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.

app/frontend/src/components/SpeechTranscriber.jsx

@@ -147,10 +147,16 @@ export default function SpeechTranscriber({
   const threads = speechSettings?.threads || 4;
   const backendPreference = speechSettings?.backendPreference || "auto";
   const translate = speechSettings?.translate === true;
+  const suppressNonSpeech = speechSettings?.suppressNonSpeech !== false;
+  const noSpeechThreshold = speechSettings?.noSpeechThreshold ?? 0.30;
+  const showAdvanced = speechSettings?.showAdvanced === true;
 
   const setLanguage = (val) => setSpeechSettings((prev) => ({ ...prev, language: val }));
   const setThreads = (val) => setSpeechSettings((prev) => ({ ...prev, threads: val }));
   const setBackendPreference = (val) => setSpeechSettings((prev) => ({ ...prev, backendPreference: val }));
+  const setSuppressNonSpeech = (val) => setSpeechSettings((prev) => ({ ...prev, suppressNonSpeech: val }));
+  const setNoSpeechThreshold = (val) => setSpeechSettings((prev) => ({ ...prev, noSpeechThreshold: val }));
+  const setShowAdvanced = (val) => setSpeechSettings((prev) => ({ ...prev, showAdvanced: val }));
 
   const [audioBlob, setAudioBlob] = useState(null);
   const [audioName, setAudioName] = useState("");
@@ -324,6 +330,8 @@ export default function SpeechTranscriber({
         backendPreference,
         translate,
         filename: audioName || "audio.wav",
+        suppressNonSpeech,
+        noSpeechThreshold,
         signal: controller.signal,
       });
       setResult(transcription);
@@ -352,54 +360,62 @@ export default function SpeechTranscriber({
     setIsTranscribing(false);
   };
 
-  const saveText = async () => {
-    if (!audioBlob) return;
-    await downloadTranscription("txt");
+  const saveText = () => {
+    if (!result) return;
+    downloadResultAsFile(result, "txt");
   };
 
-  const saveJson = async () => {
-    if (!audioBlob) return;
-    await downloadTranscription("json");
+  const saveJson = () => {
+    if (!result) return;
+    downloadResultAsFile(result, "json");
   };
 
-  const saveMarkdown = async () => {
-    if (!audioBlob) return;
-    await downloadTranscription("md");
+  const saveMarkdown = () => {
+    if (!result) return;
+    downloadResultAsFile(result, "md");
   };
 
-  const downloadTranscription = async (format) => {
-    if (!audioBlob) return;
+  const downloadResultAsFile = (transcription, format) => {
     try {
-      const params = new URLSearchParams();
-      if (selectedModel) params.set("model", selectedModel);
-      params.set("language", speechSettings?.language || "auto");
-      params.set("format", format);
-      params.set("threads", String(speechSettings?.threads || 4));
-      params.set("backendPreference", speechSettings?.backendPreference || "auto");
-      if (speechSettings?.translate) params.set("translate", "true");
-      params.set("filename", audioName || "audio.wav");
-      params.set("output_filename", `transcription-${Date.now()}`);
+      const baseName = (audioName || "transcription").replace(/\.[^.]+$/, "");
+      const stamp = new Date().toISOString().replace(/[:.]/g, "-").slice(0, 19);
+      let fileContent = "";
+      let contentType = "text/plain;charset=utf-8";
+      let extension = "txt";
 
-      const res = await fetch(`/api/speech/transcribe-file?${params.toString()}`, {
-        method: "POST",
-        headers: { "Content-Type": "audio/wav" },
-        body: audioBlob,
-      });
-
-      if (!res.ok) {
-        const errData = await res.json().catch(() => ({}));
-        throw new Error(errData.error || `Download failed (HTTP ${res.status})`);
+      if (format === "json") {
+        fileContent = JSON.stringify({
+          text: transcription?.text || "",
+          model: transcription?.model || selectedModel || "",
+          modelName: transcription?.modelName || "",
+          language: transcription?.language || speechSettings?.language || "auto",
+          backend: transcription?.backend || "",
+          backendMode: transcription?.backendMode || "",
+          sourceFilename: transcription?.sourceFilename || audioName || "",
+          durationMs: transcription?.durationMs || 0,
+          raw: transcription?.raw || {},
+          createdAt: transcription?.createdAt || new Date().toISOString(),
+        }, null, 2);
+        contentType = "application/json;charset=utf-8";
+        extension = "json";
+      } else if (format === "md") {
+        const modelName = transcription?.modelName || transcription?.model || selectedModel || "unknown";
+        const lang = transcription?.language || speechSettings?.language || "auto";
+        const backend = transcription?.backend || "unknown";
+        const backendMode = transcription?.backendMode || "";
+        const source = transcription?.sourceFilename || audioName || "unknown";
+        const durMs = transcription?.durationMs || 0;
+        const durSec = durMs ? (durMs / 1000).toFixed(1) : "?";
+        fileContent = `# Trascrizione\n\n${transcription?.text || ""}\n\n---\n*Modello: ${modelName}\n*Lingua: ${lang}\n*Backend: ${backend} (${backendMode})\n*File sorgente: ${source}\n*Durata elaborazione: ${durSec}s\n*Generato il: ${new Date().toISOString()}`;
+        contentType = "text/markdown;charset=utf-8";
+        extension = "md";
+      } else {
+        fileContent = transcription?.text || "";
+        contentType = "text/plain;charset=utf-8";
+        extension = "txt";
       }
 
-      const blob = await res.blob();
-      const url = URL.createObjectURL(blob);
-      const link = document.createElement("a");
-      link.href = url;
-      link.download = `transcription-${Date.now()}.${format}`;
-      document.body.appendChild(link);
-      link.click();
-      link.remove();
-      URL.revokeObjectURL(url);
+      downloadBlob(`${baseName}-${stamp}.${extension}`, fileContent, contentType);
     } catch (err) {
       showAlert?.({ title: "Download Failed", message: err.message || String(err), danger: true });
     }
@@ -437,6 +453,13 @@ export default function SpeechTranscriber({
             </select>
           </label>
 
+          {selectedModel && selectedModel.toLowerCase().includes("base") && (
+            <div style={{ marginTop: "6px", padding: "8px 10px", borderRadius: "6px", background: "var(--md-sys-color-error-container)", color: "var(--md-sys-color-on-error-container)", fontSize: "0.78rem", lineHeight: 1.4 }}>
+              ⚠️ Il modello <strong>Base</strong> può non trascrivere correttamente audio compresso (es. Google Meet, Zoom) — potrebbe rilevare solo "[Musica]".
+              Si consiglia il modello <strong>Small Multilingual</strong> per audio di bassa qualità.
+            </div>
+          )}
+
           <div className="speech-controls-row">
             <label className="speech-label">
               Backend
@@ -479,6 +502,45 @@ export default function SpeechTranscriber({
             </label>
           </div>
 
+          {/* Advanced settings toggle */}
+          <div style={{ marginTop: "10px", paddingLeft: "2px" }}>
+            <label style={{ display: "flex", alignItems: "center", gap: "8px", fontSize: "0.82rem", cursor: "pointer", color: "var(--md-sys-color-on-surface)" }}>
+              <input
+                type="checkbox"
+                checked={showAdvanced}
+                onChange={(event) => setShowAdvanced(event.target.checked)}
+                style={{ cursor: "pointer", width: "15px", height: "15px" }}
+              />
+              <span>Advanced options</span>
+            </label>
+          </div>
+
+          {showAdvanced && (
+            <div style={{ marginTop: "12px", padding: "12px", borderRadius: "8px", background: "var(--md-sys-color-surface-variant)", display: "flex", flexDirection: "column", gap: "10px" }}>
+              <label style={{ display: "flex", alignItems: "center", gap: "8px", fontSize: "0.82rem", cursor: "pointer", color: "var(--md-sys-color-on-surface)" }}>
+                <input
+                  type="checkbox"
+                  checked={suppressNonSpeech}
+                  onChange={(event) => setSuppressNonSpeech(event.target.checked)}
+                  style={{ cursor: "pointer", width: "15px", height: "15px" }}
+                />
+                <span>Suppress non-speech tokens (e.g. [Musica], [Applauso])</span>
+              </label>
+              <label style={{ display: "flex", flexDirection: "column", gap: "4px", fontSize: "0.82rem", color: "var(--md-sys-color-on-surface)" }}>
+                <span>No-speech threshold: <strong>{noSpeechThreshold.toFixed(2)}</strong> (lower = more sensitive)</span>
+                <input
+                  type="range"
+                  min="0.10"
+                  max="0.90"
+                  step="0.05"
+                  value={noSpeechThreshold}
+                  onChange={(event) => setNoSpeechThreshold(Number(event.target.value))}
+                  style={{ cursor: "pointer", width: "100%" }}
+                />
+              </label>
+            </div>
+          )}
+
           <div className="speech-button-row">
             <button
               className="m3-btn m3-btn-filled"

app/frontend/src/services/api.js

@@ -682,6 +682,8 @@ export async function transcribeSpeech(fileOrBlob, options = {}) {
   if (options.threads) params.set("threads", String(options.threads));
   if (options.backendPreference) params.set("backendPreference", options.backendPreference);
   if (options.translate) params.set("translate", "true");
+  if (options.suppressNonSpeech !== undefined) params.set("suppressNonSpeech", options.suppressNonSpeech ? "true" : "false");
+  if (options.noSpeechThreshold !== undefined) params.set("noSpeechThreshold", String(options.noSpeechThreshold));
 
   let contentType = "audio/wav";
   if (fileOrBlob instanceof File && fileOrBlob.name) {

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 = "";