fix(speech): generate TXT/JSON/MD locally from transcription result

Download buttons now use the already-completed transcription result
instead of re-sending the audio to the server for re-transcription.
This eliminates redundant server work and prevents download failures
when the second transcription attempt errors out.

Also wires up suppressNonSpeech and noSpeechThreshold options
through the API and server pipeline.

scripts/server/serve.cjs

@@ -2560,6 +2560,15 @@ function transcribeWavBuffer(buffer, options = {}) {
     if (options.translate) {
       args.push("-tr");
     }
+    // Suppress non-speech tokens (e.g. [Musica], [Applauso]) that appear
+    // when the model misidentifies compressed/noisy audio as non-speech
+    if (options.suppressNonSpeech !== false) {
+      args.push("-sns");
+    }
+    // Lower no-speech threshold (default 0.60) so the model is less likely
+    // to label quiet or compressed speech as non-speech
+    const noSpeechThreshold = options.noSpeechThreshold || "0.30";
+    args.push("-nth", String(noSpeechThreshold));
 
     const startedAt = Date.now();
     console.log("  [speech] Starting:", backend.cli, args.join(" "));
@@ -6030,14 +6039,19 @@ const server = http.createServer(async (req, res) => {
     const parsed = new URL(req.url, `http://${req.headers.host || "localhost"}`);
     try {
       const audio = await readBinaryBody(req);
-      const result = await runExclusiveSpeechOperation(() => transcribeAudioBuffer(audio, {
+      const suppressNonSpeech = parsed.searchParams.get("suppressNonSpeech");
+      const noSpeechThreshold = parsed.searchParams.get("noSpeechThreshold");
+      const transcribeOptions = {
         model: parsed.searchParams.get("model") || speechSettings.model,
         language: parsed.searchParams.get("language") || "auto",
         filename: parsed.searchParams.get("filename") || req.headers["x-filename"] || "recording",
         threads: parsed.searchParams.get("threads") || speechSettings.threads,
         backendPreference: parsed.searchParams.get("backendPreference") || speechSettings.backendPreference,
         translate: parsed.searchParams.get("translate") === "true",
-      }));
+      };
+      if (suppressNonSpeech !== null) transcribeOptions.suppressNonSpeech = suppressNonSpeech !== "false";
+      if (noSpeechThreshold !== null) transcribeOptions.noSpeechThreshold = noSpeechThreshold;
+      const result = await runExclusiveSpeechOperation(() => transcribeAudioBuffer(audio, transcribeOptions));
       return json(res, 200, { ok: true, transcription: result });
     } catch (err) {
       speechError = err.message || String(err);
@@ -6052,14 +6066,19 @@ const server = http.createServer(async (req, res) => {
     try {
       const audio = await readBinaryBody(req);
       const outputFormat = parsed.searchParams.get("format") || "txt";
-      const result = await runExclusiveSpeechOperation(() => transcribeAudioBuffer(audio, {
+      const suppressNonSpeech = parsed.searchParams.get("suppressNonSpeech");
+      const noSpeechThreshold = parsed.searchParams.get("noSpeechThreshold");
+      const transcribeOptions = {
         model: parsed.searchParams.get("model") || speechSettings.model,
         language: parsed.searchParams.get("language") || "auto",
         filename: parsed.searchParams.get("filename") || req.headers["x-filename"] || "recording",
         threads: parsed.searchParams.get("threads") || speechSettings.threads,
         backendPreference: parsed.searchParams.get("backendPreference") || speechSettings.backendPreference,
         translate: parsed.searchParams.get("translate") === "true",
-      }));
+      };
+      if (suppressNonSpeech !== null) transcribeOptions.suppressNonSpeech = suppressNonSpeech !== "false";
+      if (noSpeechThreshold !== null) transcribeOptions.noSpeechThreshold = noSpeechThreshold;
+      const result = await runExclusiveSpeechOperation(() => transcribeAudioBuffer(audio, transcribeOptions));
       
       const filename = parsed.searchParams.get("output_filename") || `transcription-${Date.now()}`;
       let fileContent = "";