3: meaning-language sentence first, then the same idea in the target language. This post is about the alignment and rendering pipeline behind that output, and the tradeoffs that made it reliable enough to be usable.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Transcribe with timestamps, not just text
&lt;/h2&gt;

&lt;p&gt;The server sends each file to Groq's Whisper-large-v3-turbo endpoint and requests both word and segment timestamps. Segment timestamps are useful as a first pass, but word timings matter because they let the later stages trim silence and verify that a repaired boundary still contains the words it claims to contain.&lt;/p&gt;

&lt;p&gt;A transcript with text but no timing is not enough. The product needs four numbers for every candidate pair: start and end on track A, and start and end on track B.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Align short speech units, not whole subtitles
&lt;/h2&gt;

&lt;p&gt;The original version used dynamic programming over entire ASR segments. It worked for clean, similarly paced recordings, but failed when one language split a sentence differently. A long segment on one side could correspond to two short segments on the other, and a single wrong grouping could push the rest of the path out of sync.&lt;/p&gt;

&lt;p&gt;The current approach first turns each transcript into short speech units. Units are merged only within strict limits: about 6.5 seconds per unit, three units per candidate group, and an eight-second group maximum. Explicit noise utterances such as “um” or “嗯” are removed. This makes the alignment problem smaller without pretending that every ASR segment is a sentence.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Use multilingual embeddings, but keep time in the loop
&lt;/h2&gt;

&lt;p&gt;Text alone is ambiguous. The same phrase can occur several times, and different dubs can describe the same moment with different wording. The service embeds candidate windows with Cloudflare Workers AI's multilingual BGE-M3 model, then compares both the original timestamps and normalized progress through each track.&lt;/p&gt;

&lt;p&gt;The time distance is deliberately not just “closest timestamp wins.” Absolute time is useful when the tracks are tightly synchronized. Normalized progress is useful when one version runs consistently faster. The implementation evaluates both and keeps a candidate only when it is inside a time band of roughly 22–45 seconds.&lt;/p&gt;

&lt;p&gt;There is also a minimum semantic similarity threshold. It becomes stricter as the time gap grows and is slightly stricter for one- or two-word utterances, where a chance embedding match is more likely. When a candidate merges several units, every individual unit must support the match; a strong phrase cannot carry a weak neighbor along with it.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. Make the path monotonic
&lt;/h2&gt;

&lt;p&gt;After candidates are generated, a dynamic-programming pass selects a monotonic path through both transcripts. A match advances both sides. Unmatched speech can be skipped on either side. That prevents the most damaging error mode: jumping backward or across large unrelated regions just because an embedding pair looked similar.&lt;/p&gt;

&lt;p&gt;The selector also checks ambiguity. If two nearly equal candidates compete for one side of the same unit, the pair is treated as risky and can be excluded from the default result. That is important for repeated dialogue, refrains, and generic greetings.&lt;/p&gt;

&lt;h2&gt;
  
  
  5. Publish a conservative quality signal
&lt;/h2&gt;

&lt;p&gt;The UI does not claim to show translation accuracy. It shows a conservative alignment score: the minimum of the mean semantic similarity and the duration coverage on either track. If one language loses a large section, the score drops even when the remaining pairs look confident.&lt;/p&gt;

&lt;p&gt;The user still gets a review screen. They can listen to each original clip, inspect the proposed text, and uncheck bad pairs before any audio is generated. Automatic alignment is a draft, not a promise that every semantic match is correct.&lt;/p&gt;

&lt;h2&gt;
  
  
  6. Render locally with Web Audio and ffmpeg.wasm
&lt;/h2&gt;

&lt;p&gt;Once the user confirms the pairs, the browser decodes both source files with Web Audio. The confirmed time windows are converted to PCM WAV clips, grouped into batches of about 12 pairs, encoded to 192 kbps MP3 with ffmpeg.wasm, and concatenated. The final concat is a stream copy because all batches use the same encoder settings.&lt;/p&gt;

&lt;p&gt;This keeps the expensive audio work in the browser and makes it possible to generate the finished track without uploading a second copy of the audio for rendering. The server does the transcription and semantic alignment; the client performs the final cut and join.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would improve next
&lt;/h2&gt;

&lt;p&gt;Two areas deserve more work. The first is better handling of overlapping speech and voice-over. The second is a more calibrated confidence model; the current score is deliberately conservative rather than a statistically calibrated probability.&lt;/p&gt;

&lt;p&gt;For a focused language tool, that tradeoff is acceptable: a smaller result with visible coverage gaps is more useful than a longer track that confidently pairs the wrong sentences. The complete workflow, including free trial limits and sample tracks, is at &lt;a href="https://bilingualmix.com/en/" rel="noopener noreferrer"&gt;Bilingual Mix&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;If you are building anything with ASR timestamps, keep both absolute time and relative progress, make the matching path monotonic, and surface uncertainty before rendering. Those three decisions prevented a surprising number of silent failures.&lt;/p&gt;

</description>
      <category>javascript</category>
      <category>ai</category>
      <category>webdev</category>
      <category>showdev</category>
    </item>
  </channel>
</rss>
