[eric] voice: whisper's non-speech captions no longer inject as dictation (ENG-167)

electron/voice/streamingSession.js

@@ -16,6 +16,12 @@ const SAMPLE_RATE = 16000;
 const SILENT_KEEP_BYTES = SAMPLE_RATE * 2 * 2;
 const SILENT_TRIM_BYTES = SAMPLE_RATE * 2 * 10;
 
+// Whisper captions non-speech in brackets/parens ("[ Background sounds ]", "(laughs)"); strip them so
+// neither the live preview nor a committed phrase ever carries a caption instead of dictation.
+function stripSoundCaptions(text) {
+  return String(text || '').replace(/\[[^\]]*\]|\([^)]*\)|\*[^*]*\*|(?:^|\s)>>\s?/g, ' ').replace(/\s+/g, ' ').trim();
+}
+
 function wavFromPcm16(pcm) {
   const buf = Buffer.alloc(44 + pcm.length);
   buf.write('RIFF', 0); buf.writeUInt32LE(36 + pcm.length, 4); buf.write('WAVE', 8);
@@ -68,7 +74,7 @@ function createStreamingSession({ resourceDir, userDataDir, onPartial, previewIn
     inflight = decodePcm(pcm)
       .then((text) => {
         if (closedDown || epoch !== segEpoch) return; // the segment closed mid-decode; its final wins
-        tentative = text;
+        tentative = stripSoundCaptions(text);
         emit();
       })
       .catch(() => { skipNext = true; })
@@ -94,7 +100,7 @@ function createStreamingSession({ resourceDir, userDataDir, onPartial, previewIn
     commitChain = commitChain.then(async () => {
       if (inflight) await inflight;
       try {
-        const text = await decodePcm(pcm);
+        const text = stripSoundCaptions(await decodePcm(pcm));
         if (text) committed.push(text);
       } catch (_) {
         degraded = true; // a lost phrase final means the caller must fall back to the full-clip decode
@@ -137,4 +143,4 @@ function createStreamingSession({ resourceDir, userDataDir, onPartial, previewIn
   };
 }
 
-module.exports = { createStreamingSession, wavFromPcm16 };
+module.exports = { createStreamingSession, wavFromPcm16, stripSoundCaptions };

electron/voice/streamingVoice.test.js

@@ -2,7 +2,7 @@ const { test } = require('node:test');
 const assert = require('node:assert');
 const { createStreamSegmenter } = require('./streamSegmenter');
 const whisperService = require('./whisperService');
-const { createStreamingSession, wavFromPcm16 } = require('./streamingSession');
+const { createStreamingSession, wavFromPcm16, stripSoundCaptions } = require('./streamingSession');
 
 const RATE = 16000;
 
@@ -91,6 +91,24 @@ test('session: a failed segment decode marks the result degraded', async () => {
   assert.strictEqual(out.degraded, true);
 });
 
+test('stripSoundCaptions: captions and speaker marks go, words stay', () => {
+  assert.strictEqual(stripSoundCaptions('[ Background sounds ]'), '');
+  assert.strictEqual(stripSoundCaptions('[ Silence ] >> Hello world. [ Silence ]'), 'Hello world.');
+  assert.strictEqual(stripSoundCaptions('(laughs) okay *music* done'), 'okay done');
+  assert.strictEqual(stripSoundCaptions('plain dictated text'), 'plain dictated text');
+});
+
+test('session: a caption-only decode never becomes a committed phrase', async () => {
+  const restore = stubTranscribe(async () => '[ Background sounds ]');
+  const s = createStreamingSession({ resourceDir: '', userDataDir: '', onPartial: () => {}, previewIntervalMs: 3600000 });
+  s.pushChunk(asBuffer(tone(400)));
+  const out = await s.stop();
+  restore();
+  assert.strictEqual(out.ok, true);
+  assert.strictEqual(out.text, '');
+  assert.strictEqual(out.degraded, false);
+});
+
 test('session: chunks after cancel are dropped and stop reports stopped', async () => {
   const restore = stubTranscribe(async () => 'never');
   const s = createStreamingSession({ resourceDir: '', userDataDir: '', onPartial: () => {}, previewIntervalMs: 3600000 });

frontend/src/shared/voice/useVoiceDictation.ts

@@ -200,6 +200,8 @@ export function useVoiceDictation() {
           res = await window.openswarm?.voiceTranscribe?.(wav);
         }
       }
+      // Whisper captions non-speech in brackets/parens ("[ Background sounds ]", "(laughs)") and marks speaker turns with ">>"; those are annotations, not dictation.
+      if (res?.ok && res.text) res = { ok: true, text: res.text.replace(/\[[^\]]*\]|\([^)]*\)|\*[^*]*\*|(?:^|\s)>>\s?/g, ' ').replace(/\s+/g, ' ').trim() };
       // A transcript with no letter or digit in it is a hallucination artifact (the lone comma), not dictation.
       if (res?.ok && res.text && !/[\p{L}\p{N}]/u.test(res.text)) res = { ok: true, text: '' };
       if (res?.ok && res.text) {