fix(voice): browser speech recognition on Android no longer repeats phrases

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

frontend/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "mindsift-frontend",
-  "version": "0.22.1",
+  "version": "0.22.2",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "mindsift-frontend",
-      "version": "0.22.1",
+      "version": "0.22.2",
       "license": "Unlicense",
       "dependencies": {
         "@fontsource-variable/inter": "5.3.0",

frontend/package.json

@@ -1,7 +1,7 @@
 {
   "name": "mindsift-frontend",
   "private": true,
-  "version": "0.22.1",
+  "version": "0.22.2",
   "license": "Unlicense",
   "type": "module",
   "scripts": {

frontend/src/lib/__tests__/voice.join.test.ts

@@ -0,0 +1,21 @@
+import { describe, it, expect } from 'vitest';
+import { joinSpeechResults } from '../voice';
+
+describe('joinSpeechResults (browser speech recognition)', () => {
+  it('Android: growing interim states of one phrase count once', () => {
+    expect(joinSpeechResults(['Test', 'Test', 'Test', 'Test', 'Test', 'Test 1 2 3'])).toBe('Test 1 2 3');
+    expect(joinSpeechResults(['Test', 'Test 1', 'Test 1 2', 'Test 1 2 3'])).toBe('Test 1 2 3');
+  });
+
+  it('desktop: consecutive pieces are joined', () => {
+    expect(joinSpeechResults(['Milch kaufen', ' und Brot'])).toBe('Milch kaufen und Brot');
+  });
+
+  it('Android: a new phrase after a finished one is appended', () => {
+    expect(joinSpeechResults(['Milch', 'Milch kaufen', ' und Brot', ' und Brot holen'])).toBe('Milch kaufen und Brot holen');
+  });
+
+  it('ignores empty results and earlier shorter states', () => {
+    expect(joinSpeechResults(['', 'Test 1 2', 'Test 1', ''])).toBe('Test 1 2');
+  });
+});

frontend/src/lib/useVoiceInput.ts

@@ -1,6 +1,6 @@
 import { useCallback, useEffect, useRef, useState } from 'react';
 import { api } from './api';
-import { browserRecognition, currentVoiceConfig, EffectiveVoiceMode, loadVoiceConfig, VoiceRecorder } from './voice';
+import { browserRecognition, currentVoiceConfig, EffectiveVoiceMode, joinSpeechResults, loadVoiceConfig, VoiceRecorder } from './voice';
 
 export type VoiceState = 'idle' | 'listening' | 'transcribing';
 
@@ -234,8 +234,9 @@ export function useVoiceInput({ onText, onError, continuous = false, live = fals
     recognition.interimResults = true;
     recognition.continuous = continuous || live;
     recognition.onresult = (event: any) => {
-      let sessionText = '';
-      for (let i = 0; i < event.results.length; i++) sessionText += event.results[i][0].transcript;
+      const parts: string[] = [];
+      for (let i = 0; i < event.results.length; i++) parts.push(event.results[i][0].transcript);
+      const sessionText = joinSpeechResults(parts);
       if (live && baseTranscriptRef.current) {
         const base = baseTranscriptRef.current.trimEnd();
         const next = sessionText.trimStart();

frontend/src/lib/voice.ts

@@ -186,3 +186,27 @@ export class VoiceRecorder {
     this.analyser = null;
   }
 }
+
+/**
+ * Text of one speech recognition session from its results. Desktop browsers deliver consecutive
+ * pieces ("Test 1 2 3", " und weiter"); Chrome on Android also keeps the growing interim states of
+ * the same phrase as results of their own ("Test", "Test", "Test 1 2 3"). A result that only
+ * continues the one before it therefore replaces it instead of being appended.
+ */
+export function joinSpeechResults(transcripts: string[]): string {
+  const norm = (s: string) => s.trim().toLowerCase();
+  let text = '';
+  let last = '';
+  for (const t of transcripts) {
+    if (!t) continue;
+    if (last && norm(t).startsWith(norm(last))) {
+      text = text.slice(0, text.length - last.length) + t; // grown version of the same phrase
+    } else if (last && norm(last).startsWith(norm(t))) {
+      continue; // an earlier, shorter state of it
+    } else {
+      text += t;
+    }
+    last = t;
+  }
+  return text;
+}