release: bump version to 0.0.5.2

Neuer Container aria-whisper-bridge auf der Gamebox — faster-whisper
CUDA mit float16. Der Container verbindet sich per WebSocket an den RVS,
nimmt stt_request entgegen, laeuft ffmpeg+Whisper, antwortet mit
stt_response. Hoert zusaetzlich auf config-Broadcasts und lädt das
Modell hot-swap bei Diagnostic-Wechsel.

aria-bridge ruft jetzt primaer die Gamebox an; nur wenn die nicht binnen
45s antwortet, faellt auf lokales Whisper (CPU) zurueck. Das lokale
Modell wird lazy geladen, spart RAM auf der VM.

RVS: stt_request/stt_response zur ALLOWED_TYPES-Liste.

Diagnostic-Voice-Config (whisperModel-Feld) bleibt unveraendert —
die Auswahl wird an die Gamebox durchgereicht.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

android/android/app/build.gradle

@@ -79,8 +79,8 @@ android {
         applicationId "com.ariacockpit"
         minSdkVersion rootProject.ext.minSdkVersion
         targetSdkVersion rootProject.ext.targetSdkVersion
-        versionCode 406
-        versionName "0.0.4.6"
+        versionCode 502
+        versionName "0.0.5.2"
         // Fallback fuer Libraries mit Product Flavors
         missingDimensionStrategy 'react-native-camera', 'general'
     }

android/android/app/src/main/java/com/ariacockpit/PcmStreamPlayerModule.kt

@@ -30,9 +30,13 @@ import java.util.concurrent.LinkedBlockingQueue
 class PcmStreamPlayerModule(reactContext: ReactApplicationContext) : ReactContextBaseJavaModule(reactContext) {
     companion object {
         private const val TAG = "PcmStreamPlayer"
-        // Sekunden Audio die VOR play()-Start gepuffert sein muessen.
-        // 2.5s Vorrat = genug um XTTS-Render-Pausen zwischen Chunks zu puffern.
-        private const val PREROLL_SECONDS = 2.5
+        // Fallback wenn JS keinen Wert uebergibt.
+        private const val DEFAULT_PREROLL_SECONDS = 3.5
+        private const val MIN_PREROLL_SECONDS = 0.5
+        private const val MAX_PREROLL_SECONDS = 10.0
+        // Stille am Stream-Anfang, damit AudioTrack sauber anfaehrt und die
+        // ersten Samples nicht abgeschnitten werden (XTTS-Warmup + play()-Latenz).
+        private const val LEADING_SILENCE_SECONDS = 0.2
     }
 
     override fun getName() = "PcmStreamPlayer"
@@ -50,17 +54,21 @@ class PcmStreamPlayerModule(reactContext: ReactApplicationContext) : ReactContex
     // ── Lifecycle ──
 
     @ReactMethod
-    fun start(sampleRate: Int, channels: Int, promise: Promise) {
+    fun start(sampleRate: Int, channels: Int, prerollSeconds: Double, promise: Promise) {
         try {
             // Alte Session beenden falls vorhanden
             stopInternal()
 
+            val prerollSec = prerollSeconds
+                .coerceIn(MIN_PREROLL_SECONDS, MAX_PREROLL_SECONDS)
+                .let { if (it.isFinite() && it > 0) it else DEFAULT_PREROLL_SECONDS }
+
             val channelConfig = if (channels == 2) AudioFormat.CHANNEL_OUT_STEREO else AudioFormat.CHANNEL_OUT_MONO
             val encoding = AudioFormat.ENCODING_PCM_16BIT
             val minBuf = AudioTrack.getMinBufferSize(sampleRate, channelConfig, encoding)
             val bytesPerSecond = sampleRate * channels * 2 // 16-bit = 2 bytes
             // Buffer muss mindestens PREROLL + etwas Spielraum fassen.
-            val prerollTarget = (bytesPerSecond * PREROLL_SECONDS).toInt()
+            val prerollTarget = (bytesPerSecond * prerollSec).toInt()
             val bufferSize = (minBuf * 32).coerceAtLeast(prerollTarget * 2)
             prerollBytes = prerollTarget
             bytesBuffered = 0
@@ -94,6 +102,18 @@ class PcmStreamPlayerModule(reactContext: ReactApplicationContext) : ReactContex
             writerThread = Thread({
                 val t = track ?: return@Thread
                 try {
+                    // Leading-Silence in den Buffer — gibt AudioTrack Zeit anzufahren.
+                    val silenceBytes = ((sampleRate * channels * 2) * LEADING_SILENCE_SECONDS).toInt() and 0x7FFFFFFE
+                    if (silenceBytes > 0) {
+                        val silence = ByteArray(silenceBytes)
+                        var silOff = 0
+                        while (silOff < silence.size && !writerShouldStop) {
+                            val w = t.write(silence, silOff, silence.size - silOff)
+                            if (w <= 0) break
+                            silOff += w
+                        }
+                        bytesBuffered += silence.size
+                    }
                     while (!writerShouldStop) {
                         val data = queue.poll(50, java.util.concurrent.TimeUnit.MILLISECONDS) ?: run {
                             if (endRequested) {
@@ -158,7 +178,7 @@ class PcmStreamPlayerModule(reactContext: ReactApplicationContext) : ReactContex
                 }
             }, "PcmStreamWriter").apply { start() }
 
-            Log.i(TAG, "Stream gestartet: ${sampleRate}Hz ch=$channels buf=${bufferSize}B preroll=${prerollBytes}B")
+            Log.i(TAG, "Stream gestartet: ${sampleRate}Hz ch=$channels buf=${bufferSize}B preroll=${prerollBytes}B (${prerollSec}s)")
             promise.resolve(true)
         } catch (e: Exception) {
             Log.e(TAG, "start fehlgeschlagen", e)

android/package.json

@@ -1,6 +1,6 @@
 {
   "name": "aria-cockpit",
-  "version": "0.0.4.6",
+  "version": "0.0.5.2",
   "private": true,
   "scripts": {
     "android": "react-native run-android",

android/src/screens/ChatScreen.tsx

@@ -18,6 +18,7 @@ import {
   Image,
   ScrollView,
   Modal,
+  ToastAndroid,
 } from 'react-native';
 import AsyncStorage from '@react-native-async-storage/async-storage';
 import RNFS from 'react-native-fs';
@@ -325,6 +326,26 @@ const ChatScreen: React.FC = () => {
         const tool = (message.payload.tool as string) || '';
         setAgentActivity({ activity, tool });
       }
+
+      // Voice-Config aus Diagnostic — setzt die lokale App-Stimme auf den
+      // gerade in Diagnostic gewaehlten Wert zurueck. User-Wahl in der App
+      // wird dadurch ueberschrieben.
+      if (message.type === ('config' as any)) {
+        const newVoice = ((message.payload as any).xttsVoice as string) ?? '';
+        localXttsVoiceRef.current = newVoice;
+        AsyncStorage.setItem('aria_xtts_voice', newVoice);
+      }
+
+      // XTTS-Bridge meldet Stimme fertig geladen (kurzer Status-Toast)
+      if (message.type === ('voice_ready' as any)) {
+        const v = ((message.payload as any).voice as string) ?? '';
+        const err = (message.payload as any).error as string | undefined;
+        if (err) {
+          ToastAndroid.show(`Stimme "${v}" Fehler: ${err}`, ToastAndroid.LONG);
+        } else {
+          ToastAndroid.show(`Stimme "${v || 'Standard'}" bereit`, ToastAndroid.SHORT);
+        }
+      }
     });
 
     const unsubState = rvs.onStateChange((state) => {

android/src/screens/SettingsScreen.tsx

@@ -15,11 +15,19 @@ import {
   StyleSheet,
   Alert,
   Platform,
+  ToastAndroid,
+  ActivityIndicator,
 } from 'react-native';
 import AsyncStorage from '@react-native-async-storage/async-storage';
 import RNFS from 'react-native-fs';
 import DocumentPicker from 'react-native-document-picker';
 import rvs, { ConnectionState, RVSMessage, ConnectionConfig, ConnectionLogEntry } from '../services/rvs';
+import {
+  TTS_PREROLL_DEFAULT_SEC,
+  TTS_PREROLL_MIN_SEC,
+  TTS_PREROLL_MAX_SEC,
+  TTS_PREROLL_STORAGE_KEY,
+} from '../services/audio';
 import ModeSelector from '../components/ModeSelector';
 import QRScanner from '../components/QRScanner';
 import VoiceCloneModal from '../components/VoiceCloneModal';
@@ -73,8 +81,10 @@ const SettingsScreen: React.FC = () => {
   const [autoDownload, setAutoDownload] = useState(true);
   const [storageSize, setStorageSize] = useState('...');
   const [ttsEnabled, setTtsEnabled] = useState(true);
+  const [ttsPrerollSec, setTtsPrerollSec] = useState<number>(TTS_PREROLL_DEFAULT_SEC);
   const [editingPath, setEditingPath] = useState(false);
   const [xttsVoice, setXttsVoice] = useState('');
+  const [loadingVoice, setLoadingVoice] = useState<string | null>(null);
   const [availableVoices, setAvailableVoices] = useState<Array<{name: string, size: number}>>([]);
   const [voiceCloneVisible, setVoiceCloneVisible] = useState(false);
   const [tempPath, setTempPath] = useState('');
@@ -99,6 +109,14 @@ const SettingsScreen: React.FC = () => {
     AsyncStorage.getItem('aria_tts_enabled').then(saved => {
       if (saved !== null) setTtsEnabled(saved === 'true');
     });
+    AsyncStorage.getItem(TTS_PREROLL_STORAGE_KEY).then(saved => {
+      if (saved != null) {
+        const n = parseFloat(saved);
+        if (isFinite(n) && n >= TTS_PREROLL_MIN_SEC && n <= TTS_PREROLL_MAX_SEC) {
+          setTtsPrerollSec(n);
+        }
+      }
+    });
     AsyncStorage.getItem('aria_xtts_voice').then(saved => {
       if (saved) setXttsVoice(saved);
     });
@@ -250,6 +268,31 @@ const SettingsScreen: React.FC = () => {
         }
         rvs.send('xtts_list_voices' as any, {});
       }
+
+      // Diagnostic-Voice-Wechsel → lokale App-Stimme auf den neuen Default zuruecksetzen.
+      // Zusaetzlich Preload triggern, damit der User weiss wann's geladen ist.
+      if (message.type === ('config' as any)) {
+        const newVoice = ((message.payload as any).xttsVoice as string) ?? '';
+        setXttsVoice(newVoice);
+        AsyncStorage.setItem('aria_xtts_voice', newVoice);
+        if (newVoice) {
+          setLoadingVoice(newVoice);
+        }
+      }
+
+      // XTTS-Bridge meldet: Stimme fertig geladen
+      if (message.type === ('voice_ready' as any)) {
+        const v = ((message.payload as any).voice as string) ?? '';
+        const err = (message.payload as any).error as string | undefined;
+        const ms = (message.payload as any).loadMs as number | undefined;
+        setLoadingVoice(null);
+        if (err) {
+          ToastAndroid.show(`Stimme "${v}" konnte nicht geladen werden: ${err}`, ToastAndroid.LONG);
+        } else {
+          const suffix = ms ? ` (${(ms / 1000).toFixed(1)}s)` : '';
+          ToastAndroid.show(`Stimme "${v || 'Standard'}" bereit${suffix}`, ToastAndroid.SHORT);
+        }
+      }
     });
 
     return () => {
@@ -318,6 +361,13 @@ const SettingsScreen: React.FC = () => {
   const selectVoice = useCallback((voiceName: string) => {
     setXttsVoice(voiceName);
     AsyncStorage.setItem('aria_xtts_voice', voiceName);
+    // Preload nur fuer Custom-Voices — "Standard" braucht keinen Ladevorgang
+    if (voiceName) {
+      setLoadingVoice(voiceName);
+      rvs.send('voice_preload' as any, { voice: voiceName, source: 'app' });
+    } else {
+      setLoadingVoice(null);
+    }
   }, []);
 
   const deleteVoice = useCallback((name: string) => {
@@ -527,6 +577,42 @@ const SettingsScreen: React.FC = () => {
           />
         </View>
 
+        {ttsEnabled && (
+          <View style={{marginTop: 20}}>
+            <Text style={styles.toggleLabel}>Puffer vor Wiedergabestart</Text>
+            <Text style={styles.toggleHint}>
+              Wie viel Audio gesammelt wird bevor die Wiedergabe startet.
+              Hoeher = robuster gegen Render-Pausen, aber mehr Startverzoegerung.
+              Default: {TTS_PREROLL_DEFAULT_SEC.toFixed(1)}s.
+            </Text>
+            <View style={styles.prerollRow}>
+              <TouchableOpacity
+                style={styles.prerollButton}
+                onPress={() => {
+                  const next = Math.max(TTS_PREROLL_MIN_SEC, Math.round((ttsPrerollSec - 0.5) * 10) / 10);
+                  setTtsPrerollSec(next);
+                  AsyncStorage.setItem(TTS_PREROLL_STORAGE_KEY, String(next));
+                }}
+                disabled={ttsPrerollSec <= TTS_PREROLL_MIN_SEC}
+              >
+                <Text style={styles.prerollButtonText}>−0.5</Text>
+              </TouchableOpacity>
+              <Text style={styles.prerollValue}>{ttsPrerollSec.toFixed(1)} s</Text>
+              <TouchableOpacity
+                style={styles.prerollButton}
+                onPress={() => {
+                  const next = Math.min(TTS_PREROLL_MAX_SEC, Math.round((ttsPrerollSec + 0.5) * 10) / 10);
+                  setTtsPrerollSec(next);
+                  AsyncStorage.setItem(TTS_PREROLL_STORAGE_KEY, String(next));
+                }}
+                disabled={ttsPrerollSec >= TTS_PREROLL_MAX_SEC}
+              >
+                <Text style={styles.prerollButtonText}>+0.5</Text>
+              </TouchableOpacity>
+            </View>
+          </View>
+        )}
+
         {ttsEnabled && (
           <View style={{marginTop: 20}}>
             <Text style={styles.toggleLabel}>Stimme (geraetelokal)</Text>
@@ -561,7 +647,10 @@ const SettingsScreen: React.FC = () => {
                     </Text>
                     <Text style={styles.voiceRowMeta}>{(v.size / 1024).toFixed(0)} KB</Text>
                   </TouchableOpacity>
-                  {xttsVoice === v.name && <Text style={styles.voiceRowCheck}>{'\u2713'}</Text>}
+                  {loadingVoice === v.name && (
+                    <ActivityIndicator size="small" color="#0096FF" style={{marginRight: 8}} />
+                  )}
+                  {xttsVoice === v.name && loadingVoice !== v.name && <Text style={styles.voiceRowCheck}>{'\u2713'}</Text>}
                   <TouchableOpacity onPress={() => deleteVoice(v.name)} style={styles.voiceRowDelete}>
                     <Text style={styles.voiceRowDeleteIcon}>X</Text>
                   </TouchableOpacity>
@@ -1118,6 +1207,34 @@ const styles = StyleSheet.create({
   bottomSpacer: {
     height: 40,
   },
+
+  prerollRow: {
+    flexDirection: 'row',
+    alignItems: 'center',
+    justifyContent: 'center',
+    marginTop: 12,
+    gap: 16,
+  },
+  prerollButton: {
+    backgroundColor: '#2A2A3E',
+    paddingHorizontal: 18,
+    paddingVertical: 10,
+    borderRadius: 8,
+    minWidth: 72,
+    alignItems: 'center',
+  },
+  prerollButtonText: {
+    color: '#FFFFFF',
+    fontSize: 16,
+    fontWeight: '600',
+  },
+  prerollValue: {
+    color: '#FFFFFF',
+    fontSize: 20,
+    fontWeight: '700',
+    minWidth: 80,
+    textAlign: 'center',
+  },
 });
 
 export default SettingsScreen;

android/src/services/audio.ts

@@ -9,6 +9,7 @@
 import { Platform, PermissionsAndroid, NativeModules } from 'react-native';
 import Sound from 'react-native-sound';
 import RNFS from 'react-native-fs';
+import AsyncStorage from '@react-native-async-storage/async-storage';
 import AudioRecorderPlayer, {
   AudioEncoderAndroidType,
   AudioSourceAndroidType,
@@ -41,7 +42,7 @@ const { AudioFocus, PcmStreamPlayer } = NativeModules as {
     release: () => Promise<boolean>;
   };
   PcmStreamPlayer?: {
-    start: (sampleRate: number, channels: number) => Promise<boolean>;
+    start: (sampleRate: number, channels: number, prerollSeconds: number) => Promise<boolean>;
     writeChunk: (base64Pcm: string) => Promise<boolean>;
     end: () => Promise<boolean>;
     stop: () => Promise<boolean>;
@@ -80,6 +81,26 @@ const VAD_SPEECH_MIN_MS = 500;         // ms Sprache bevor Aufnahme zaehlt — l
 // Max-Dauer einer Aufnahme in Gespraechsmodus (Notbremse gegen Runaway-Loops)
 const MAX_RECORDING_MS = 30000;
 
+// Pre-Roll: Wie lange Audio im AudioTrack-Buffer liegt bevor play() startet.
+// Einstellbar via Diagnostic/Settings (Key: aria_tts_preroll_sec).
+export const TTS_PREROLL_DEFAULT_SEC = 3.5;
+export const TTS_PREROLL_MIN_SEC = 1.0;
+export const TTS_PREROLL_MAX_SEC = 6.0;
+export const TTS_PREROLL_STORAGE_KEY = 'aria_tts_preroll_sec';
+
+async function loadPrerollSec(): Promise<number> {
+  try {
+    const raw = await AsyncStorage.getItem(TTS_PREROLL_STORAGE_KEY);
+    if (raw != null) {
+      const n = parseFloat(raw);
+      if (isFinite(n) && n >= TTS_PREROLL_MIN_SEC && n <= TTS_PREROLL_MAX_SEC) {
+        return n;
+      }
+    }
+  } catch {}
+  return TTS_PREROLL_DEFAULT_SEC;
+}
+
 // --- Audio-Service ---
 
 class AudioService {
@@ -373,8 +394,9 @@ class AudioService {
       this.pcmBuffer = [];
       this.pcmBytesCollected = 0;
       if (!silent) {
+        const prerollSec = await loadPrerollSec();
         try {
-          await PcmStreamPlayer!.start(sampleRate, channels);
+          await PcmStreamPlayer!.start(sampleRate, channels, prerollSec);
         } catch (err) {
           console.error('[Audio] PcmStreamPlayer.start fehlgeschlagen:', err);
           this.pcmStreamActive = false;

bridge/aria_bridge.py

@@ -150,6 +150,15 @@ def _small_range_to_words(m):
     return f"{_num_to_words_de(a)} bis {_num_to_words_de(b)}"
 
 
+def _decimal_to_words(m):
+    """'0.1' / '0,1' → 'null komma eins', '1,25' → 'eins komma zwei fuenf'."""
+    int_part = int(m.group(1))
+    dec_part = m.group(2)
+    int_word = _num_to_words_de(int_part) if 0 <= int_part <= 59 else str(int_part)
+    dec_words = " ".join(_num_to_words_de(int(d)) for d in dec_part)
+    return f"{int_word} komma {dec_words}"
+
+
 _UNIT_WORDS = [
     (r'\bTB\b', 'Terabyte'),
     (r'\bGB\b', 'Gigabyte'),
@@ -236,6 +245,11 @@ def clean_text_for_tts(text: str) -> str:
     # Kleine Zahlen-Bereiche ohne "Uhr": "5-6" → "fuenf bis sechs"
     t = _re_tts.sub(r'\b(\d{1,2})\s*[-–]\s*(\d{1,2})\b', _small_range_to_words, t)
 
+    # Dezimalzahlen: "0.1" / "0,5" / "1,25" → "null komma eins" / "null komma fuenf" / ...
+    # Muss vor "Zahl+Einheit" laufen, sonst frisst die Unit-Regel den Nachkommaanteil.
+    # Lookahead verhindert Match auf IP-artigen Strings wie 192.168.1.1.
+    t = _re_tts.sub(r'\b(\d+)[.,](\d+)(?![.,\d])', _decimal_to_words, t)
+
     # Zahlen + Einheit: "22GB" → "22 Gigabyte" (Leerzeichen einfuegen)
     t = _re_tts.sub(r'(\d+)([A-Za-z]{1,4})\b', r'\1 \2', t)
 
@@ -243,6 +257,12 @@ def clean_text_for_tts(text: str) -> str:
     for pat, repl in _UNIT_WORDS:
         t = _re_tts.sub(pat, repl, t)
 
+    # Generisches Buchstabieren: alle verbleibenden 2-5-Zeichen-Grossbuchstaben-Woerter
+    # (XTTS, USB, DNS, JSON, HTML, ...) → "X T T S". Laeuft NACH der expliziten Liste,
+    # damit TTS/GPU/... schon aufgeloest sind. "WLAN"-artige, die als Wort gesprochen
+    # werden, koennen bei Bedarf explizit in _UNIT_WORDS uebersteuert werden.
+    t = _re_tts.sub(r'\b([A-Z]{2,5})\b', lambda m: " ".join(m.group(1)), t)
+
     # Anfuehrungszeichen
     t = _re_tts.sub(r'["""„`]', '', t)
 
@@ -305,8 +325,16 @@ class STTEngine:
             Erkannter Text oder leerer String.
         """
         if self.model is None:
-            logger.error("Whisper-Modell nicht initialisiert")
-            return ""
+            # Lazy-Load: normalerweise laeuft STT remote auf der Gamebox.
+            # Erst wenn das Fallback hier zuschlaegt, laden wir lokal.
+            logger.info("Lokales Whisper-Fallback — Modell wird nachgeladen...")
+            try:
+                self.initialize()
+            except Exception:
+                logger.exception("Lokales Whisper konnte nicht geladen werden")
+                return ""
+            if self.model is None:
+                return ""
 
         try:
             # Audio als float32 normalisieren
@@ -503,6 +531,9 @@ class ARIABridge:
         # Wird fuer die direkt folgende ARIA-Antwort genutzt und dann zurueckgesetzt.
         # So kann jedes Geraet seine bevorzugte Stimme bekommen (pro Request).
         self._next_voice_override: Optional[str] = None
+        # STT-Requests die aktuell auf Antwort von der whisper-bridge (Gamebox) warten.
+        # requestId → Future mit dem Text (oder None bei Fehler).
+        self._pending_stt: dict[str, asyncio.Future] = {}
 
     def initialize(self) -> None:
         """Initialisiert alle Komponenten.
@@ -515,8 +546,9 @@ class ARIABridge:
         logger.info("ARIA Voice Bridge startet...")
         logger.info("=" * 50)
 
-        # STT IMMER laden — verarbeitet Audio von der App (braucht kein Sounddevice)
-        self.stt_engine.initialize()
+        # STT wird standardmaessig von der whisper-bridge (Gamebox) erledigt.
+        # Lokales Whisper ist nur Fallback und wird lazy geladen wenn remote nicht
+        # antwortet. Das spart RAM auf der VM und Startup-Zeit.
 
         # Audio-Hardware pruefen (fuer lokales Mikro/Lautsprecher)
         self.audio_available = False
@@ -1175,11 +1207,16 @@ class ARIABridge:
                 changed = True
             if "whisperModel" in payload:
                 new_model = payload["whisperModel"]
-                if new_model and new_model != self.stt_engine.model_size:
-                    logger.info("[rvs] Whisper-Modell Wechsel: %s -> %s (laedt...)", self.stt_engine.model_size, new_model)
-                    loop = asyncio.get_event_loop()
-                    if await loop.run_in_executor(None, self.stt_engine.reload, new_model):
-                        changed = True
+                allowed = {"tiny", "base", "small", "medium", "large-v3"}
+                if new_model in allowed and new_model != self.stt_engine.model_size:
+                    # Merken und mitschicken an whisper-bridge (Gamebox).
+                    # Lokales Modell wird NICHT geladen — nur das Fallback braucht's,
+                    # und das passiert erst on-demand wenn Remote nicht antwortet.
+                    logger.info("[rvs] Whisper-Modell → %s (nur Config; Modell laedt Gamebox)",
+                                new_model)
+                    self.stt_engine.model_size = new_model
+                    self.stt_engine.model = None
+                    changed = True
             # Persistent speichern in Shared Volume
             if changed:
                 try:
@@ -1339,22 +1376,111 @@ class ARIABridge:
                         mime_type, duration_ms, len(audio_b64) // 1365)
             asyncio.create_task(self._process_app_audio(audio_b64, mime_type))
 
+        elif msg_type == "stt_response":
+            # Antwort der whisper-bridge auf unseren stt_request
+            request_id = payload.get("requestId", "")
+            future = self._pending_stt.get(request_id)
+            if future is None or future.done():
+                return
+            error = payload.get("error", "")
+            if error:
+                logger.warning("[rvs] stt_response Fehler: %s", error)
+                future.set_result(None)
+            else:
+                text = payload.get("text", "")
+                stt_ms = payload.get("sttMs", 0)
+                model = payload.get("model", "?")
+                logger.info("[rvs] Remote-STT OK (%s, %dms): '%s'", model, stt_ms, (text or "")[:80])
+                future.set_result(text)
+            return
+
         else:
             logger.debug("[rvs] Unbekannter Typ: %s", msg_type)
 
+    # STT-Orchestrierung: zuerst Remote (Gamebox), Fallback lokal.
+    # Timeout grosszuegig gewaehlt, damit auch ein erstmaliger Modell-Load
+    # auf der Gamebox (bis ~30s bei large-v3) durchgeht.
+    _STT_REMOTE_TIMEOUT_S = 45.0
+
     async def _process_app_audio(self, audio_b64: str, mime_type: str) -> None:
-        """Decodiert App-Audio (Base64 AAC/MP4), konvertiert zu 16kHz PCM, STT, sendet an core."""
+        """App-Audio → STT → aria-core. Primaer via whisper-bridge (RVS), Fallback lokal."""
+        # Erst Remote versuchen
+        text = await self._stt_remote(audio_b64, mime_type)
+        if text is None:
+            # Remote hat nicht geantwortet → lokales Whisper
+            logger.warning("[rvs] Remote-STT nicht verfuegbar — Fallback auf lokales Whisper")
+            text = await self._stt_local(audio_b64, mime_type)
+        if text is None:
+            return
+
+        if text.strip():
+            logger.info("[rvs] STT Ergebnis: '%s'", text[:80])
+            # ERST an aria-core senden (wichtigster Schritt)
+            await self.send_to_core(text, source="app-voice")
+            # STT-Text an RVS senden (fuer Anzeige in App + Diagnostic)
+            # sender="stt" damit Bridge es ignoriert (kein Loop)
+            try:
+                await self._send_to_rvs({
+                    "type": "chat",
+                    "payload": {
+                        "text": text,
+                        "sender": "stt",
+                    },
+                    "timestamp": int(asyncio.get_event_loop().time() * 1000),
+                })
+            except Exception as e:
+                logger.warning("[rvs] STT-Text konnte nicht an RVS gesendet werden: %s", e)
+        else:
+            logger.info("[rvs] Keine Sprache erkannt — ignoriert")
+
+    async def _stt_remote(self, audio_b64: str, mime_type: str) -> Optional[str]:
+        """Schickt Audio an die whisper-bridge und wartet auf stt_response.
+
+        Rueckgabe:
+            str  — erkannter Text (kann leer sein)
+            None — Remote-STT nicht erreichbar oder Fehler/Timeout (→ Fallback)
+        """
+        if self.ws_rvs is None:
+            return None
+
+        request_id = str(uuid.uuid4())
+        loop = asyncio.get_event_loop()
+        future: asyncio.Future = loop.create_future()
+        self._pending_stt[request_id] = future
+
+        try:
+            await self._send_to_rvs({
+                "type": "stt_request",
+                "payload": {
+                    "requestId": request_id,
+                    "audio": audio_b64,
+                    "mimeType": mime_type,
+                    "model": getattr(self.stt_engine, "model_size", "small"),
+                    "language": getattr(self.stt_engine, "language", "de"),
+                },
+                "timestamp": int(loop.time() * 1000),
+            })
+            return await asyncio.wait_for(future, timeout=self._STT_REMOTE_TIMEOUT_S)
+        except asyncio.TimeoutError:
+            logger.warning("[rvs] Remote-STT Timeout (%.0fs)", self._STT_REMOTE_TIMEOUT_S)
+            return None
+        except Exception as e:
+            logger.warning("[rvs] Remote-STT Fehler: %s", e)
+            return None
+        finally:
+            self._pending_stt.pop(request_id, None)
+
+    async def _stt_local(self, audio_b64: str, mime_type: str) -> Optional[str]:
+        """Lokales Whisper-Fallback: FFmpeg → float32 → stt_engine.transcribe."""
         loop = asyncio.get_event_loop()
         tmp_in = None
         tmp_out = None
         try:
-            # Base64 → temp-Datei
             ext = ".mp4" if "mp4" in mime_type else ".wav" if "wav" in mime_type else ".ogg"
             tmp_in = tempfile.NamedTemporaryFile(suffix=ext, delete=False)
             tmp_in.write(base64.b64decode(audio_b64))
             tmp_in.close()
 
-            # FFmpeg: beliebiges Format → 16kHz mono PCM (raw float32)
             tmp_out = tempfile.NamedTemporaryFile(suffix=".raw", delete=False)
             tmp_out.close()
 
@@ -1369,45 +1495,21 @@ class ARIABridge:
             )
             if result.returncode != 0:
                 logger.error("[rvs] FFmpeg Fehler: %s", result.stderr.decode()[:200])
-                return
+                return None
 
-            # PCM lesen → numpy float32
             audio_data = np.fromfile(tmp_out.name, dtype=np.float32)
             if len(audio_data) == 0:
                 logger.warning("[rvs] Leere Audio-Daten nach Konvertierung")
-                return
+                return None
 
             duration_s = len(audio_data) / 16000.0
-            logger.info("[rvs] Audio konvertiert: %.1fs, %d samples", duration_s, len(audio_data))
-
-            # STT
-            text = await loop.run_in_executor(None, self.stt_engine.transcribe, audio_data)
-
-            if text.strip():
-                logger.info("[rvs] STT Ergebnis: '%s'", text[:80])
-                # ERST an aria-core senden (wichtigster Schritt)
-                await self.send_to_core(text, source="app-voice")
-                # STT-Text an RVS senden (fuer Anzeige in App + Diagnostic)
-                # sender="stt" damit Bridge es ignoriert (kein Loop)
-                try:
-                    await self._send_to_rvs({
-                        "type": "chat",
-                        "payload": {
-                            "text": text,
-                            "sender": "stt",
-                        },
-                        "timestamp": int(asyncio.get_event_loop().time() * 1000),
-                    })
-                except Exception as e:
-                    logger.warning("[rvs] STT-Text konnte nicht an RVS gesendet werden: %s", e)
-            else:
-                logger.info("[rvs] Keine Sprache erkannt — ignoriert")
-
+            logger.info("[rvs] Lokal-STT: %.1fs Audio, model=%s", duration_s, self.stt_engine.model_size)
+            return await loop.run_in_executor(None, self.stt_engine.transcribe, audio_data)
         except Exception:
-            logger.exception("[rvs] Audio-Verarbeitung fehlgeschlagen")
+            logger.exception("[rvs] Lokales STT fehlgeschlagen")
+            return None
         finally:
-            # Temp-Dateien aufraeumen
-            for f in [tmp_in, tmp_out]:
+            for f in (tmp_in, tmp_out):
                 if f:
                     try:
                         os.unlink(f.name)

diagnostic/index.html

@@ -438,13 +438,14 @@
         </div>
 
         <!-- XTTS Stimme -->
-        <div style="display:flex;align-items:center;gap:12px;margin-bottom:12px;">
+        <div style="display:flex;align-items:center;gap:12px;margin-bottom:6px;">
           <label style="color:#8888AA;font-size:12px;">XTTS Stimme:</label>
           <select id="diag-xtts-voice" onchange="sendVoiceConfig()" style="background:#1E1E2E;color:#fff;border:1px solid #2A2A3E;border-radius:6px;padding:6px 10px;font-size:13px;">
             <option value="">Standard (XTTS Default)</option>
           </select>
           <button class="btn secondary" onclick="loadXTTSVoices()" style="padding:4px 10px;font-size:11px;">Laden</button>
         </div>
+        <div id="voice-status" style="font-size:11px;min-height:14px;margin-bottom:12px;color:#8888AA;"></div>
 
         <!-- Gecloned Stimmen — Liste mit Loeschen -->
         <div id="xtts-voice-list" style="margin-bottom:12px;"></div>
@@ -851,6 +852,25 @@
           return;
         }
 
+        if (msg.type === 'voice_ready') {
+          const v = msg.payload?.voice || '';
+          const err = msg.payload?.error;
+          const ms = msg.payload?.loadMs;
+          const statusEl = document.getElementById('voice-status');
+          if (statusEl) {
+            if (err) {
+              statusEl.textContent = `⚠️ Stimme "${v}" Fehler: ${err}`;
+              statusEl.style.color = '#FF3B30';
+            } else {
+              statusEl.textContent = `✅ Stimme "${v || 'Standard'}" bereit${ms ? ` (${(ms/1000).toFixed(1)}s)` : ''}`;
+              statusEl.style.color = '#34C759';
+            }
+            setTimeout(() => { if (statusEl) statusEl.textContent = ''; }, 5000);
+          }
+          addLog('info', 'xtts', err ? `Voice "${v}": ${err}` : `Voice "${v || 'Standard'}" bereit`);
+          return;
+        }
+
         if (msg.type === 'watchdog') {
           const colors = { warning: '#FFD60A', fixing: '#FF9500', fixed: '#34C759', error: '#FF3B30' };
           const color = colors[msg.status] || '#FFD60A';
@@ -1551,6 +1571,11 @@
       const xttsVoice = document.getElementById('diag-xtts-voice').value;
       const whisperModel = document.getElementById('diag-whisper-model').value;
       send({ action: 'send_voice_config', ttsEnabled, xttsVoice, whisperModel });
+      const statusEl = document.getElementById('voice-status');
+      if (statusEl && xttsVoice) {
+        statusEl.textContent = `⏳ Stimme "${xttsVoice}" wird geladen...`;
+        statusEl.style.color = '#FFD60A';
+      }
     }
 
     // ── Passwort-Feld Anzeigen/Verbergen ─────────────────────

diagnostic/server.js

@@ -626,6 +626,17 @@ function connectRVS(forcePlain) {
         // Mode-Broadcast von der Bridge → an Browser-Clients weiterreichen
         log("info", "rvs", `Mode-Broadcast: ${msg.payload?.mode} (${msg.payload?.name})`);
         broadcast({ type: "mode", payload: msg.payload });
+      } else if (msg.type === "voice_ready") {
+        // XTTS-Bridge meldet Stimme fertig geladen → an Browser durchreichen
+        const v = msg.payload?.voice || "";
+        const err = msg.payload?.error;
+        const ms = msg.payload?.loadMs;
+        if (err) {
+          log("warn", "rvs", `Voice-Ready Fehler fuer "${v}": ${err}`);
+        } else {
+          log("info", "rvs", `Voice "${v || "default"}" geladen${ms ? ` in ${(ms/1000).toFixed(1)}s` : ""}`);
+        }
+        broadcast({ type: "voice_ready", payload: msg.payload });
       } else {
         log("debug", "rvs", `Nachricht: ${JSON.stringify(msg).slice(0, 150)}`);
       }

rvs/server.js

@@ -19,6 +19,8 @@ const ALLOWED_TYPES = new Set([
   "agent_activity", "cancel_request",
   "audio_pcm",
   "xtts_delete_voice",
+  "voice_preload", "voice_ready",
+  "stt_request", "stt_response",
 ]);
 
 // Token-Raum: token -> { clients: Set<ws> }

xtts/bridge.js

@@ -69,6 +69,18 @@ function connectRVS(forcePlain) {
         await handleListVoices();
       } else if (msg.type === "xtts_delete_voice") {
         await handleDeleteVoice(msg.payload);
+      } else if (msg.type === "voice_preload") {
+        await handleVoicePreload(msg.payload);
+      } else if (msg.type === "config") {
+        // Diagnostic hat globale Voice gewechselt → Preload damit der naechste
+        // Render ohne Ladewartezeit startet + alle Clients "voice_ready" sehen
+        const v = msg.payload && msg.payload.xttsVoice;
+        if (v && v !== lastDiagnosticVoice) {
+          lastDiagnosticVoice = v;
+          await handleVoicePreload({ voice: v, source: "diagnostic" });
+        } else if (!v) {
+          lastDiagnosticVoice = "";
+        }
       }
     } catch (err) {
       log(`Fehler: ${err.message}`);
@@ -95,12 +107,35 @@ function connectRVS(forcePlain) {
 
 // ── TTS Request Handler ─────────────────────────────
 
+/**
+ * Linearer Fade-In auf einen base64-PCM-Chunk (s16le).
+ * Mascht XTTS-Warmup-Glitches am Anfang eines Renders.
+ */
+function applyFadeIn(base64Pcm, sampleRate, channels, fadeMs) {
+  const buf = Buffer.from(base64Pcm, "base64");
+  const totalSamples = buf.length / 2; // s16le
+  const fadeSamples = Math.min(
+    Math.floor((fadeMs / 1000) * sampleRate) * channels,
+    totalSamples
+  );
+  for (let i = 0; i < fadeSamples; i++) {
+    const sample = buf.readInt16LE(i * 2);
+    const gain = i / fadeSamples;
+    buf.writeInt16LE(Math.round(sample * gain), i * 2);
+  }
+  return buf.toString("base64");
+}
+
 // ── TTS-Queue ──────────────────────────────────────
 // XTTS verarbeitet Requests sequenziell, damit Streams sich nicht ueberlappen.
 // Ohne Queue wuerden parallele Requests parallel streamen → App bekommt
 // interleaved PCM-Chunks aus zwei Rendern → klingt wie Chaos.
 let ttsQueue = Promise.resolve();
 
+// Merkt sich die letzte in Diagnostic gewaehlte Voice, damit wir nicht bei jedem
+// config-Broadcast (auch ohne Aenderung) einen Preload triggern.
+let lastDiagnosticVoice = "";
+
 function handleTTSRequest(payload) {
   ttsQueue = ttsQueue.then(() => _runTTSRequest(payload)).catch(err => {
     log(`TTS-Queue Fehler: ${err.message}`);
@@ -132,20 +167,38 @@ async function _runTTSRequest(payload) {
   log(`TTS-Request (streaming): "${cleanText.slice(0, 80)}..." (${cleanText.length} chars, voice: ${voice || "default"})`);
 
   try {
-    const voiceSample = voice ? path.join(VOICES_DIR, `${voice}.wav`) : null;
-    const hasCustomVoice = voiceSample && fs.existsSync(voiceSample);
+    // Im local-Mode erwartet daswer123 XTTS speaker_wav als Basename (ohne .wav,
+    // ohne Pfad) — der Server prefixt EXAMPLE_FOLDER selbst. Wir checken hier
+    // nur das physische File ab um Warnungen zu loggen; runter ans API geht
+    // nur der Name.
+    const voiceFilePath = voice ? path.join(VOICES_DIR, `${voice}.wav`) : null;
+    const hasCustomVoice = voiceFilePath && fs.existsSync(voiceFilePath);
+    const speakerName = hasCustomVoice ? voice : "";
+    if (voice && !hasCustomVoice) {
+      log(`WARNUNG: Voice "${voice}" angefordert, aber ${voiceFilePath} existiert nicht — nehme Default`);
+    } else if (hasCustomVoice) {
+      log(`Voice "${voice}" verwendet (speaker_wav="${speakerName}")`);
+    }
 
     let chunkIndex = 0;
     let pcmMeta = null;
+    let firstChunkSeen = false;
 
     const onChunk = (pcmBase64, meta) => {
       if (!pcmMeta) pcmMeta = meta;
+      let outBase64 = pcmBase64;
+      // Fade-In auf den ersten Chunk — maskiert XTTS-Warmup-Glitches
+      // (autoregressiver Generator hat am Anfang wenig Kontext → Artefakte).
+      if (!firstChunkSeen && pcmBase64) {
+        firstChunkSeen = true;
+        outBase64 = applyFadeIn(pcmBase64, meta.sampleRate, meta.channels, 120);
+      }
       sendToRVS({
         type: "audio_pcm",
         payload: {
           requestId: requestId || "",
           messageId: messageId || "",
-          base64: pcmBase64,
+          base64: outBase64,
           format: "pcm_s16le",
           sampleRate: meta.sampleRate,
           channels: meta.channels,
@@ -163,7 +216,7 @@ async function _runTTSRequest(payload) {
       await streamXTTSAsPCM(
         cleanText,
         language || "de",
-        hasCustomVoice ? voiceSample : null,
+        speakerName,
         onChunk,
       );
     } catch (streamErr) {
@@ -171,7 +224,7 @@ async function _runTTSRequest(payload) {
       await streamXTTSBatch(
         cleanText,
         language || "de",
-        hasCustomVoice ? voiceSample : null,
+        speakerName,
         onChunk,
       );
     }
@@ -216,15 +269,14 @@ function streamXTTSAsPCM(text, language, speakerWav, onPcmChunk) {
   return new Promise((resolve, reject) => {
     // Wichtig: speaker_wav MUSS als Query-Key dabei sein (Pydantic required) —
     // auch bei default-voice mit leerem Wert. Sonst gibt's HTTP 422.
-    // stream_chunk_size=100: Kompromiss zwischen first-audio-latency und
-    // gap-risk. Bei RTX 3060 (RTF 1.48) ~3s bis erster Audio, Chunks gross
-    // genug dass der AudioTrack-Buffer (128KB ≈ 2.7s) zwischen Chunks nicht
-    // leerlauft.
+    // stream_chunk_size=250: grosse Chunks = wenige Chunk-Grenzen = wenig
+    // Render-Artefakte. daswer123 erzeugt an Chunk-Boundaries haeufig Glitches
+    // in den Worten die ueber die Grenze gehen. Hoehere Latenz ist OK.
     const qs = new URLSearchParams();
     qs.set("text", text);
     qs.set("language", language || "de");
     qs.set("speaker_wav", speakerWav || "");
-    qs.set("stream_chunk_size", "100");
+    qs.set("stream_chunk_size", "250");
 
     const url = new URL(XTTS_API_URL);
     const fullPath = `/tts_stream?${qs.toString()}`;
@@ -434,6 +486,63 @@ async function handleDeleteVoice(payload) {
 
 // ── Voice List Handler ──────────────────────────────
 
+/**
+ * Preload einer Stimme — rendert stumm ein kurzes Dummy-Audio, damit XTTS
+ * die Speaker-Latents laedt und der naechste echte Request ohne Wartezeit
+ * loslegen kann. Broadcastet "voice_ready" wenn fertig (oder mit error).
+ */
+async function handleVoicePreload(payload) {
+  const voice = (payload && payload.voice) || "";
+  const source = (payload && payload.source) || "unknown";
+  const requestId = (payload && payload.requestId) || "";
+  log(`Voice-Preload angefordert: "${voice}" (source=${source})`);
+
+  try {
+    let speakerName = "";
+    if (voice) {
+      const voiceFilePath = path.join(VOICES_DIR, `${voice}.wav`);
+      if (!fs.existsSync(voiceFilePath)) {
+        sendToRVS({
+          type: "voice_ready",
+          payload: { voice, requestId, error: "voice-file-not-found" },
+          timestamp: Date.now(),
+        });
+        log(`Preload abgebrochen: ${voiceFilePath} existiert nicht`);
+        return;
+      }
+      speakerName = voice;
+    }
+
+    // Dummy-Request via Queue — damit sich Preload nicht mit echtem TTS ueberholt.
+    const t0 = Date.now();
+    await new Promise((resolve, reject) => {
+      ttsQueue = ttsQueue.then(async () => {
+        try {
+          await streamXTTSAsPCM("ja.", "de", speakerName, () => {});
+          resolve();
+        } catch (err) {
+          reject(err);
+        }
+      }).catch(reject);
+    });
+    const ms = Date.now() - t0;
+    log(`Voice "${voice || "default"}" geladen in ${ms}ms`);
+
+    sendToRVS({
+      type: "voice_ready",
+      payload: { voice, requestId, loadMs: ms },
+      timestamp: Date.now(),
+    });
+  } catch (err) {
+    log(`Voice-Preload Fehler: ${err.message}`);
+    sendToRVS({
+      type: "voice_ready",
+      payload: { voice, requestId, error: err.message.slice(0, 200) },
+      timestamp: Date.now(),
+    });
+  }
+}
+
 async function handleListVoices() {
   try {
     const files = fs.existsSync(VOICES_DIR)

xtts/docker-compose.yml

@@ -58,5 +58,37 @@ services:
       - RVS_TOKEN=${RVS_TOKEN}
     restart: unless-stopped
 
+  # ─── Whisper STT (GPU) ────────────────────────
+  # Faster-Whisper auf der Gamebox statt auf der VM (CPU) —
+  # deutlich schneller. Verbindet sich selbst per WebSocket an
+  # den RVS und nimmt dort stt_request Nachrichten der aria-bridge
+  # entgegen, antwortet mit stt_response. Laedt das Modell beim
+  # Start vor; auf Config-Broadcasts (Diagnostic → whisperModel)
+  # wird zur Laufzeit hot-swapped.
+  whisper-bridge:
+    build: ./whisper
+    container_name: aria-whisper-bridge
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    environment:
+      - RVS_HOST=${RVS_HOST}
+      - RVS_PORT=${RVS_PORT:-443}
+      - RVS_TLS=${RVS_TLS:-true}
+      - RVS_TLS_FALLBACK=${RVS_TLS_FALLBACK:-true}
+      - RVS_TOKEN=${RVS_TOKEN}
+      - WHISPER_MODEL=${WHISPER_MODEL:-small}
+      - WHISPER_DEVICE=${WHISPER_DEVICE:-cuda}
+      - WHISPER_COMPUTE_TYPE=${WHISPER_COMPUTE_TYPE:-float16}
+      - WHISPER_LANGUAGE=${WHISPER_LANGUAGE:-de}
+    volumes:
+      - whisper-models:/root/.cache/huggingface # Model-Cache persistieren
+    restart: unless-stopped
+
 volumes:
   xtts-models:
+  whisper-models:

xtts/whisper/Dockerfile

@@ -0,0 +1,14 @@
+FROM nvidia/cuda:12.2.2-cudnn8-runtime-ubuntu22.04
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    python3 python3-pip ffmpeg \
+    && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /app
+
+COPY requirements.txt .
+RUN pip3 install --no-cache-dir -r requirements.txt
+
+COPY bridge.py .
+
+CMD ["python3", "bridge.py"]

xtts/whisper/bridge.py

@@ -0,0 +1,247 @@
+#!/usr/bin/env python3
+"""
+ARIA Whisper Bridge — laeuft auf der Gamebox (RTX 3060).
+
+Empfaengt stt_request via RVS → FFmpeg-Konvertierung → faster-whisper auf GPU
+→ sendet stt_response zurueck an die aria-bridge.
+
+Env:
+  RVS_HOST, RVS_PORT, RVS_TLS, RVS_TLS_FALLBACK, RVS_TOKEN
+  WHISPER_MODEL          Default: small
+  WHISPER_DEVICE         Default: cuda
+  WHISPER_COMPUTE_TYPE   Default: float16
+  WHISPER_LANGUAGE       Default: de
+"""
+import asyncio
+import base64
+import json
+import logging
+import os
+import subprocess
+import sys
+import tempfile
+import time
+from typing import Optional
+
+import numpy as np
+import websockets
+from faster_whisper import WhisperModel
+
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s [%(levelname)s] %(message)s",
+    datefmt="%H:%M:%S",
+)
+logger = logging.getLogger("whisper-bridge")
+
+RVS_HOST = os.getenv("RVS_HOST", "").strip()
+RVS_PORT = int(os.getenv("RVS_PORT", "443"))
+RVS_TLS = os.getenv("RVS_TLS", "true").lower() == "true"
+RVS_TLS_FALLBACK = os.getenv("RVS_TLS_FALLBACK", "true").lower() == "true"
+RVS_TOKEN = os.getenv("RVS_TOKEN", "").strip()
+
+WHISPER_MODEL = os.getenv("WHISPER_MODEL", "small")
+WHISPER_DEVICE = os.getenv("WHISPER_DEVICE", "cuda")
+WHISPER_COMPUTE_TYPE = os.getenv("WHISPER_COMPUTE_TYPE", "float16")
+WHISPER_LANGUAGE = os.getenv("WHISPER_LANGUAGE", "de")
+
+ALLOWED_MODELS = {"tiny", "base", "small", "medium", "large-v3"}
+
+
+class WhisperRunner:
+    """Haelt das Whisper-Modell. Hot-Swap bei Konfig-Wechsel via ensure_loaded()."""
+
+    def __init__(self) -> None:
+        self.model_size: str = WHISPER_MODEL
+        self.model: Optional[WhisperModel] = None
+        self._lock = asyncio.Lock()
+
+    def _load_blocking(self, size: str) -> None:
+        logger.info(
+            "Lade Whisper '%s' (device=%s, compute=%s)",
+            size, WHISPER_DEVICE, WHISPER_COMPUTE_TYPE,
+        )
+        t0 = time.time()
+        self.model = WhisperModel(
+            size, device=WHISPER_DEVICE, compute_type=WHISPER_COMPUTE_TYPE,
+        )
+        self.model_size = size
+        logger.info("Whisper '%s' geladen in %.1fs", size, time.time() - t0)
+
+    async def ensure_loaded(self, desired_size: str) -> None:
+        if desired_size not in ALLOWED_MODELS:
+            logger.warning("Ungueltiges Whisper-Modell '%s' — nutze %s", desired_size, WHISPER_MODEL)
+            desired_size = WHISPER_MODEL
+        async with self._lock:
+            if self.model is not None and self.model_size == desired_size:
+                return
+            loop = asyncio.get_event_loop()
+            await loop.run_in_executor(None, self._load_blocking, desired_size)
+
+    async def transcribe(self, audio: np.ndarray, language: str) -> tuple[str, float]:
+        if self.model is None:
+            return "", 0.0
+
+        def _run():
+            segments, info = self.model.transcribe(
+                audio, language=language, beam_size=5, vad_filter=True,
+            )
+            text = " ".join(seg.text.strip() for seg in segments)
+            return text, info.duration
+
+        loop = asyncio.get_event_loop()
+        return await loop.run_in_executor(None, _run)
+
+
+def ffmpeg_to_float32(audio_b64: str, mime_type: str) -> np.ndarray:
+    """Dekodiert beliebiges Audio-Format → 16kHz mono float32 PCM."""
+    if "mp4" in mime_type or "m4a" in mime_type or "aac" in mime_type:
+        ext = ".mp4"
+    elif "wav" in mime_type:
+        ext = ".wav"
+    elif "ogg" in mime_type or "opus" in mime_type:
+        ext = ".ogg"
+    else:
+        ext = ".bin"
+
+    in_fh = tempfile.NamedTemporaryFile(suffix=ext, delete=False)
+    try:
+        in_fh.write(base64.b64decode(audio_b64))
+        in_fh.close()
+        out_path = in_fh.name + ".raw"
+        cmd = ["ffmpeg", "-y", "-i", in_fh.name, "-ar", "16000", "-ac", "1", "-f", "f32le", out_path]
+        result = subprocess.run(cmd, capture_output=True, timeout=30)
+        if result.returncode != 0:
+            logger.error("FFmpeg Fehler: %s", result.stderr.decode(errors="replace")[:300])
+            return np.zeros(0, dtype=np.float32)
+        try:
+            return np.fromfile(out_path, dtype=np.float32)
+        finally:
+            try:
+                os.unlink(out_path)
+            except OSError:
+                pass
+    finally:
+        try:
+            os.unlink(in_fh.name)
+        except OSError:
+            pass
+
+
+async def _send(ws, mtype: str, payload: dict) -> None:
+    try:
+        await ws.send(json.dumps({
+            "type": mtype,
+            "payload": payload,
+            "timestamp": int(time.time() * 1000),
+        }))
+    except Exception as e:
+        logger.warning("Send fehlgeschlagen (%s): %s", mtype, e)
+
+
+async def handle_stt_request(ws, payload: dict, runner: WhisperRunner) -> None:
+    request_id = payload.get("requestId", "")
+    audio_b64 = payload.get("audio", "")
+    mime_type = payload.get("mimeType", "audio/mp4")
+    model = payload.get("model") or WHISPER_MODEL
+    language = payload.get("language") or WHISPER_LANGUAGE
+
+    if not audio_b64:
+        await _send(ws, "stt_response", {"requestId": request_id, "error": "no-audio"})
+        return
+
+    try:
+        t_load = time.time()
+        await runner.ensure_loaded(model)
+        load_ms = int((time.time() - t_load) * 1000)
+
+        audio = ffmpeg_to_float32(audio_b64, mime_type)
+        if audio.size == 0:
+            await _send(ws, "stt_response", {"requestId": request_id, "error": "ffmpeg-failed"})
+            return
+        duration_s = len(audio) / 16000.0
+        logger.info("STT-Request: %.1fs Audio, model=%s, lang=%s", duration_s, runner.model_size, language)
+
+        t_stt = time.time()
+        text, detected_duration = await runner.transcribe(audio, language)
+        stt_ms = int((time.time() - t_stt) * 1000)
+
+        logger.info("STT-Ergebnis (%dms): '%s'", stt_ms, text[:100])
+
+        await _send(ws, "stt_response", {
+            "requestId": request_id,
+            "text": text.strip(),
+            "durationS": duration_s,
+            "sttMs": stt_ms,
+            "loadMs": load_ms,
+            "model": runner.model_size,
+        })
+    except Exception as e:
+        logger.exception("STT-Request fehlgeschlagen")
+        await _send(ws, "stt_response", {
+            "requestId": request_id,
+            "error": str(e)[:200],
+        })
+
+
+async def run_loop(runner: WhisperRunner) -> None:
+    # Modell vorab laden damit erste Anfrage flott ist
+    try:
+        await runner.ensure_loaded(WHISPER_MODEL)
+    except Exception as e:
+        logger.error("Preload fehlgeschlagen: %s — Fortsetzung, wird bei erstem Request nachgeladen", e)
+
+    use_tls = RVS_TLS
+    retry_s = 2
+    tls_fallback_tried = False
+
+    while True:
+        scheme = "wss" if use_tls else "ws"
+        url = f"{scheme}://{RVS_HOST}:{RVS_PORT}/ws?token={RVS_TOKEN}"
+        masked = url.replace(RVS_TOKEN, "***") if RVS_TOKEN else url
+        try:
+            logger.info("Verbinde zu RVS: %s", masked)
+            async with websockets.connect(url, ping_interval=20, ping_timeout=10) as ws:
+                logger.info("RVS verbunden")
+                retry_s = 2
+                tls_fallback_tried = False
+                async for raw in ws:
+                    try:
+                        msg = json.loads(raw)
+                    except Exception:
+                        continue
+                    mtype = msg.get("type", "")
+                    payload = msg.get("payload", {}) or {}
+
+                    if mtype == "stt_request":
+                        asyncio.create_task(handle_stt_request(ws, payload, runner))
+                    elif mtype == "config":
+                        new_model = payload.get("whisperModel")
+                        if new_model and new_model != runner.model_size:
+                            logger.info("Config-Broadcast: Whisper-Modell → %s", new_model)
+                            asyncio.create_task(runner.ensure_loaded(new_model))
+                    # andere Types (chat, heartbeat, ...) einfach ignorieren
+        except Exception as e:
+            logger.warning("Verbindung verloren: %s", e)
+            if use_tls and RVS_TLS_FALLBACK and not tls_fallback_tried:
+                logger.info("TLS-Verbindung fehlgeschlagen — Fallback auf ws://")
+                use_tls = False
+                tls_fallback_tried = True
+                continue
+            await asyncio.sleep(min(retry_s, 30))
+            retry_s = min(retry_s * 2, 30)
+
+
+async def main() -> None:
+    if not RVS_HOST:
+        logger.error("RVS_HOST ist nicht gesetzt — Abbruch")
+        sys.exit(1)
+    runner = WhisperRunner()
+    await run_loop(runner)
+
+
+if __name__ == "__main__":
+    try:
+        asyncio.run(main())
+    except KeyboardInterrupt:
+        sys.exit(0)

xtts/whisper/requirements.txt

@@ -0,0 +1,3 @@
+faster-whisper==1.0.3
+websockets>=12.0
+numpy>=1.24