feat(voice): all microphone buttons use the server's speech recognition

- New useVoiceInput hook (lib/voice.ts): "server" records with MediaRecorder
  and sends to /voice/transcribe, "browser" keeps the live recognition,
  "off" or missing browser support hides the microphone.
- Quick dump bar, mobile capture sheet and hold-to-record use it; live
  waves from the real input level (VoiceWave) and "Wird erkannt…".
- Recordings stop after STT_MAX_SECONDS.

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

CHANGELOG.md

@@ -10,6 +10,9 @@ Das Format folgt [Keep a Changelog](https://keepachangelog.com/de/1.1.0/), die V
 - Eigene Spracherkennung (optional): Mit `STT_URL` schickt MindSift Aufnahmen an einen selbst betriebenen Whisper-Server (OpenAI-kompatibel, z. B. speaches) statt an Google oder Microsoft. Einstellung nur per Umgebungsvariablen (`STT_*`) inklusive Grenzwerten; `docker-compose.yml` bringt den Dienst `whisper` als Profil mit, das Modell lädt MindSift beim ersten Start selbst. Administration → Server & Isolation zeigt den aktiven Modus und bietet eine Testaufnahme. In der Demo bleibt es bei der Browser-Erkennung.
 - API: `POST /voice/transcribe`, `GET /admin/voice`; `/auth/status` meldet `voice.mode`.
 
+### Changed
+- Spracheingabe in Quick Dump, im mobilen Erfassen-Sheet und beim Halten des +-Knopfs folgt der Server-Einstellung: mit eigenem Whisper-Server funktioniert sie in allen Browsern (auch Firefox), mit Live-Wellen während der Aufnahme und „Wird erkannt…“ danach. Ohne Unterstützung im Browser oder mit `STT_DISABLED=true` ist das Mikrofon ausgeblendet statt einer Fehlermeldung.
+
 ## [0.6.1] - 2026-10-29
 
 ### Fixed

Doku/architecture.md

@@ -92,7 +92,7 @@ Hinweis: `DATABASE_URL` mit `postgres…` führt dazu, dass der Standardpfad in
 
 ## Spracherkennung
 
-Nur über Umgebungsvariablen (`STT_*`, siehe oben), nicht in der Oberfläche; Administration → Server & Isolation zeigt den aktiven Modus, den Zustand des Whisper-Servers und bietet eine Testaufnahme. Im Server-Modus nimmt der Browser per `MediaRecorder` auf und schickt die Aufnahme an `POST /voice/transcribe`; MindSift reicht sie an `STT_URL/v1/audio/transcriptions` weiter (`model`, `language`) und gibt nur den Text zurück. Nichts wird gespeichert oder protokolliert außer Fehlern (`internal/api/voice.go`). Beim Start wartet MindSift auf den Whisper-Server und stößt den Download des Modells an, falls es fehlt (`POST /v1/models/<id>` bei speaches).
+Nur über Umgebungsvariablen (`STT_*`, siehe oben), nicht in der Oberfläche; Administration → Server & Isolation zeigt den aktiven Modus, den Zustand des Whisper-Servers und bietet eine Testaufnahme. Alle Mikrofon-Knöpfe (Quick-Dump-Leiste, mobiles Erfassen-Sheet, Halten des +-Knopfs) nutzen denselben Hook `useVoiceInput` (`frontend/src/lib/voice.ts`, `useVoiceInput.ts`): `server` nimmt auf und zeigt „Wird erkannt…“, `browser` liefert Text schon beim Sprechen, `off` oder ein Browser ohne die nötige Funktion blendet das Mikrofon aus. Im Server-Modus zeigen Wellen den echten Mikrofonpegel (`VoiceWave`), die Aufnahme endet spätestens nach `STT_MAX_SECONDS`. Der Browser nimmt per `MediaRecorder` auf (WebM/Opus, in Safari MP4/AAC) und schickt die Aufnahme an `POST /voice/transcribe`; MindSift reicht sie an `STT_URL/v1/audio/transcriptions` weiter (`model`, `language`) und gibt nur den Text zurück. Nichts wird gespeichert oder protokolliert außer Fehlern (`internal/api/voice.go`). Beim Start wartet MindSift auf den Whisper-Server und stößt den Download des Modells an, falls es fehlt (`POST /v1/models/<id>` bei speaches).
 
 Docker Compose bringt den optionalen Dienst `whisper` (speaches 0.8.3, CPU, Profil `whisper`) mit; das Modell liegt in `./whisper` (Ordner gehört UID 1000). Gemessen auf einem Desktop mit 16 Kernen (int8) für einen Satz von 4 s: `Systran/faster-whisper-small` 1,6 s, `deepdml/faster-whisper-large-v3-turbo-ct2` 8 s bei besserer Erkennung (RAM ca. 1 bzw. 2 GB). Auf geteilten vCPUs ist mit dem Mehrfachen zu rechnen.
 

frontend/src/App.tsx

@@ -23,6 +23,7 @@ import {
   Flame,
   Mic,
   MicOff,
+  Loader2,
   Search,
 } from 'lucide-react';
 import { AuthStatus, DashboardStats, PriorityLevel, SharedList, Thought, User, UserTheme } from './lib/types';
@@ -35,6 +36,8 @@ import {
   disconnectWebSocket,
   isSharingDisabled,
 } from './lib/api';
+import { useVoiceInput } from './lib/useVoiceInput';
+import { VoiceWave } from './components/VoiceWave';
 import {
   AVAILABLE_THEMES,
   customThemeOption,
@@ -175,49 +178,31 @@ export const App: React.FC = () => {
   const [mobileTriageContent, setMobileTriageContent] = useState('');
   const [mobileTriagePriority, setMobileTriagePriority] = useState<PriorityLevel>('normal');
   const [mobileTriageDueDate, setMobileTriageDueDate] = useState('');
-  const [mobileVoiceActive, setMobileVoiceActive] = useState(false);
   const [isVoiceCaptured, setIsVoiceCaptured] = useState(false);
-  const mobileRecognitionRef = useRef<any>(null);
+  const [mobileVoiceError, setMobileVoiceError] = useState<string | null>(null);
   const mobileBaseVoiceRef = useRef('');
+  // Voice input of the mobile capture sheet: own Whisper server or the browser (lib/voice.ts)
+  const mobileVoice = useVoiceInput({
+    onText: (text) => {
+      if (!text) return;
+      const base = mobileBaseVoiceRef.current;
+      setMobileTriageContent(base ? `${base} ${text}` : text);
+      setIsVoiceCaptured(true);
+    },
+    onError: (message) => {
+      setMobileVoiceError(message);
+      window.setTimeout(() => setMobileVoiceError(null), 4000);
+    },
+  });
+  const mobileVoiceActive = mobileVoice.state === 'listening';
 
   const toggleMobileVoice = () => {
-    const SpeechRecognition =
-      (window as any).SpeechRecognition || (window as any).webkitSpeechRecognition;
-    if (!SpeechRecognition) {
-      alert('Spracheingabe wird von diesem Browser leider nicht unterstützt.');
-      return;
-    }
-    if (mobileVoiceActive && mobileRecognitionRef.current) {
-      mobileRecognitionRef.current.stop();
-      setMobileVoiceActive(false);
-      return;
-    }
-    try {
+    if (mobileVoice.state === 'idle') {
       mobileBaseVoiceRef.current = mobileTriageContent.replace(/\s*\[…\]$/, '').trim();
-      const recognition = new SpeechRecognition();
-      recognition.lang = 'de-DE';
-      recognition.continuous = false;
-      recognition.interimResults = true;
-      recognition.onstart = () => {
-        setMobileVoiceActive(true);
-        triggerHaptic(20);
-      };
-      recognition.onresult = (event: any) => {
-        let transcript = '';
-        for (let i = 0; i < event.results.length; i++) {
-          transcript += event.results[i][0].transcript;
-        }
-        const base = mobileBaseVoiceRef.current;
-        setMobileTriageContent(base ? `${base} ${transcript}` : transcript);
-        setIsVoiceCaptured(true);
-      };
-      recognition.onerror = () => setMobileVoiceActive(false);
-      recognition.onend = () => setMobileVoiceActive(false);
-      mobileRecognitionRef.current = recognition;
-      recognition.start();
-    } catch {
-      setMobileVoiceActive(false);
+      setMobileVoiceError(null);
+      triggerHaptic(20);
     }
+    mobileVoice.toggle();
   };
 
   const handleMobileCreateTriage = async (e: React.FormEvent) => {
@@ -249,7 +234,7 @@ export const App: React.FC = () => {
     setMobileTriagePriority('normal');
     setMobileTriageDueDate('');
     setIsVoiceCaptured(false);
-    setMobileVoiceActive(false);
+    if (mobileVoice.state !== 'idle') mobileVoice.cancel();
     setShowMobileCapture(true);
   };
 
@@ -1341,21 +1326,49 @@ export const App: React.FC = () => {
                     autoFocus
                     value={mobileTriageContent}
                     onChange={(e) => setMobileTriageContent(e.target.value)}
-                    placeholder="Was geht dir durch den Kopf?…"
-                    className="w-full input-field px-3 py-2 pr-10 text-sm"
+                    placeholder={
+                      mobileVoiceError
+                        ? mobileVoiceError
+                        : mobileVoiceActive
+                          ? 'Ich höre zu…'
+                          : mobileVoice.state === 'transcribing'
+                            ? 'Wird erkannt…'
+                            : 'Was geht dir durch den Kopf?…'
+                    }
+                    className={
+                      mobileVoiceActive && mobileVoice.mode === 'server'
+                        ? 'w-full input-field px-3 py-2 pr-24 text-sm'
+                        : 'w-full input-field px-3 py-2 pr-10 text-sm'
+                    }
                   />
-                  <button
-                    type="button"
-                    onClick={toggleMobileVoice}
-                    className={`absolute right-1.5 top-1/2 -translate-y-1/2 w-7 h-7 rounded-full flex items-center justify-center transition-colors cursor-pointer ${
-                      mobileVoiceActive
-                        ? 'bg-accent text-on-accent animate-pulse'
-                        : 'text-muted hover:text-main hover:bg-panel-elevated'
-                    }`}
-                    title="Spracheingabe"
-                  >
-                    {mobileVoiceActive ? <MicOff className="w-3.5 h-3.5" /> : <Mic className="w-3.5 h-3.5" />}
-                  </button>
+                  {mobileVoice.available && (
+                    <div className="absolute right-1.5 top-1/2 -translate-y-1/2 flex items-center gap-2">
+                      {mobileVoiceActive && mobileVoice.mode === 'server' && (
+                        <VoiceWave getLevel={mobileVoice.level} bars={7} className="h-4 text-accent" />
+                      )}
+                      <button
+                        type="button"
+                        onClick={toggleMobileVoice}
+                        disabled={mobileVoice.state === 'transcribing'}
+                        className={`w-9 h-9 rounded-full flex items-center justify-center transition-colors cursor-pointer ${
+                          mobileVoiceActive
+                            ? mobileVoice.mode === 'server'
+                              ? 'bg-status-err text-on-accent'
+                              : 'bg-accent text-on-accent animate-pulse'
+                            : 'text-muted hover:text-main hover:bg-panel-elevated'
+                        }`}
+                        title={mobileVoiceActive ? 'Aufnahme beenden' : 'Spracheingabe'}
+                      >
+                        {mobileVoice.state === 'transcribing' ? (
+                          <Loader2 className="w-4 h-4 animate-spin" />
+                        ) : mobileVoiceActive ? (
+                          <MicOff className="w-4 h-4" />
+                        ) : (
+                          <Mic className="w-4 h-4" />
+                        )}
+                      </button>
+                    </div>
+                  )}
                 </div>
               </div>
 

frontend/src/components/VoiceWave.tsx

@@ -0,0 +1,58 @@
+import React, { useEffect, useRef } from 'react';
+
+interface VoiceWaveProps {
+  /** Input level 0..1; null (unknown, e.g. browser recognition) shows a calm idle wave. */
+  getLevel: () => number | null;
+  bars?: number;
+  /** Bar colour and size come from the caller, e.g. "h-4 text-accent". */
+  className?: string;
+}
+
+// Live wave reacting to the voice. Heights are set directly on the bars every frame, so React does
+// not re-render 60 times a second.
+export const VoiceWave: React.FC<VoiceWaveProps> = ({ getLevel, bars = 5, className = 'h-4 text-accent' }) => {
+  const barRefs = useRef<(HTMLSpanElement | null)[]>([]);
+  const getLevelRef = useRef(getLevel);
+  getLevelRef.current = getLevel;
+
+  useEffect(() => {
+    const history: number[] = new Array(bars).fill(0);
+    let frame = 0;
+    let raf = 0;
+    const reduceMotion = window.matchMedia?.('(prefers-reduced-motion: reduce)').matches;
+    const tick = () => {
+      frame++;
+      const measured = getLevelRef.current();
+      const level = measured ?? 0.25 + 0.2 * Math.sin(frame / 9);
+      // Newest level enters in the middle and travels outwards
+      if (frame % 3 === 0) {
+        history.pop();
+        history.unshift(level);
+      }
+      const mid = Math.floor(bars / 2);
+      barRefs.current.forEach((el, i) => {
+        if (!el) return;
+        const v = reduceMotion ? 0.5 : history[Math.abs(i - mid)] ?? 0;
+        el.style.transform = `scaleY(${Math.max(0.15, Math.min(1, v * 1.4))})`;
+      });
+      raf = requestAnimationFrame(tick);
+    };
+    raf = requestAnimationFrame(tick);
+    return () => cancelAnimationFrame(raf);
+  }, [bars]);
+
+  return (
+    <span className={`inline-flex items-center gap-[3px] ${className}`} aria-hidden="true">
+      {Array.from({ length: bars }, (_, i) => (
+        <span
+          key={i}
+          ref={(el) => {
+            barRefs.current[i] = el;
+          }}
+          className="w-[3px] h-full rounded-full bg-current origin-center transition-transform duration-75"
+          style={{ transform: 'scaleY(0.15)' }}
+        />
+      ))}
+    </span>
+  );
+};

frontend/src/features/capture/HoldToRecordFab.tsx

@@ -1,6 +1,8 @@
 import React, { useEffect, useRef, useState } from 'react';
 import { Mic, Plus } from 'lucide-react';
 import { triggerHaptic } from '../../lib/api';
+import { useVoiceInput } from '../../lib/useVoiceInput';
+import { VoiceWave } from '../../components/VoiceWave';
 
 export interface HoldToRecordFabProps {
   onTap: () => void;
@@ -24,14 +26,10 @@ export const HoldToRecordFab: React.FC<HoldToRecordFabProps> = ({ onTap, onTrans
   const didEnterHoldRef = useRef(false);
   const fallbackHandledRef = useRef(false);
   const isCancelZoneRef = useRef(false);
-  const abortedRef = useRef(false);
-  const isStoppingRef = useRef(false);
-  const hasEndedRef = useRef(false);
   const finalizedRef = useRef(false);
   const startYRef = useRef(0);
   const transcriptRef = useRef('');
 
-  const recognitionRef = useRef<any>(null);
   const holdTimerRef = useRef<number | null>(null);
   const stopTimerRef = useRef<number | null>(null);
   const toastTimerRef = useRef<number | null>(null);
@@ -50,17 +48,66 @@ export const HoldToRecordFab: React.FC<HoldToRecordFabProps> = ({ onTap, onTrans
     }
   };
 
-  const showToast = (msg: string) => {
+  const showToast = (msg: string, sticky = false) => {
     if (toastTimerRef.current !== null) {
       window.clearTimeout(toastTimerRef.current);
+      toastTimerRef.current = null;
     }
     setToastMessage(msg);
+    if (sticky) return;
     toastTimerRef.current = window.setTimeout(() => {
       setToastMessage(null);
       toastTimerRef.current = null;
     }, TOAST_DURATION_MS);
   };
 
+  const finalizeRecording = async (text: string) => {
+    if (finalizedRef.current) return;
+    finalizedRef.current = true;
+    clearStopTimer();
+    setLiveTranscript('');
+    const trimmed = text.trim();
+    if (!trimmed) {
+      showToast('Nichts erkannt');
+      return;
+    }
+    try {
+      await onTranscript(trimmed);
+      triggerHaptic(25);
+      showToast('Im Quick Dump gespeichert');
+    } catch {
+      showToast('Fehler beim Speichern');
+    }
+  };
+
+  // Own Whisper server (text after release) or the browser's recognition (text while speaking)
+  const voice = useVoiceInput({
+    continuous: true,
+    onText: (text, final) => {
+      transcriptRef.current = text;
+      if (!final) {
+        setLiveTranscript(text);
+        return;
+      }
+      void finalizeRecording(text);
+    },
+    onError: (message) => {
+      clearStopTimer();
+      setIsRecording(false);
+      setIsCancelZone(false);
+      setLiveTranscript('');
+      if (didEnterHoldRef.current && isPressedRef.current && !fallbackHandledRef.current) {
+        // Recording could not start: open the normal capture sheet instead
+        fallbackHandledRef.current = true;
+        showToast(message);
+        onTap();
+        return;
+      }
+      finalizedRef.current = true;
+      showToast(message);
+    },
+  });
+
   useEffect(() => {
     return () => {
       clearHoldTimer();
@@ -68,112 +115,28 @@ export const HoldToRecordFab: React.FC<HoldToRecordFabProps> = ({ onTap, onTrans
       if (toastTimerRef.current !== null) {
         window.clearTimeout(toastTimerRef.current);
       }
-      if (recognitionRef.current) {
-        try {
-          recognitionRef.current.abort?.();
-        } catch {
-          // ignore cleanup errors
-        }
-      }
     };
   }, []);
 
-  const finalizeRecording = async () => {
-    if (finalizedRef.current) return;
-    finalizedRef.current = true;
-    clearStopTimer();
-    recognitionRef.current = null;
-
-    if (abortedRef.current) return;
-
-    const text = transcriptRef.current.trim();
-    setLiveTranscript('');
-    if (!text) {
-      showToast('Nichts erkannt');
-      return;
-    }
-
-    try {
-      await onTranscript(text);
-      triggerHaptic(25);
-      showToast('Im Quick Dump gespeichert');
-    } catch {
-      showToast('Fehler beim Speichern');
-    }
-  };
+  useEffect(() => {
+    if (voice.state === 'transcribing') showToast('Wird erkannt…', true);
+  }, [voice.state]);
 
   const startVoiceRecording = () => {
-    const SpeechRecognition =
-      (window as any).SpeechRecognition || (window as any).webkitSpeechRecognition;
-
-    if (!SpeechRecognition) {
+    if (voice.mode === 'off' || voice.mode === 'unsupported') {
       fallbackHandledRef.current = true;
-      showToast('Spracheingabe wird von diesem Browser nicht unterstützt');
+      if (voice.mode === 'unsupported') showToast('Spracheingabe wird von diesem Browser nicht unterstützt');
       onTap();
       return;
     }
-
-    try {
-      const recognition = new SpeechRecognition();
-      recognition.lang = 'de-DE';
-      recognition.interimResults = true;
-      recognition.continuous = true;
-
-      transcriptRef.current = '';
-      abortedRef.current = false;
-      isStoppingRef.current = false;
-      hasEndedRef.current = false;
-      finalizedRef.current = false;
-      isCancelZoneRef.current = false;
-
-      recognition.onresult = (event: any) => {
-        let combined = '';
-        for (let i = 0; i < event.results.length; i++) {
-          combined += event.results[i][0].transcript;
-        }
-        transcriptRef.current = combined;
-        setLiveTranscript(combined);
-      };
-
-      recognition.onerror = (event: any) => {
-        const errCode = event?.error;
-        if (errCode === 'not-allowed' || errCode === 'service-not-allowed') {
-          abortedRef.current = true;
-          clearStopTimer();
-          setIsRecording(false);
-          setIsCancelZone(false);
-          setLiveTranscript('');
-          if (!fallbackHandledRef.current) {
-            fallbackHandledRef.current = true;
-            showToast('Mikrofon nicht freigegeben');
-            onTap();
-          }
-          return;
-        }
-        if (isStoppingRef.current) {
-          void finalizeRecording();
-        }
-      };
-
-      recognition.onend = () => {
-        hasEndedRef.current = true;
-        if (isStoppingRef.current) {
-          void finalizeRecording();
-        }
-      };
-
-      recognitionRef.current = recognition;
-      setLiveTranscript('');
-      setIsCancelZone(false);
-      setIsRecording(true);
-      triggerHaptic([15, 60, 25]);
-      recognition.start();
-    } catch {
-      fallbackHandledRef.current = true;
-      setIsRecording(false);
-      showToast('Spracheingabe wird von diesem Browser nicht unterstützt');
-      onTap();
-    }
+    transcriptRef.current = '';
+    finalizedRef.current = false;
+    isCancelZoneRef.current = false;
+    setLiveTranscript('');
+    setIsCancelZone(false);
+    setIsRecording(true);
+    triggerHaptic([15, 60, 25]);
+    void voice.start();
   };
 
   const handlePointerDown = (e: React.PointerEvent<HTMLButtonElement>) => {
@@ -240,34 +203,23 @@ export const HoldToRecordFab: React.FC<HoldToRecordFabProps> = ({ onTap, onTrans
     setIsCancelZone(false);
 
     if (isCancelZoneRef.current) {
-      abortedRef.current = true;
+      finalizedRef.current = true;
       clearStopTimer();
       setLiveTranscript('');
       triggerHaptic(15);
-      try {
-        recognitionRef.current?.abort?.();
-      } catch {
-        // ignore
-      }
-      recognitionRef.current = null;
+      voice.cancel();
       return;
     }
 
-    isStoppingRef.current = true;
-    if (hasEndedRef.current || !recognitionRef.current) {
-      void finalizeRecording();
-      return;
-    }
-
-    stopTimerRef.current = window.setTimeout(() => {
-      void finalizeRecording();
-    }, STOP_TIMEOUT_MS);
-
-    try {
-      recognitionRef.current.stop();
-    } catch {
-      void finalizeRecording();
+    // The browser's recognition sometimes never reports its end: save what was heard after a moment.
+    // The own server answers when it is done (with its own time limit).
+    if (voice.mode !== 'server') {
+      stopTimerRef.current = window.setTimeout(() => {
+        voice.cancel();
+        void finalizeRecording(transcriptRef.current);
+      }, STOP_TIMEOUT_MS);
     }
+    void voice.stop();
   };
 
   const handlePointerCancel = (e: React.PointerEvent<HTMLButtonElement>) => {
@@ -283,16 +235,11 @@ export const HoldToRecordFab: React.FC<HoldToRecordFabProps> = ({ onTap, onTrans
     clearStopTimer();
 
     if (didEnterHoldRef.current && !fallbackHandledRef.current) {
-      abortedRef.current = true;
+      finalizedRef.current = true;
       setIsRecording(false);
       setIsCancelZone(false);
       setLiveTranscript('');
-      try {
-        recognitionRef.current?.abort?.();
-      } catch {
-        // ignore
-      }
-      recognitionRef.current = null;
+      voice.cancel();
     }
   };
 
@@ -313,7 +260,8 @@ export const HoldToRecordFab: React.FC<HoldToRecordFabProps> = ({ onTap, onTrans
             <div className="text-xs font-semibold">🗑️ Zum Abbrechen loslassen</div>
           ) : (
             <div className="space-y-1">
-              <div className="text-xs font-semibold text-main">
+              <div className="text-xs font-semibold text-main flex items-center justify-center gap-2">
+                {voice.mode === 'server' && <VoiceWave getLevel={voice.level} bars={7} className="h-4 text-accent" />}
                 Ich höre zu… (loslassen speichert)
               </div>
               {displayedTranscript ? (

frontend/src/features/capture/InstantCaptureBar.tsx

@@ -13,9 +13,12 @@ import {
   ChevronDown,
   AlertTriangle,
   Flame,
+  Loader2,
 } from 'lucide-react';
 import { PriorityLevel, SharedList, Thought } from '../../lib/types';
 import { api, triggerHaptic } from '../../lib/api';
+import { useVoiceInput } from '../../lib/useVoiceInput';
+import { VoiceWave } from '../../components/VoiceWave';
 
 export type CaptureTarget = 'inbox' | 'task' | 'list_item' | 'document';
 
@@ -46,17 +49,31 @@ export const InstantCaptureBar: React.FC<InstantCaptureBarProps> = ({
   const [target, setTarget] = useState<CaptureTarget>(fixedTarget || 'inbox');
   const [targetListId, setTargetListId] = useState<string>('');
   const [showOptions, setShowOptions] = useState(defaultShowOptions || hideTargetPicker);
-  const [isListening, setIsListening] = useState(false);
+  const [voiceError, setVoiceError] = useState<string | null>(null);
   const [isSubmitting, setIsSubmitting] = useState(false);
   const [justSavedLabel, setJustSavedLabel] = useState<string | null>(null);
   const [isMultiline, setIsMultiline] = useState(false);
 
   const inputRef = useRef<HTMLTextAreaElement>(null);
   const containerRef = useRef<HTMLFormElement>(null);
-  const recognitionRef = useRef<any>(null);
   const baseVoiceTextRef = useRef('');
   const hasVoiceInputRef = useRef(false);
 
+  // Voice input: own Whisper server or the browser's recognition (lib/voice.ts)
+  const voice = useVoiceInput({
+    onText: (text, final) => {
+      const base = baseVoiceTextRef.current;
+      if (text) setContent(base ? `${base} ${text}` : text);
+      if (final) inputRef.current?.focus();
+    },
+    onError: (message) => {
+      setVoiceError(message);
+      window.setTimeout(() => setVoiceError(null), 4000);
+    },
+  });
+  const isListening = voice.state === 'listening';
+  const isTranscribing = voice.state === 'transcribing';
+
   const adjustHeight = (el?: HTMLTextAreaElement | null) => {
     const targetEl = el || inputRef.current;
     if (!targetEl) return;
@@ -120,57 +137,13 @@ export const InstantCaptureBar: React.FC<InstantCaptureBarProps> = ({
   }, []);
 
   const toggleVoiceDictation = () => {
-    const SpeechRecognition =
-      (window as any).SpeechRecognition || (window as any).webkitSpeechRecognition;
-
-    if (!SpeechRecognition) {
-      alert('Spracheingabe wird von diesem Browser nicht direkt unterstützt (bitte Chrome, Edge oder Safari nutzen).');
-      return;
-    }
-
-    if (isListening && recognitionRef.current) {
-      recognitionRef.current.stop();
-      setIsListening(false);
-      return;
-    }
-
-    try {
+    if (voice.state === 'idle') {
       baseVoiceTextRef.current = content.replace(/\s*\[…\]$/, '').trim();
       hasVoiceInputRef.current = true;
-
-      const recognition = new SpeechRecognition();
-      recognition.lang = 'de-DE';
-      recognition.continuous = false;
-      recognition.interimResults = true;
-
-      recognition.onstart = () => {
-        setIsListening(true);
-        triggerHaptic(20);
-      };
-
-      recognition.onresult = (event: any) => {
-        let transcript = '';
-        for (let i = 0; i < event.results.length; i++) {
-          transcript += event.results[i][0].transcript;
-        }
-        const base = baseVoiceTextRef.current;
-        setContent(base ? `${base} ${transcript}` : transcript);
-      };
-
-      recognition.onerror = () => {
-        setIsListening(false);
-      };
-
-      recognition.onend = () => {
-        setIsListening(false);
-        inputRef.current?.focus();
-      };
-
-      recognitionRef.current = recognition;
-      recognition.start();
-    } catch {
-      setIsListening(false);
+      setVoiceError(null);
+      triggerHaptic(20);
     }
+    voice.toggle();
   };
 
   const executeCapture = async (
@@ -187,6 +160,8 @@ export const InstantCaptureBar: React.FC<InstantCaptureBarProps> = ({
     const capturedDueDate = dueDate;
     const capturedSource = hasVoiceInputRef.current || isListening ? 'voice' : 'manual';
     hasVoiceInputRef.current = false;
+    // A recognition still running would refill the field after it was saved
+    if (voice.state !== 'idle') voice.cancel();
 
     // Optimistically clear input immediately so the UI feels instant (<10ms) in all browsers
     setContent('');
@@ -285,8 +260,12 @@ export const InstantCaptureBar: React.FC<InstantCaptureBarProps> = ({
               }
             }}
             placeholder={
-              isListening
+              voiceError
+                ? voiceError
+                : isListening
                 ? 'Ich höre zu…'
+                : isTranscribing
+                ? 'Wird erkannt…'
                 : hideTargetPicker || fixedTarget === 'inbox'
                 ? 'Quick Dump erfassen… (Enter zum Speichern)'
                 : target === 'task'
@@ -336,19 +315,33 @@ export const InstantCaptureBar: React.FC<InstantCaptureBarProps> = ({
               </button>
             )}
 
-            {/* Voice Dictation Button */}
-            <button
-              type="button"
-              onClick={toggleVoiceDictation}
-              title="Spracheingabe"
-              className={`w-6 h-6 capture-pill rounded-full flex items-center justify-center transition-colors ${
-                isListening
-                  ? 'bg-accent-dim text-accent'
-                  : 'text-muted hover:text-main hover:bg-panel-elevated'
-              }`}
-            >
-              {isListening ? <MicOff className="w-3 h-3" /> : <Mic className="w-3 h-3" />}
-            </button>
+            {/* Voice Dictation Button (hidden when voice input is off or not supported) */}
+            {voice.available && (
+              <>
+                {isListening && voice.mode === 'server' && (
+                  <VoiceWave getLevel={voice.level} className="h-3.5 text-accent" />
+                )}
+                <button
+                  type="button"
+                  onClick={toggleVoiceDictation}
+                  disabled={isTranscribing}
+                  title={isListening ? 'Aufnahme beenden' : isTranscribing ? 'Wird erkannt…' : 'Spracheingabe'}
+                  className={`w-6 h-6 capture-pill rounded-full flex items-center justify-center transition-colors ${
+                    isListening || isTranscribing
+                      ? 'bg-accent-dim text-accent'
+                      : 'text-muted hover:text-main hover:bg-panel-elevated'
+                  }`}
+                >
+                  {isTranscribing ? (
+                    <Loader2 className="w-3 h-3 animate-spin" />
+                  ) : isListening ? (
+                    <MicOff className="w-3 h-3" />
+                  ) : (
+                    <Mic className="w-3 h-3" />
+                  )}
+                </button>
+              </>
+            )}
 
             {/* Instant Send Button */}
             <button

frontend/src/lib/__tests__/useVoiceInput.test.tsx

@@ -0,0 +1,133 @@
+import { describe, it, expect, vi, beforeEach, afterEach } from 'vitest';
+import { act, renderHook, waitFor } from '@testing-library/react';
+
+const getAuthStatus = vi.fn();
+const transcribe = vi.fn();
+vi.mock('../api', () => ({ api: { getAuthStatus: (...a: any[]) => getAuthStatus(...a), transcribe: (...a: any[]) => transcribe(...a) } }));
+
+import { useVoiceInput } from '../useVoiceInput';
+import { resetVoiceConfig } from '../voice';
+
+class FakeMediaRecorder {
+  static isTypeSupported = (t: string) => t === 'audio/webm;codecs=opus';
+  state: 'inactive' | 'recording' = 'inactive';
+  mimeType = 'audio/webm;codecs=opus';
+  ondataavailable: ((e: { data: Blob }) => void) | null = null;
+  onstop: (() => void) | null = null;
+  constructor(public stream: any) {}
+  start() {
+    this.state = 'recording';
+  }
+  stop() {
+    this.state = 'inactive';
+    this.ondataavailable?.({ data: new Blob(['audio'], { type: this.mimeType }) });
+    this.onstop?.();
+  }
+}
+
+const stopTrack = vi.fn();
+
+beforeEach(() => {
+  resetVoiceConfig();
+  getAuthStatus.mockReset();
+  transcribe.mockReset();
+  stopTrack.mockReset();
+  (window as any).MediaRecorder = FakeMediaRecorder;
+  Object.defineProperty(navigator, 'mediaDevices', {
+    configurable: true,
+    value: { getUserMedia: vi.fn(async () => ({ getTracks: () => [{ stop: stopTrack }] })) },
+  });
+});
+
+afterEach(() => {
+  delete (window as any).MediaRecorder;
+  delete (window as any).SpeechRecognition;
+  vi.useRealTimers();
+});
+
+describe('useVoiceInput', () => {
+  it('server mode: records, sends the audio to the own server and returns its text', async () => {
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'server', maxSeconds: 600 } });
+    transcribe.mockResolvedValue('Milch kaufen');
+    const onText = vi.fn();
+    const { result } = renderHook(() => useVoiceInput({ onText }));
+    await waitFor(() => expect(result.current.mode).toBe('server'));
+    expect(result.current.available).toBe(true);
+
+    await act(async () => {
+      await result.current.start();
+    });
+    expect(result.current.state).toBe('listening');
+
+    await act(async () => {
+      await result.current.stop();
+    });
+    expect(transcribe).toHaveBeenCalledTimes(1);
+    expect((transcribe.mock.calls[0][0] as Blob).type).toBe('audio/webm;codecs=opus');
+    expect(onText).toHaveBeenCalledWith('Milch kaufen', true);
+    expect(result.current.state).toBe('idle');
+    expect(stopTrack).toHaveBeenCalled(); // microphone released
+  });
+
+  it('server mode: cancel throws the recording away', async () => {
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'server', maxSeconds: 600 } });
+    const onText = vi.fn();
+    const { result } = renderHook(() => useVoiceInput({ onText }));
+    await waitFor(() => expect(result.current.mode).toBe('server'));
+    await act(async () => {
+      await result.current.start();
+    });
+    act(() => result.current.cancel());
+    expect(transcribe).not.toHaveBeenCalled();
+    expect(onText).not.toHaveBeenCalled();
+    expect(result.current.state).toBe('idle');
+  });
+
+  it('server mode: a failing server reports an error instead of text', async () => {
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'server', maxSeconds: 600 } });
+    transcribe.mockRejectedValue(new Error('Spracherkennung fehlgeschlagen'));
+    const onText = vi.fn();
+    const onError = vi.fn();
+    const { result } = renderHook(() => useVoiceInput({ onText, onError }));
+    await waitFor(() => expect(result.current.mode).toBe('server'));
+    await act(async () => {
+      await result.current.start();
+    });
+    await act(async () => {
+      await result.current.stop();
+    });
+    expect(onError).toHaveBeenCalledWith('Spracherkennung fehlgeschlagen');
+    expect(onText).not.toHaveBeenCalled();
+    expect(result.current.state).toBe('idle');
+  });
+
+  it('server mode: stops by itself after the maximum length', async () => {
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'server', maxSeconds: 5 } });
+    transcribe.mockResolvedValue('Zu lang');
+    const onText = vi.fn();
+    const { result } = renderHook(() => useVoiceInput({ onText }));
+    await waitFor(() => expect(result.current.mode).toBe('server'));
+    vi.useFakeTimers();
+    await act(async () => {
+      await result.current.start();
+    });
+    await act(async () => {
+      vi.advanceTimersByTime(5000);
+    });
+    vi.useRealTimers();
+    await waitFor(() => expect(onText).toHaveBeenCalledWith('Zu lang', true));
+  });
+
+  it('hides the microphone when switched off, and without browser recognition', async () => {
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'off', maxSeconds: 600 } });
+    const off = renderHook(() => useVoiceInput({ onText: vi.fn() }));
+    await waitFor(() => expect(off.result.current.mode).toBe('off'));
+    expect(off.result.current.available).toBe(false);
+
+    resetVoiceConfig();
+    getAuthStatus.mockResolvedValue({ voice: { mode: 'browser', maxSeconds: 600 } });
+    const none = renderHook(() => useVoiceInput({ onText: vi.fn() }));
+    await waitFor(() => expect(none.result.current.mode).toBe('unsupported'));
+    expect(none.result.current.available).toBe(false);
+  });
+});

frontend/src/lib/useVoiceInput.ts

@@ -0,0 +1,200 @@
+import { useCallback, useEffect, useRef, useState } from 'react';
+import { api } from './api';
+import { browserRecognition, currentVoiceConfig, EffectiveVoiceMode, loadVoiceConfig, VoiceRecorder } from './voice';
+
+export type VoiceState = 'idle' | 'listening' | 'transcribing';
+
+export interface VoiceInputOptions {
+  /** Text of the current recording; in browser mode repeatedly while speaking (final=false), always once at the end (final=true). */
+  onText: (text: string, final: boolean) => void;
+  onError?: (message: string) => void;
+  /** Browser mode: keep listening through pauses (hold-to-record) instead of stopping after a sentence. */
+  continuous?: boolean;
+}
+
+/**
+ * Microphone input for one button: start, stop (text arrives), cancel (nothing arrives). How the text
+ * is made depends on the server setting, see lib/voice.ts. mode is null until known, "off" or
+ * "unsupported" means: do not show a microphone.
+ */
+export function useVoiceInput({ onText, onError, continuous = false }: VoiceInputOptions) {
+  const [mode, setMode] = useState<EffectiveVoiceMode | null>(null);
+  const [state, setState] = useState<VoiceState>('idle');
+  const maxSecondsRef = useRef(600);
+  const recorderRef = useRef<VoiceRecorder | null>(null);
+  const recognitionRef = useRef<any>(null);
+  const timerRef = useRef<number | null>(null);
+  const cancelledRef = useRef(false);
+  const transcriptRef = useRef('');
+  const mountedRef = useRef(true);
+  const onTextRef = useRef(onText);
+  onTextRef.current = onText;
+  const onErrorRef = useRef(onError);
+  onErrorRef.current = onError;
+
+  useEffect(() => {
+    mountedRef.current = true;
+    void loadVoiceConfig().then((cfg) => {
+      if (!mountedRef.current) return;
+      setMode(cfg.mode);
+      maxSecondsRef.current = cfg.maxSeconds;
+    });
+    return () => {
+      mountedRef.current = false;
+      if (timerRef.current !== null) window.clearTimeout(timerRef.current);
+      recorderRef.current?.cancel();
+      try {
+        recognitionRef.current?.abort?.();
+      } catch {
+        // ignore
+      }
+    };
+  }, []);
+
+  const setStateSafe = (s: VoiceState) => {
+    if (mountedRef.current) setState(s);
+  };
+  const fail = (message: string) => {
+    setStateSafe('idle');
+    onErrorRef.current?.(message);
+  };
+  const clearTimer = () => {
+    if (timerRef.current !== null) window.clearTimeout(timerRef.current);
+    timerRef.current = null;
+  };
+
+  const stop = useCallback(async () => {
+    clearTimer();
+    if (recognitionRef.current) {
+      try {
+        recognitionRef.current.stop();
+      } catch {
+        // onend still follows
+      }
+      return;
+    }
+    const rec = recorderRef.current;
+    if (!rec) return;
+    recorderRef.current = null;
+    const blob = await rec.stop();
+    if (cancelledRef.current) return;
+    if (!blob) {
+      setStateSafe('idle');
+      onTextRef.current('', true);
+      return;
+    }
+    setStateSafe('transcribing');
+    try {
+      const text = await api.transcribe(blob);
+      setStateSafe('idle');
+      onTextRef.current(text, true);
+    } catch (err: any) {
+      fail(err?.message || 'Spracherkennung fehlgeschlagen');
+    }
+    // fail only touches refs and state setters
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, []);
+
+  const startServer = async () => {
+    const rec = new VoiceRecorder();
+    recorderRef.current = rec;
+    setStateSafe('listening');
+    try {
+      await rec.start();
+    } catch {
+      recorderRef.current = null;
+      fail('Kein Zugriff auf das Mikrofon (Berechtigung, oder die Seite läuft nicht über HTTPS)');
+      return;
+    }
+    if (cancelledRef.current || recorderRef.current !== rec) {
+      rec.cancel();
+      return;
+    }
+    timerRef.current = window.setTimeout(() => void stop(), maxSecondsRef.current * 1000);
+  };
+
+  const startBrowser = () => {
+    const Recognition = browserRecognition();
+    if (!Recognition) {
+      fail('Spracheingabe wird von diesem Browser nicht unterstützt');
+      return;
+    }
+    const recognition = new Recognition();
+    recognition.lang = 'de-DE';
+    recognition.interimResults = true;
+    recognition.continuous = continuous;
+    recognition.onresult = (event: any) => {
+      let combined = '';
+      for (let i = 0; i < event.results.length; i++) combined += event.results[i][0].transcript;
+      transcriptRef.current = combined;
+      onTextRef.current(combined, false);
+    };
+    recognition.onerror = (event: any) => {
+      if (event?.error === 'not-allowed' || event?.error === 'service-not-allowed') {
+        cancelledRef.current = true;
+        recognitionRef.current = null;
+        fail('Mikrofon nicht freigegeben');
+      }
+    };
+    recognition.onend = () => {
+      clearTimer();
+      recognitionRef.current = null;
+      if (cancelledRef.current) return;
+      setStateSafe('idle');
+      onTextRef.current(transcriptRef.current, true);
+    };
+    recognitionRef.current = recognition;
+    setStateSafe('listening');
+    try {
+      recognition.start();
+      timerRef.current = window.setTimeout(() => void stop(), maxSecondsRef.current * 1000);
+    } catch {
+      recognitionRef.current = null;
+      fail('Spracheingabe konnte nicht gestartet werden');
+    }
+  };
+
+  const start = useCallback(async () => {
+    if (recorderRef.current || recognitionRef.current) return;
+    cancelledRef.current = false;
+    transcriptRef.current = '';
+    const current = mode ?? currentVoiceConfig().mode;
+    if (current === 'server') await startServer();
+    else if (current === 'browser') startBrowser();
+    else fail(current === 'off' ? 'Spracheingabe ist ausgeschaltet' : 'Spracheingabe wird von diesem Browser nicht unterstützt');
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [mode, continuous]);
+
+  const cancel = useCallback(() => {
+    cancelledRef.current = true;
+    clearTimer();
+    recorderRef.current?.cancel();
+    recorderRef.current = null;
+    try {
+      recognitionRef.current?.abort?.();
+    } catch {
+      // ignore
+    }
+    recognitionRef.current = null;
+    setStateSafe('idle');
+  }, []);
+
+  const toggle = useCallback(() => {
+    if (state === 'listening') void stop();
+    else if (state === 'idle') void start();
+  }, [state, start, stop]);
+
+  /** Current input level 0..1 for waves (server mode), null when unknown. */
+  const level = useCallback(() => recorderRef.current?.level() ?? null, []);
+
+  return {
+    mode,
+    available: mode === 'server' || mode === 'browser',
+    state,
+    start,
+    stop,
+    cancel,
+    toggle,
+    level,
+  };
+}

frontend/src/lib/voice.ts

@@ -0,0 +1,157 @@
+// Voice input for all places with a microphone button. The server decides how it works (STT_* in the
+// environment, reported by /auth/status):
+//   server   record with MediaRecorder, the own Whisper server turns it into text after stopping
+//   browser  the browser's speech recognition (Chrome, Edge, Safari; text appears while speaking)
+//   off      no voice input
+import { api } from './api';
+import type { VoiceMode } from './types';
+
+export type EffectiveVoiceMode = VoiceMode | 'unsupported';
+
+export interface VoiceConfig {
+  mode: EffectiveVoiceMode;
+  maxSeconds: number;
+}
+
+let configPromise: Promise<VoiceConfig> | null = null;
+let loadedConfig: VoiceConfig | null = null;
+
+export function browserRecognition(): any {
+  if (typeof window === 'undefined') return null;
+  return (window as any).SpeechRecognition || (window as any).webkitSpeechRecognition || null;
+}
+
+export function canRecord(): boolean {
+  return typeof window !== 'undefined' && typeof (window as any).MediaRecorder !== 'undefined' && !!navigator.mediaDevices?.getUserMedia;
+}
+
+/** What this browser can actually do with the server's setting. */
+export function effectiveMode(mode: VoiceMode | undefined): EffectiveVoiceMode {
+  if (mode === 'off') return 'off';
+  if (mode === 'server') return canRecord() ? 'server' : 'unsupported';
+  return browserRecognition() ? 'browser' : 'unsupported';
+}
+
+/** Loaded once per page; the setting only changes with a server restart. */
+export function loadVoiceConfig(): Promise<VoiceConfig> {
+  if (!configPromise) {
+    configPromise = api
+      .getAuthStatus()
+      .then((st) => {
+        loadedConfig = { mode: effectiveMode(st.voice?.mode), maxSeconds: st.voice?.maxSeconds || 600 };
+        return loadedConfig;
+      })
+      .catch(() => {
+        configPromise = null; // offline at start: ask again next time
+        return { mode: effectiveMode('browser'), maxSeconds: 600 };
+      });
+  }
+  return configPromise;
+}
+
+/** The setting if already loaded; until then the browser's recognition (a button must react at once). */
+export function currentVoiceConfig(): VoiceConfig {
+  return loadedConfig ?? { mode: effectiveMode('browser'), maxSeconds: 600 };
+}
+
+/** For tests. */
+export function resetVoiceConfig(): void {
+  configPromise = null;
+  loadedConfig = null;
+}
+
+// Containers in order of preference; Safari only records MP4/AAC
+const RECORDING_TYPES = ['audio/webm;codecs=opus', 'audio/webm', 'audio/mp4', 'audio/ogg;codecs=opus'];
+
+function recordingType(): string | undefined {
+  const MR = (window as any).MediaRecorder;
+  if (!MR?.isTypeSupported) return undefined;
+  return RECORDING_TYPES.find((t) => MR.isTypeSupported(t));
+}
+
+/**
+ * One recording from the microphone, with a live input level for wave animations. stop() resolves
+ * with the audio; cancel() throws it away.
+ */
+export class VoiceRecorder {
+  private stream: MediaStream | null = null;
+  private recorder: MediaRecorder | null = null;
+  private chunks: Blob[] = [];
+  private audioCtx: AudioContext | null = null;
+  private analyser: AnalyserNode | null = null;
+  private samples: Uint8Array | null = null;
+  private done: ((blob: Blob | null) => void) | null = null;
+
+  async start(): Promise<void> {
+    this.stream = await navigator.mediaDevices.getUserMedia({
+      audio: { echoCancellation: true, noiseSuppression: true, autoGainControl: true },
+    });
+    const type = recordingType();
+    this.recorder = new MediaRecorder(this.stream, type ? { mimeType: type } : undefined);
+    this.chunks = [];
+    this.recorder.ondataavailable = (e) => {
+      if (e.data.size > 0) this.chunks.push(e.data);
+    };
+    this.recorder.onstop = () => {
+      const blob = this.chunks.length ? new Blob(this.chunks, { type: this.recorder?.mimeType || type || 'audio/webm' }) : null;
+      this.release();
+      this.done?.(blob);
+      this.done = null;
+    };
+    this.recorder.start(1000);
+    try {
+      const Ctx = (window as any).AudioContext || (window as any).webkitAudioContext;
+      this.audioCtx = new Ctx();
+      const source = this.audioCtx!.createMediaStreamSource(this.stream);
+      this.analyser = this.audioCtx!.createAnalyser();
+      this.analyser.fftSize = 512;
+      source.connect(this.analyser);
+      this.samples = new Uint8Array(this.analyser.fftSize);
+      void this.audioCtx!.resume?.();
+    } catch {
+      this.analyser = null; // no level, the recording still works
+    }
+  }
+
+  /** Input level 0..1 (roughly loudness), or null when unknown. */
+  level(): number | null {
+    if (!this.analyser || !this.samples) return null;
+    this.analyser.getByteTimeDomainData(this.samples as Uint8Array<ArrayBuffer>);
+    let sum = 0;
+    for (const v of this.samples) {
+      const x = (v - 128) / 128;
+      sum += x * x;
+    }
+    return Math.min(1, Math.sqrt(sum / this.samples.length) * 4);
+  }
+
+  stop(): Promise<Blob | null> {
+    return new Promise((resolve) => {
+      if (!this.recorder || this.recorder.state === 'inactive') {
+        this.release();
+        resolve(null);
+        return;
+      }
+      this.done = resolve;
+      this.recorder.stop();
+    });
+  }
+
+  cancel(): void {
+    this.done = null;
+    if (this.recorder && this.recorder.state !== 'inactive') {
+      this.recorder.onstop = () => this.release();
+      this.recorder.stop();
+    } else {
+      this.release();
+    }
+  }
+
+  private release(): void {
+    this.stream?.getTracks().forEach((t) => t.stop());
+    this.stream = null;
+    void this.audioCtx?.close?.().catch(() => {});
+    this.audioCtx = null;
+    this.analyser = null;
+  }
+}