release: bump version to 0.2.3.4

VoxtralProcessor.apply_transcription_request -> load_audio_as braucht librosa zum Einlesen der Audiodatei. Fehlte im Image.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

ai-box/bootstrap.sh

@@ -11,6 +11,8 @@
 #
 # Optionen:
 #   --up            am Ende den xtts-Stack starten (Default-Profil, OHNE voxtral)
+#   --upgrade-driver  NVIDIA-Treiber aus trixie-backports (modernes CUDA fuer Voxtral).
+#                     Danach REBOOT noetig. Default-Bootstrap laesst 550 unangetastet.
 #   --token <TOK>   RVS_TOKEN in xtts/.env eintragen (alternativ: env RVS_TOKEN=...)
 #   --rvs-host <H>  RVS_HOST setzen (Default aus .env.example)
 #
@@ -21,11 +23,13 @@ set -euo pipefail
 
 # ── CLI ──
 DO_UP=0
+DO_UPGRADE_DRIVER=0
 RVS_TOKEN_ARG="${RVS_TOKEN:-}"
 RVS_HOST_ARG="${RVS_HOST:-}"
 while [[ $# -gt 0 ]]; do
   case "$1" in
     --up) DO_UP=1; shift ;;
+    --upgrade-driver) DO_UPGRADE_DRIVER=1; shift ;;
     --token) RVS_TOKEN_ARG="${2:-}"; shift 2 ;;
     --rvs-host) RVS_HOST_ARG="${2:-}"; shift 2 ;;
     -h|--help) grep '^#' "$0" | sed 's/^# \{0,1\}//'; exit 0 ;;
@@ -90,6 +94,45 @@ fi
 [[ $ENABLED_ANY -eq 0 ]] && ok "non-free schon aktiv"
 apt-get update -qq   # immer neu einlesen, damit der Kandidat sicher da ist
 
+# ── 2b. (opt-in) Treiber-Upgrade via trixie-backports — fuer Voxtral/modernes CUDA ──
+# Der Trixie-Standardtreiber (550, CUDA 12.4) ist zu alt fuer den modernen Stack
+# (torch 2.13, vLLM). Backports bringt einen neueren, apt-verwalteten Treiber.
+# NUR mit --upgrade-driver, damit ein laufendes Setup nicht ungewollt angefasst wird.
+if [[ "$DO_UPGRADE_DRIVER" -eq 1 ]]; then
+  step "NVIDIA-Treiber-Upgrade (trixie-backports)"
+  BP_FILE="/etc/apt/sources.list.d/backports.sources"
+  if ! grep -rqs "trixie-backports" /etc/apt/sources.list /etc/apt/sources.list.d/ 2>/dev/null; then
+    cat > "$BP_FILE" <<'EOF'
+Types: deb
+URIs: http://deb.debian.org/debian
+Suites: trixie-backports
+Components: main contrib non-free non-free-firmware
+Signed-By: /usr/share/keyrings/debian-archive-keyring.gpg
+EOF
+    ok "trixie-backports hinzugefuegt"
+  else
+    ok "trixie-backports bereits aktiv"
+  fi
+  apt-get update
+  apt-get install -y linux-headers-amd64 || true
+  apt-get install -y "linux-headers-$(uname -r)" || true
+  if apt-get install -y -t trixie-backports nvidia-driver; then
+    dkms autoinstall >/dev/null 2>&1 || true
+    NEWV="$(dpkg-query -W -f='${Version}' nvidia-driver 2>/dev/null || echo '?')"
+    ok "nvidia-driver aus backports installiert: ${NEWV}"
+    echo
+    echo -e "${c_y}==> REBOOT noetig, damit der neue Treiber laedt:${c_0}"
+    echo -e "${c_y}    sudo reboot   &&   danach:  cd ai-box && sudo ./bootstrap.sh --up --token <TOKEN>${c_0}"
+    echo -e "${c_y}    (nvidia-smi zeigt dann die neue Version + CUDA-Level)${c_0}"
+    exit 0
+  else
+    warn "backports-Install fehlgeschlagen — 550er bleibt aktiv."
+    warn "Alternative fuer den neuesten Treiber: NVIDIAs CUDA-Repo fuer debian13"
+    warn "  (developer.download.nvidia.com/compute/cuda/repos/debian13/x86_64) → Paket 'cuda-drivers'."
+    die "Treiber-Upgrade nicht moeglich — siehe oben."
+  fi
+fi
+
 # ── 3. NVIDIA-Treiber ──
 step "NVIDIA-Treiber"
 if nvidia-smi >/dev/null 2>&1; then

android/android/app/build.gradle

@@ -79,8 +79,8 @@ android {
         applicationId "com.ariacockpit"
         minSdkVersion rootProject.ext.minSdkVersion
         targetSdkVersion rootProject.ext.targetSdkVersion
-        versionCode 20302
-        versionName "0.2.3.2"
+        versionCode 20304
+        versionName "0.2.3.4"
         // Fallback fuer Libraries mit Product Flavors
         missingDimensionStrategy 'react-native-camera', 'general'
     }

android/package.json

@@ -1,6 +1,6 @@
 {
   "name": "aria-cockpit",
-  "version": "0.2.3.2",
+  "version": "0.2.3.4",
   "private": true,
   "scripts": {
     "android": "react-native run-android",

android/src/screens/ChatScreen.tsx

@@ -50,7 +50,7 @@ import VoiceButton from '../components/VoiceButton';
 import FileUpload, { FileData } from '../components/FileUpload';
 import CameraUpload, { PhotoData } from '../components/CameraUpload';
 import MessageText from '../components/MessageText';
-import { loadConvWindowMs, loadTtsSpeed, TTS_SPEED_DEFAULT, loadSttEndpointMs } from '../services/audio';
+import { loadConvWindowMs, loadTtsSpeed, TTS_SPEED_DEFAULT, loadSttEndpointMs, loadMaxRecordingMs } from '../services/audio';
 import Geolocation from '@react-native-community/geolocation';
 
 // --- Typen ---
@@ -93,6 +93,9 @@ interface ChatMessage {
    *  gespiegelt damit wir die EXAKT richtige Placeholder-Bubble ersetzen,
    *  auch wenn mehrere Aufnahmen parallel offen sind. */
   audioRequestId?: string;
+  /** Laenge der Sprachaufnahme in Sekunden (aus dem stt_endpoint) — fuer die
+   *  Dauer-Anzeige an der Voice-Bubble. */
+  durationS?: number;
   /** Skill-Created-Bubble: ARIA hat einen neuen Skill angelegt */
   skillCreated?: {
     name: string;
@@ -184,6 +187,14 @@ function stripSystemHints(text: string): string {
   }
   return out;
 }
+/** Sekunden → "M:SS" fuer die Sprachnachricht-Dauer. */
+function formatDur(sec: number): string {
+  const s = Math.max(0, Math.round(sec));
+  const m = Math.floor(s / 60);
+  const r = s % 60;
+  return `${m}:${r.toString().padStart(2, '0')}`;
+}
+
 const DEFAULT_ATTACHMENT_DIR = `${RNFS.DocumentDirectoryPath}/chat_attachments`;
 const STORAGE_PATH_KEY = 'aria_attachment_storage_path';
 
@@ -1668,7 +1679,9 @@ const ChatScreen: React.FC = () => {
         location: location || null,
         noSpeechTimeoutMs: windowMs,
         endpointMs: await loadSttEndpointMs(),
-        hardCapMs: 60000,
+        // Notbremse 5 min (nicht 1 min) — der Stille-Endpoint beendet normale
+        // Turns eh sofort; der Cap darf lange Diktate nicht mitten drin kappen.
+        hardCapMs: await loadMaxRecordingMs(),
         projectId: focusedProjectIdRef.current,
       });
       import('../services/logger').then(m => m.reportAppDebug('wake.cb', `startStreamingRecording returned ok=${ok}`)).catch(()=>{});
@@ -1696,6 +1709,13 @@ const ChatScreen: React.FC = () => {
       if (ev.text && ev.text.trim()) {
         console.log('[Chat] STT-Endpoint: %r (reason=%s, %dms, %.1fs Audio)',
                     ev.text.slice(0, 80), ev.reason, ev.sttMs, ev.durationS);
+        // Aufnahme-Dauer an die passende Voice-Bubble haengen (Anzeige). Der
+        // spaetere STT-Text-Update spreadet die Message, die Dauer bleibt.
+        if (ev.audioRequestId && typeof ev.durationS === 'number' && ev.durationS > 0) {
+          const dur = ev.durationS;
+          setMessages(prev => prev.map(m =>
+            m.audioRequestId === ev.audioRequestId ? { ...m, durationS: dur } : m));
+        }
         // Wenn passive lauschend: User hat tatsaechlich was gesagt → uebergang
         // zu 'conversing' damit der normale Flow greift (TTS, resume, etc.)
         if (wakeWordService.getState() === 'listening') {
@@ -1771,7 +1791,8 @@ const ChatScreen: React.FC = () => {
         location: location || null,
         noSpeechTimeoutMs: windowMs,
         endpointMs: await loadSttEndpointMs(),
-        hardCapMs: 60000,
+        // Notbremse 5 min (s.o.) — lange Diktate nicht bei 1 min abschneiden.
+        hardCapMs: await loadMaxRecordingMs(),
         projectId: focusedProjectIdRef.current,
       });
       if (ok) {
@@ -2255,7 +2276,7 @@ const ChatScreen: React.FC = () => {
       // die Session auch app-seitig haben wir +2s Toleranz.
       noSpeechTimeoutMs: 0,
       endpointMs: await loadSttEndpointMs(),
-      hardCapMs: 300000,
+      hardCapMs: await loadMaxRecordingMs(),
       projectId: focusedProjectIdRef.current,
     });
     if (!ok) {
@@ -2647,6 +2668,16 @@ const ChatScreen: React.FC = () => {
                   {att.serverPath ? '(tippen zum Laden)' : '(nicht verfuegbar)'}
                 </Text>
               </TouchableOpacity>
+            ) : att.type === 'audio' ? (
+              <View style={styles.attachmentFile}>
+                <Text style={styles.attachmentFileIcon}>{'🎙'}</Text>
+                <Text style={styles.attachmentFileName} numberOfLines={1}>
+                  {att.name || 'Sprachaufnahme'}
+                </Text>
+                {typeof item.durationS === 'number' && item.durationS > 0 ? (
+                  <Text style={styles.attachmentFileSize}>{formatDur(item.durationS)}</Text>
+                ) : null}
+              </View>
             ) : (
               <TouchableOpacity
                 style={styles.attachmentFile}

android/src/screens/SettingsScreen.tsx

@@ -67,6 +67,10 @@ import {
   CONV_WINDOW_MIN_SEC,
   CONV_WINDOW_MAX_SEC,
   CONV_WINDOW_STORAGE_KEY,
+  STT_ENDPOINT_DEFAULT_MS,
+  STT_ENDPOINT_MIN_MS,
+  STT_ENDPOINT_MAX_MS,
+  STT_ENDPOINT_STORAGE_KEY,
   MAX_RECORDING_DEFAULT_SEC,
   MAX_RECORDING_MIN_SEC,
   MAX_RECORDING_MAX_SEC,
@@ -195,6 +199,9 @@ const SettingsScreen: React.FC = () => {
   const [ttsEnabled, setTtsEnabled] = useState(true);
   const [ttsPrerollSec, setTtsPrerollSec] = useState<number>(TTS_PREROLL_DEFAULT_SEC);
   const [vadSilenceSec, setVadSilenceSec] = useState<number>(VAD_SILENCE_DEFAULT_SEC);
+  // Aktive Streaming-Pausen-Toleranz (STT_ENDPOINT) — der "Stille-Toleranz"-Regler
+  // steuert jetzt DIESEN Wert (der alte vadSilenceSec war der tote Legacy-dB-Pfad).
+  const [sttEndpointSec, setSttEndpointSec] = useState<number>(STT_ENDPOINT_DEFAULT_MS / 1000);
   const [convWindowSec, setConvWindowSec] = useState<number>(CONV_WINDOW_DEFAULT_SEC);
   const [maxRecordingSec, setMaxRecordingSec] = useState<number>(MAX_RECORDING_DEFAULT_SEC);
   // null = automatisch (adaptive Baseline), sonst manueller dB-Override
@@ -306,6 +313,14 @@ const SettingsScreen: React.FC = () => {
         }
       }
     });
+    AsyncStorage.getItem(STT_ENDPOINT_STORAGE_KEY).then(saved => {
+      if (saved != null) {
+        const n = parseInt(saved, 10);
+        if (isFinite(n) && n >= STT_ENDPOINT_MIN_MS && n <= STT_ENDPOINT_MAX_MS) {
+          setSttEndpointSec(n / 1000);
+        }
+      }
+    });
     AsyncStorage.getItem(MAX_RECORDING_STORAGE_KEY).then(saved => {
       if (saved != null) {
         const n = parseFloat(saved);
@@ -1655,30 +1670,30 @@ const SettingsScreen: React.FC = () => {
         <Text style={styles.toggleHint}>
           Wie lange du eine Sprechpause machen darfst, bevor die Aufnahme
           automatisch beendet und gesendet wird. Hoeher = mehr Zeit zum
-          Nachdenken; niedriger = schnelleres Senden.
-          Default: {VAD_SILENCE_DEFAULT_SEC.toFixed(1)}s.
+          Nachdenken (z.B. im Auto); niedriger = schnelleres Senden.
+          Default: {(STT_ENDPOINT_DEFAULT_MS / 1000).toFixed(1)}s.
         </Text>
         <View style={styles.prerollRow}>
           <TouchableOpacity
             style={styles.prerollButton}
             onPress={() => {
-              const next = Math.max(VAD_SILENCE_MIN_SEC, Math.round((vadSilenceSec - 0.5) * 10) / 10);
-              setVadSilenceSec(next);
-              AsyncStorage.setItem(VAD_SILENCE_STORAGE_KEY, String(next));
+              const next = Math.max(STT_ENDPOINT_MIN_MS / 1000, Math.round((sttEndpointSec - 0.5) * 10) / 10);
+              setSttEndpointSec(next);
+              AsyncStorage.setItem(STT_ENDPOINT_STORAGE_KEY, String(Math.round(next * 1000)));
             }}
-            disabled={vadSilenceSec <= VAD_SILENCE_MIN_SEC}
+            disabled={sttEndpointSec <= STT_ENDPOINT_MIN_MS / 1000}
           >
             <Text style={styles.prerollButtonText}>−0.5</Text>
           </TouchableOpacity>
-          <Text style={styles.prerollValue}>{vadSilenceSec.toFixed(1)} s</Text>
+          <Text style={styles.prerollValue}>{sttEndpointSec.toFixed(1)} s</Text>
           <TouchableOpacity
             style={styles.prerollButton}
             onPress={() => {
-              const next = Math.min(VAD_SILENCE_MAX_SEC, Math.round((vadSilenceSec + 0.5) * 10) / 10);
-              setVadSilenceSec(next);
-              AsyncStorage.setItem(VAD_SILENCE_STORAGE_KEY, String(next));
+              const next = Math.min(STT_ENDPOINT_MAX_MS / 1000, Math.round((sttEndpointSec + 0.5) * 10) / 10);
+              setSttEndpointSec(next);
+              AsyncStorage.setItem(STT_ENDPOINT_STORAGE_KEY, String(Math.round(next * 1000)));
             }}
-            disabled={vadSilenceSec >= VAD_SILENCE_MAX_SEC}
+            disabled={sttEndpointSec >= STT_ENDPOINT_MAX_MS / 1000}
           >
             <Text style={styles.prerollButtonText}>+0.5</Text>
           </TouchableOpacity>
@@ -1749,56 +1764,10 @@ const SettingsScreen: React.FC = () => {
           </TouchableOpacity>
         </View>
 
-        <View style={{flexDirection: 'row', alignItems: 'center', marginTop: 24, gap: 8}}>
-          <Text style={styles.toggleLabel}>Stille-Pegel (dB)</Text>
-          <TouchableOpacity onPress={() => setShowVadInfo(true)} style={styles.infoBtn}>
-            <Text style={styles.infoBtnText}>i</Text>
-          </TouchableOpacity>
-        </View>
-        <Text style={styles.toggleHint}>
-          Welcher Mikro-Pegel als "Stille" gilt. Standard: automatisch (Baseline aus
-          den ersten 500ms). Manuell setzen wenn Auto nicht zuverlaessig greift.
-        </Text>
-        <View style={styles.prerollRow}>
-          <TouchableOpacity
-            style={styles.prerollButton}
-            onPress={() => {
-              const next = vadSilenceDb == null
-                ? VAD_SILENCE_DB_DEFAULT - 1
-                : Math.max(VAD_SILENCE_DB_MIN, vadSilenceDb - 1);
-              setVadSilenceDb(next);
-              AsyncStorage.setItem(VAD_SILENCE_DB_OVERRIDE_KEY, String(next));
-            }}
-          >
-            <Text style={styles.prerollButtonText}>−1</Text>
-          </TouchableOpacity>
-          <Text style={styles.prerollValue}>
-            {vadSilenceDb == null ? 'auto' : `${vadSilenceDb} dB`}
-          </Text>
-          <TouchableOpacity
-            style={styles.prerollButton}
-            onPress={() => {
-              const next = vadSilenceDb == null
-                ? VAD_SILENCE_DB_DEFAULT + 1
-                : Math.min(VAD_SILENCE_DB_MAX, vadSilenceDb + 1);
-              setVadSilenceDb(next);
-              AsyncStorage.setItem(VAD_SILENCE_DB_OVERRIDE_KEY, String(next));
-            }}
-          >
-            <Text style={styles.prerollButtonText}>+1</Text>
-          </TouchableOpacity>
-        </View>
-        {vadSilenceDb != null && (
-          <TouchableOpacity
-            onPress={() => {
-              setVadSilenceDb(null);
-              AsyncStorage.removeItem(VAD_SILENCE_DB_OVERRIDE_KEY);
-            }}
-            style={{alignSelf: 'center', marginTop: 8, paddingVertical: 6, paddingHorizontal: 12}}
-          >
-            <Text style={{color: '#0096FF', fontSize: 13}}>↻ Auf automatisch zuruecksetzen</Text>
-          </TouchableOpacity>
-        )}
+        {/* "Stille-Pegel (dB)"-Regler entfernt: der aktive Streaming-STT nutzt
+            einen adaptiven Rausch-Boden (automatisch), ein manueller dB-Wert war
+            wirkungslos. Rauschen-als-Wort verhindert das STT-Modell selbst
+            (no_speech_prob-Filter), nicht die dB-Schwelle. */}
       </View>
 
       <Modal

xtts/docker-compose.yml

@@ -63,6 +63,7 @@ services:
   whisper-bridge:
     build: ./whisper
     container_name: aria-whisper-bridge
+    profiles: ["whisper"]        # Fallback-STT — startet nur mit --profile whisper
     deploy:
       resources:
         reservations:
@@ -138,55 +139,32 @@ services:
       - LLM_TIMEOUT_SEC=${LLM_TIMEOUT_SEC:-600}
     restart: unless-stopped
 
-  # ─── Voxtral STT (GPU, Realtime) — PROFIL "voxtral" ───────────
-  # Ersetzt whisper als STT sobald die 24-GB-Karte da ist. Startet NUR mit
-  #   docker compose --profile voxtral up -d
-  # (sonst kollidiert es mit whisper — beide wuerden stt_* beantworten).
-  #
-  # ⚠️ VRAM: Voxtral-Mini-4B-Realtime-2602 braucht >=16 GB (BF16, laut vLLM-
-  #    Rezept keine Quant). Laeuft NICHT auf der 3060 (12 GB) — erst 24-GB-Karte.
-  # ⚠️ vLLM: Version >=0.20.0 noetig. Entrypoint/Serve-Form beim ersten Lauf
-  #    gegen das offizielle Rezept pruefen (siehe voxtral/README.md).
-  voxtral-vllm:
-    image: vllm/vllm-openai:latest
-    container_name: aria-voxtral-vllm
-    profiles: ["voxtral"]
+  # ─── Voxtral STT-3B (Transformers, GPU) — DEFAULT-STT ─────────
+  # Ersetzt whisper als STT. Laeuft auf Treiber 550/CUDA 12.4 (torch cu124, KEIN
+  # Treiber-Upgrade noetig). Modell Voxtral-Mini-3B-2507 (~9 GB bf16) → GPU 1.
+  # Startet bei jedem `docker compose up -d`. Whisper ist der opt-in Fallback
+  # (Profil "whisper") — beide zusammen wuerden stt_* doppelt beantworten, also
+  # immer nur EINEN STT laufen lassen.
+  voxtral-bridge:
+    build: ./voxtral
+    container_name: aria-voxtral-bridge
     deploy:
       resources:
         reservations:
           devices:
             - driver: nvidia
-              count: 1
+              device_ids: ["1"]      # 12-GB-Karte (STT-3B ~9 GB); GPU 0 (8 GB) bleibt fuer F5/LLM
               capabilities: [gpu]
     volumes:
       - ./hf-cache:/root/.cache/huggingface      # gleicher Modell-Cache wie whisper/f5
-    environment:
-      - VLLM_DISABLE_COMPILE_CACHE=1
-      - HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}
-    # Serve-Command aus dem vLLM-Rezept (Voxtral-Mini-4B-Realtime-2602).
-    command:
-      - --model
-      - mistralai/Voxtral-Mini-4B-Realtime-2602
-      - --tokenizer-mode
-      - mistral
-      - --compilation_config
-      - '{"cudagraph_mode":"PIECEWISE"}'
-    restart: unless-stopped
-
-  # ─── Voxtral-Bridge — RVS <-> vLLM-Realtime-WS (CPU-Glue) ─────
-  voxtral-bridge:
-    build: ./voxtral
-    container_name: aria-voxtral-bridge
-    profiles: ["voxtral"]
-    depends_on:
-      - voxtral-vllm
     environment:
       - RVS_HOST=${RVS_HOST}
       - RVS_PORT=${RVS_PORT:-443}
       - RVS_TLS=${RVS_TLS:-true}
       - RVS_TLS_FALLBACK=${RVS_TLS_FALLBACK:-true}
       - RVS_TOKEN=${RVS_TOKEN}
-      - VOXTRAL_VLLM_URL=ws://voxtral-vllm:8000/v1/realtime
-      - VOXTRAL_MODEL=mistralai/Voxtral-Mini-4B-Realtime-2602
+      - VOXTRAL_MODEL=mistralai/Voxtral-Mini-3B-2507
       - VOXTRAL_LANGUAGE=${WHISPER_LANGUAGE:-de}
+      - HUGGING_FACE_HUB_TOKEN=${HF_TOKEN:-}   # falls das Modell HF-gated ist
+      - PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True   # weniger VRAM-Fragmentierung
     restart: unless-stopped

xtts/voxtral/Dockerfile

@@ -1,14 +1,26 @@
-# Voxtral-BRIDGE (nicht das Modell!) — leichte CPU-Glue zwischen RVS und dem
-# vLLM-Realtime-Server. Das eigentliche Voxtral-Modell laeuft im Container
-# `voxtral-vllm` (GPU, vllm/vllm-openai). Deshalb hier kein CUDA-Base noetig.
-FROM python:3.11-slim
+# Voxtral-STT-3B Bridge (Transformers). Laeuft auf Treiber 550/CUDA 12.4 via
+# torch cu124 — KEIN Treiber-Upgrade noetig (gleicher Trick wie f5tts).
+# Modell: Voxtral-Mini-3B-2507 (~9 GB in bf16) → passt auf die 12-GB-Karte (GPU 1).
+FROM nvidia/cuda:12.2.2-cudnn8-runtime-ubuntu22.04
 
+ENV DEBIAN_FRONTEND=noninteractive
 ENV PYTHONUNBUFFERED=1
 WORKDIR /app
 
-COPY requirements.txt .
-RUN pip install --no-cache-dir -r requirements.txt
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    python3 python3-pip ffmpeg git \
+    && rm -rf /var/lib/apt/lists/*
 
-COPY bridge.py ./
+# torch FEST auf cu124 (Treiber 550 = CUDA 12.4). 2.6.0 ist der neueste cu124-Build;
+# torch 2.7+ gibt es nur fuer cu126+ und braeuchte einen neueren Treiber.
+RUN pip3 install --no-cache-dir torch==2.6.0 torchaudio==2.6.0 \
+    --index-url https://download.pytorch.org/whl/cu124
+
+COPY requirements.txt .
+# Constraint haelt transformers/mistral-common davon ab, torch wieder hochzuziehen.
+RUN printf 'torch==2.6.0\ntorchaudio==2.6.0\n' > /tmp/torch-constraint.txt && \
+    pip3 install --no-cache-dir -c /tmp/torch-constraint.txt -r requirements.txt
+
+COPY bridge.py .
 
 CMD ["python3", "bridge.py"]

xtts/voxtral/README.md

@@ -1,70 +1,47 @@
-# Voxtral-STT-Satellit (M0.3)
+# Voxtral-STT-3B-Satellit (Transformers)
 
-Streaming-STT via **Voxtral-Mini-4B-Realtime-2602** (Mistral, Apache 2.0) auf
-vLLM. Ersetzt whisper als STT — genauer (~5,9 % WER vs 7,4 % FLEURS) und mit
-echtem Realtime-Streaming. Deutsch ist in den 13 Sprachen abgedeckt.
+Streaming-STT via **Voxtral-Mini-3B-2507** (Mistral, Apache 2.0) über
+🤗 Transformers. Ersetzt whisper als STT — genauer, und **ohne Treiber-Upgrade**:
+läuft auf dem Trixie-Standardtreiber (550/CUDA 12.4) via **torch 2.6.0+cu124**
+(derselbe Trick wie bei f5tts).
 
-Zwei Container:
-- **`voxtral-vllm`** — das Modell auf vLLM (GPU). Exponiert die Realtime-WS-API.
-- **`voxtral-bridge`** — CPU-Glue: RVS ⇄ vLLM-Realtime-WS. Macht das Endpointing
-  selbst (adaptiver Rausch-Boden-VAD, identisch zur whisper-Bridge / M0.1).
+- Modell ~9 GB (bf16) → **GPU 1** (die 12-GB-Karte; per Compose gepinnt).
+- **F5-TTS + LLM** bleiben auf GPU 0 (8 GB).
+- Arbeitsweise = chunked wie whisper: live PCM → alle ~1 s transkribieren
+  (Partials) → **adaptiver Endpointer** (Rausch-Boden-VAD + semantische
+  Stagnation, aus M0.1) feuert `stt_endpoint`. RVS-Protokoll identisch → drop-in.
 
-## ⚠️ Hardware-Realität — läuft NICHT auf der 3060
-
-Das Realtime-Modell braucht laut [vLLM-Rezept](https://recipes.vllm.ai/mistralai/Voxtral-Mini-4B-Realtime-2602)
-**≥ 16 GB VRAM (BF16, keine Quant)**. Die RTX 3060 hat 12 GB → passt nicht.
-
-- **Interim-gpubox (nur 3060):** whisper (mit M0.1-Fix) + F5-TTS bleiben aktiv.
-  Voxtral NICHT starten.
-- **Ab der 24-GB-Karte:** Voxtral-Profil hochziehen, whisper wird Fallback.
-
-Deshalb liegen beide Services hinter dem Compose-**Profil `voxtral`** und starten
-NUR explizit — sonst würden whisper *und* voxtral dieselben `stt_*`-Messages
-beantworten (Kollision).
-
-## Starten (erst wenn die 24-GB-Karte drin ist)
+## Starten (Profil `voxtral`)
 
 ```bash
 cd xtts
+docker compose stop whisper-bridge          # sonst beantworten beide stt_*
 docker compose --profile voxtral up -d --build
-docker logs -f aria-voxtral-vllm      # laedt Modell (mehrere GB, dauert)
-docker logs -f aria-voxtral-bridge    # "RVS verbunden" + service_status ready
-```
-Whisper vorher stoppen, damit nur eine STT-Engine antwortet:
-```bash
-docker compose stop whisper-bridge
+docker logs -f aria-voxtral-bridge          # Modell laedt (mehrere GB), dann "RVS verbunden"
 ```
+Zurück zu whisper: `docker compose --profile voxtral down && docker compose up -d whisper-bridge`.
 
-## ⚠️ Auf echter Hardware verifizieren (blind gebaut, kein Test hier)
+## ⚠️ Auf echter Hardware verifizieren (blind gebaut)
 
-1. **vLLM-Version ≥ 0.20.0** und die **Serve-Form**. Das Rezept nutzt
-   `vllm serve <model> …`. Falls das `vllm/vllm-openai`-Image einen anderen
-   Entrypoint hat, das `command:` in `docker-compose.yml` anpassen
-   (Rezept-Command steht dort als Kommentar).
-2. **Realtime-WS-Frames.** Die exakten Event-Namen sind in `bridge.py` ganz oben
-   als Konstanten gebündelt (`VLLM_SEND_APPEND`, `VLLM_DELTA_SUFFIXES`, …),
-   modelliert nach dem OpenAI-Realtime-Schema. Gegen das offizielle
-   **vLLM-Realtime-Client-Beispiel** prüfen und dort anpassen — nur an dieser
-   einen Stelle. Das Response-Handling ist bereits defensiv (mehrere Feldnamen).
-3. **Endpoint-URL/Port.** Default `ws://voxtral-vllm:8000/v1/realtime` — prüfen ob
-   vLLM auf 8000 lauscht und `/v1/realtime` registriert (Log-Zeile
-   `Route: /v1/realtime`).
-4. **Endpointing.** Voxtral liefert keine eigene VAD → unser adaptiver Endpointer
-   entscheidet (akustisch + semantisch am Delta-Wachstum). `endpointMs` kommt wie
-   bei whisper aus der App; `voiceFactor` per Session tunebar.
+1. **Transformers-API.** Die exakte Voxtral-Transkriptions-API ist in `bridge.py`
+   in **einer** Methode gekapselt (`VoxtralRunner._transcribe_blocking`), modelliert
+   nach der HF-Modelcard (`apply_transcription_request` → `generate` → `batch_decode`).
+   Beim ersten Lauf gegen die Modelcard prüfen und dort anpassen.
+2. **HF-Gating.** Ist `Voxtral-Mini-3B-2507` gated, `HF_TOKEN` in `xtts/.env` setzen
+   (wird als `HUGGING_FACE_HUB_TOKEN` durchgereicht).
+3. **VRAM/Tempo.** 3B in bf16 ~9 GB auf der 12-GB-Karte — Rest fürs KV-Cache. Ist die
+   Partial-Transkription (alle 1 s) zu schwer, `STREAM_TRANSCRIBE_INTERVAL_MS` hochsetzen.
+4. **torch-Konflikt.** Falls `transformers`/`mistral-common` beim Build torch>2.6
+   erzwingen, meldet der Constraint einen Konflikt → dann brauchen wir doch das
+   Treiber-Upgrade (`bootstrap.sh --upgrade-driver` via NVIDIA-CUDA-Repo) + cu126-torch.
 
 ## Protokoll (RVS, identisch zu whisper — drop-in)
-
 Rein: `stt_stream_start`, `stt_audio_chunk` (16 kHz mono s16le, base64), `stt_stream_end`.
-Raus: `stt_partial`, `stt_endpoint` (das Event, auf das aria-bridge horcht), `stt_stream_done`.
+Raus: `stt_partial`, `stt_endpoint`, `stt_stream_done`.
 
-## TTS-Hinweis
-
-Voxtral-**TTS** (Voice-Cloning) ist hier NICHT enthalten — braucht ebenfalls
-16 GB VRAM und ist ein eigener Bau. Bis zur 24-GB-Karte bleibt **F5-TTS** aktiv.
-Danach: eigener `voxtral-tts`-Satellit (separates Ticket).
+## TTS
+Bleibt **F5-TTS** (klingt gut, passt auf GPU 0). Voxtral-TTS bräuchte ~24 GB — separates Thema.
 
 ## Quellen
-- Rezept: https://recipes.vllm.ai/mistralai/Voxtral-Mini-4B-Realtime-2602
-- vLLM Speech-to-Text: https://docs.vllm.ai/en/latest/serving/online_serving/speech_to_text/
-- Modell: https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602
+- Modell: https://huggingface.co/mistralai/Voxtral-Mini-3B-2507
+- Transformers-Nutzung: HF-Modelcard (Voxtral) + `mistral-common`

xtts/voxtral/bridge.py

@@ -1,50 +1,41 @@
 #!/usr/bin/env python3
 """
-ARIA Voxtral Bridge — Streaming-STT via Voxtral-Mini-4B-Realtime-2602 (vLLM).
+ARIA Voxtral-STT-3B Bridge (Transformers) — Ersatz fuer whisper.
 
-Zwilling der whisper-Bridge, aber das Transkribieren macht NICHT faster-whisper
-im selben Prozess, sondern der separate vLLM-Realtime-Server (Container
-`voxtral-vllm`) ueber dessen WebSocket-API `/v1/realtime`. Diese Bridge ist
-reine Glue:
+Laeuft auf Treiber 550/CUDA 12.4 via torch cu124 (kein Treiber-Upgrade noetig).
+Modell: Voxtral-Mini-3B-2507 (bf16, ~9 GB) → GPU 1 (12 GB, per Compose gepinnt).
 
-  App ──(RVS: stt_stream_start / stt_audio_chunk / stt_stream_end)──▶ diese Bridge
-  diese Bridge ──(WS /v1/realtime: PCM16-b64 append)──▶ voxtral-vllm
-  voxtral-vllm ──(transcription.delta / transcription.done)──▶ diese Bridge
-  diese Bridge ──(RVS: stt_partial / stt_endpoint / stt_stream_done)──▶ App/aria-bridge
+Arbeitsweise = Zwilling der whisper-Bridge: App schickt live PCM-Chunks; wir
+transkribieren alle ~STREAM_TRANSCRIBE_INTERVAL_MS auf dem Ringbuffer (Partials)
+und feuern stt_endpoint, sobald der ADAPTIVE Endpointer (Rausch-Boden-VAD +
+semantische Stagnation, aus M0.1) "fertig" sagt. RVS-Wire-Protokoll identisch zu
+whisper → drop-in (die App merkt nur bessere Genauigkeit).
 
-Das RVS-Wire-Protokoll ist IDENTISCH zur whisper-Bridge (drop-in). Das
-Endpointing (wann hat der User aufgehoert zu sprechen) macht diese Bridge
-selbst — mit demselben ADAPTIVEN Rausch-Boden-Endpointer wie whisper (Voxtral
-Realtime liefert laut vLLM-Doku keine eigene VAD/„speaker done"-Semantik, nur
-transcription.delta/.done). Die akustische Energie messen wir auf unserer
-eigenen PCM-Kopie, die semantische Stagnation am Delta-Textwachstum.
-
-⚠️  HARDWARE: Voxtral-Mini-4B-Realtime-2602 braucht >=16 GB VRAM (BF16). Auf der
-    RTX 3060 (12 GB) laeuft es NICHT — erst auf der 24-GB-Karte. Bis dahin
-    bleibt die whisper-Bridge aktiv (Profil-gesteuert im docker-compose).
-
-⚠️  VERIFY-ON-FIRST-RUN: Die exakten vLLM-Realtime-FRAME-Namen (Audio-Append,
-    Delta/Done-Event-Typen) sind unten als Konstanten gebuendelt und nach dem
-    OpenAI-Realtime-Schema modelliert. Gegen das offizielle vLLM-Realtime-
-    Client-Beispiel pruefen und ggf. anpassen — sie stehen bewusst an EINER
-    Stelle. Response-Handling ist defensiv (mehrere moegliche Feldnamen).
+⚠️ VERIFY-ON-FIRST-RUN: Die exakte Transformers-Transkriptions-API von Voxtral
+   (apply_transcription_request / generate / decode) ist unten in EINER Methode
+   (VoxtralRunner._transcribe_blocking) gekapselt und nach dem HF-Modelcard-Muster
+   modelliert. Beim ersten echten Lauf gegen die Voxtral-Modelcard pruefen und dort
+   anpassen. Alles andere (RVS, Endpointer) ist bewaehrt.
 
 Env:
   RVS_HOST, RVS_PORT, RVS_TLS, RVS_TLS_FALLBACK, RVS_TOKEN
-  VOXTRAL_VLLM_URL     Default: ws://voxtral-vllm:8000/v1/realtime
-  VOXTRAL_MODEL        Default: mistralai/Voxtral-Mini-4B-Realtime-2602
-  VOXTRAL_LANGUAGE     Default: de
+  VOXTRAL_MODEL      Default: mistralai/Voxtral-Mini-3B-2507
+  VOXTRAL_LANGUAGE   Default: de
+  VOXTRAL_DEVICE     Default: cuda
+  STREAM_TRANSCRIBE_INTERVAL_MS  Default 1000 (3B ist schwerer als whisper-small)
 """
 import asyncio
 import base64
 import json
 import logging
 import os
+import tempfile
 import time
 from dataclasses import dataclass, field
 from typing import Optional
 
 import numpy as np
+import soundfile as sf
 import websockets
 
 logging.basicConfig(
@@ -60,29 +51,18 @@ 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()
 
-VOXTRAL_VLLM_URL = os.getenv("VOXTRAL_VLLM_URL", "ws://voxtral-vllm:8000/v1/realtime")
-VOXTRAL_MODEL = os.getenv("VOXTRAL_MODEL", "mistralai/Voxtral-Mini-4B-Realtime-2602")
+VOXTRAL_MODEL = os.getenv("VOXTRAL_MODEL", "mistralai/Voxtral-Mini-3B-2507")
 VOXTRAL_LANGUAGE = os.getenv("VOXTRAL_LANGUAGE", "de")
+VOXTRAL_DEVICE = os.getenv("VOXTRAL_DEVICE", "cuda")
 
-# ── vLLM-Realtime-Frames — HIER anpassen falls das Client-Beispiel abweicht ──
-# Senderichtung (wir → vLLM): PCM16-16kHz-mono base64 anhaengen + committen.
-VLLM_SEND_APPEND = "input_audio_buffer.append"   # {"type":..., "audio": "<b64>"}
-VLLM_SEND_COMMIT = "input_audio_buffer.commit"   # Buffer abschliessen
-VLLM_AUDIO_FIELD = "audio"
-# Empfangsrichtung (vLLM → wir): inkrementeller Text + final. Defensiv geprueft.
-VLLM_DELTA_SUFFIXES = ("transcription.delta",)   # msg["type"] endet hierauf
-VLLM_DONE_SUFFIXES = ("transcription.done", "transcription.completed")
-VLLM_DELTA_FIELDS = ("delta", "text", "transcription")   # eins davon traegt den Text
-
-# ── Streaming-/Endpointing-Parameter (analog whisper-Bridge) ──
+STREAM_TRANSCRIBE_INTERVAL_MS = int(os.getenv("STREAM_TRANSCRIBE_INTERVAL_MS", "1000"))
 STREAM_DEFAULT_ENDPOINT_MS = 2400
-STREAM_DEFAULT_HARD_CAP_MS = 60000
+STREAM_DEFAULT_HARD_CAP_MS = 300000
 STREAM_MIN_AUDIO_MS = 600
 STREAM_SESSION_TTL_S = 120
 STREAM_ENERGY_WINDOW_MS = 300
 STREAM_SEMANTIC_BACKUP_FACTOR = 2.0
-# Adaptiver Voice-Schwellwert (siehe whisper-Bridge M0.1): Grenze relativ zum
-# gemessenen Rausch-Boden statt fix — schneidet leises Sprechen nicht ab.
+# Adaptiver Voice-Schwellwert (M0.1): relativ zum gemessenen Rausch-Boden.
 STREAM_VOICE_FACTOR = 2.5
 STREAM_VOICE_RMS_MIN = 0.005
 STREAM_VOICE_RMS_MAX = 0.020
@@ -91,21 +71,73 @@ STREAM_VOICE_RMS_MAX = 0.020
 def pcm_s16le_to_float32(data: bytes) -> np.ndarray:
     if not data:
         return np.zeros(0, dtype=np.float32)
-    arr = np.frombuffer(data, dtype=np.int16).astype(np.float32) / 32768.0
-    return arr
+    return np.frombuffer(data, dtype=np.int16).astype(np.float32) / 32768.0
 
 
 async def _send(ws, mtype: str, payload: dict) -> None:
     try:
         await ws.send(json.dumps({
-            "type": mtype,
-            "payload": payload,
-            "timestamp": int(time.time() * 1000),
+            "type": mtype, "payload": payload, "timestamp": int(time.time() * 1000),
         }))
     except Exception as e:
         logger.warning("RVS-Send fehlgeschlagen (%s): %s", mtype, e)
 
 
+class VoxtralRunner:
+    """Haelt das Voxtral-Modell (Transformers). transcribe() blockiert → aus dem
+    Event-Loop via run_in_executor aufrufen. Ein Lock serialisiert GPU-Zugriffe."""
+
+    def __init__(self) -> None:
+        self.model = None
+        self.processor = None
+        self._lock = asyncio.Lock()
+
+    def load(self) -> None:
+        import torch
+        from transformers import AutoProcessor, VoxtralForConditionalGeneration
+        t0 = time.time()
+        logger.info("Lade Voxtral '%s' (device=%s, bf16)…", VOXTRAL_MODEL, VOXTRAL_DEVICE)
+        self.processor = AutoProcessor.from_pretrained(VOXTRAL_MODEL)
+        self.model = VoxtralForConditionalGeneration.from_pretrained(
+            VOXTRAL_MODEL, torch_dtype=torch.bfloat16, device_map=VOXTRAL_DEVICE,
+        )
+        logger.info("Voxtral geladen in %.1fs", time.time() - t0)
+
+    def _transcribe_blocking(self, audio_f32: np.ndarray, language: str) -> str:
+        import torch
+        proc, model = self.processor, self.model
+        if proc is None or model is None or audio_f32.size == 0:
+            return ""
+        # VoxtralProcessor verlangt bei rohen Arrays ein 'format'. Robuster:
+        # in ein temp-WAV schreiben und den PFAD uebergeben — der Processor liest
+        # Format + Samplerate selbst, kein 'format'-Argument noetig.
+        wav_path = None
+        try:
+            with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tf:
+                wav_path = tf.name
+            sf.write(wav_path, audio_f32, 16000, subtype="PCM_16")
+            inputs = proc.apply_transcription_request(
+                language=language, audio=wav_path, model_id=VOXTRAL_MODEL,
+            )
+            inputs = inputs.to(VOXTRAL_DEVICE, dtype=torch.bfloat16)
+            with torch.no_grad():
+                outputs = model.generate(**inputs, max_new_tokens=512)
+            trimmed = outputs[:, inputs.input_ids.shape[1]:]
+            text = proc.batch_decode(trimmed, skip_special_tokens=True)
+            return (text[0] if text else "").strip()
+        finally:
+            if wav_path:
+                try:
+                    os.unlink(wav_path)
+                except Exception:
+                    pass
+
+    async def transcribe(self, audio_f32: np.ndarray, language: str) -> str:
+        loop = asyncio.get_running_loop()
+        async with self._lock:
+            return await loop.run_in_executor(None, self._transcribe_blocking, audio_f32, language)
+
+
 @dataclass
 class StreamSession:
     request_id: str
@@ -126,28 +158,26 @@ class StreamSession:
     last_chunk_at: float = field(default_factory=time.time)
     last_partial: str = ""
     last_growth_at: float = 0.0
+    last_transcribe_at: float = 0.0
     last_voice_at: float = 0.0
     noise_floor: float = 0.0
     closed: bool = False
     endpoint_sent: bool = False
-    # vLLM-Realtime-Session
-    vllm_ws: object = None
-    vllm_reader: object = None
 
 
 class SessionManager:
-    def __init__(self) -> None:
+    def __init__(self, runner: VoxtralRunner) -> None:
+        self.runner = runner
         self._sessions: dict[str, StreamSession] = {}
-        self._ws = None  # RVS
+        self._ws = None
 
     def attach_ws(self, ws) -> None:
         self._ws = ws
 
-    async def start_session(self, payload: dict) -> Optional[StreamSession]:
-        request_id = (payload.get("requestId") or "").strip()
-        if not request_id:
-            logger.warning("stt_stream_start ohne requestId — ignoriert")
-            return None
+    def start_session(self, payload: dict) -> None:
+        rid = (payload.get("requestId") or "").strip()
+        if not rid:
+            return
         try:
             endpoint_ms = int(payload.get("endpointMs") or STREAM_DEFAULT_ENDPOINT_MS)
         except (TypeError, ValueError):
@@ -160,8 +190,8 @@ class SessionManager:
             voice_factor = float(payload.get("voiceFactor") or STREAM_VOICE_FACTOR)
         except (TypeError, ValueError):
             voice_factor = STREAM_VOICE_FACTOR
-        sess = StreamSession(
-            request_id=request_id,
+        self._sessions[rid] = StreamSession(
+            request_id=rid,
             audio_request_id=payload.get("audioRequestId", "") or "",
             language=payload.get("language") or VOXTRAL_LANGUAGE,
             endpoint_ms=endpoint_ms,
@@ -173,159 +203,44 @@ class SessionManager:
             location=payload.get("location") or None,
             sample_rate=int(payload.get("sampleRate") or 16000),
         )
-        # vLLM-Realtime-Session oeffnen + Reader starten.
-        try:
-            sess.vllm_ws = await websockets.connect(VOXTRAL_VLLM_URL, max_size=8 * 1024 * 1024)
-            await self._vllm_configure(sess)
-            sess.vllm_reader = asyncio.create_task(self._vllm_read_loop(sess))
-        except Exception as e:
-            logger.exception("Stream %s: vLLM-Realtime-Connect fehlgeschlagen: %s",
-                             request_id[:8], e)
-            # ohne Backend keine Transkription → sofort leeres Endpoint melden
-            self._sessions[request_id] = sess
-            await self._finalize(sess, reason="vllm_unavailable")
-            return None
-        self._sessions[request_id] = sess
         logger.info("Voxtral-Session offen: id=%s lang=%s endpointMs=%d",
-                    request_id[:8], sess.language, sess.endpoint_ms)
-        return sess
-
-    async def _vllm_configure(self, sess: StreamSession) -> None:
-        """Optionale Session-Konfig an vLLM (Modell/Sprache/temperature=0).
-        VERIFY: exaktes session.update-Schema gegen vLLM-Realtime-Beispiel.
-        Best-effort — Fehler hier sind nicht fatal."""
-        try:
-            await sess.vllm_ws.send(json.dumps({
-                "type": "session.update",
-                "session": {
-                    "model": VOXTRAL_MODEL,
-                    "language": sess.language,
-                    "temperature": 0.0,
-                    "input_audio_format": "pcm16",
-                },
-            }))
-        except Exception:
-            pass
-
-    async def _vllm_read_loop(self, sess: StreamSession) -> None:
-        """Liest transcription.delta/.done vom vLLM-Realtime-Server."""
-        ws = sess.vllm_ws
-        try:
-            async for raw in ws:
-                try:
-                    msg = json.loads(raw)
-                except Exception:
-                    continue
-                mtype = str(msg.get("type", ""))
-                if any(mtype.endswith(s) for s in VLLM_DELTA_SUFFIXES):
-                    text = self._extract_text(msg)
-                    if text:
-                        await self._on_delta(sess, text)
-                elif any(mtype.endswith(s) for s in VLLM_DONE_SUFFIXES):
-                    text = self._extract_text(msg)
-                    if text:
-                        await self._on_delta(sess, text, final=True)
-        except Exception:
-            logger.debug("Stream %s: vLLM-Reader beendet", sess.request_id[:8])
-
-    @staticmethod
-    def _extract_text(msg: dict) -> str:
-        for f in VLLM_DELTA_FIELDS:
-            v = msg.get(f)
-            if isinstance(v, str) and v:
-                return v
-        return ""
-
-    async def _on_delta(self, sess: StreamSession, text: str, final: bool = False) -> None:
-        """Neuer/finaler Transkript-Text vom vLLM. delta = inkrementell; wir
-        haengen an, wenn er den bisherigen Partial verlaengert, sonst ersetzen
-        wir (Voxtral kann korrigieren)."""
-        if final or text.startswith(sess.last_partial):
-            new_full = text if final else text
-        else:
-            new_full = (sess.last_partial + text).strip()
-        new_full = new_full.strip()
-        if new_full and new_full != sess.last_partial:
-            sess.last_partial = new_full
-            sess.last_growth_at = time.time()
-            if self._ws is not None:
-                await _send(self._ws, "stt_partial", {
-                    "requestId": sess.request_id,
-                    "audioRequestId": sess.audio_request_id,
-                    "text": new_full,
-                })
+                    rid[:8], self._sessions[rid].language, endpoint_ms)
 
     def feed_chunk(self, payload: dict) -> bool:
-        request_id = payload.get("requestId", "")
-        sess = self._sessions.get(request_id)
+        sess = self._sessions.get(payload.get("requestId", ""))
         if sess is None or sess.closed:
             return False
         pcm_b64 = payload.get("pcm", "")
-        if not pcm_b64:
-            return True
-        try:
-            pcm = base64.b64decode(pcm_b64)
-        except Exception:
-            return True
-        sess.pcm_buffer.extend(pcm)
+        if pcm_b64:
+            try:
+                sess.pcm_buffer.extend(base64.b64decode(pcm_b64))
+            except Exception:
+                pass
         sess.last_chunk_at = time.time()
-        # An vLLM weiterreichen (fire-and-forget).
-        if sess.vllm_ws is not None:
-            asyncio.create_task(self._vllm_append(sess, pcm_b64))
         return True
 
-    async def _vllm_append(self, sess: StreamSession, pcm_b64: str) -> None:
-        try:
-            await sess.vllm_ws.send(json.dumps({
-                "type": VLLM_SEND_APPEND,
-                VLLM_AUDIO_FIELD: pcm_b64,
-            }))
-        except Exception:
-            pass
-
     def end_session(self, request_id: str) -> None:
         sess = self._sessions.get(request_id)
         if sess is not None:
             sess.closed = True
 
     def drop(self, request_id: str) -> None:
-        sess = self._sessions.pop(request_id, None)
-        if sess is not None:
-            self._teardown_vllm(sess)
+        self._sessions.pop(request_id, None)
 
-    def _teardown_vllm(self, sess: StreamSession) -> None:
-        try:
-            if sess.vllm_reader is not None:
-                sess.vllm_reader.cancel()
-        except Exception:
-            pass
-        if sess.vllm_ws is not None:
-            asyncio.create_task(self._close_ws(sess.vllm_ws))
-            sess.vllm_ws = None
-
-    @staticmethod
-    async def _close_ws(ws) -> None:
-        try:
-            await ws.close()
-        except Exception:
-            pass
-
-    # ── Endpointer (adaptiv, wie whisper-Bridge M0.1) ──
-    def _buffer_duration_ms(self, sess: StreamSession) -> float:
+    # ── Endpointer (adaptiv, M0.1) ──
+    def _buffer_ms(self, sess: StreamSession) -> float:
         samples = len(sess.pcm_buffer) // 2
         return (samples / sess.sample_rate) * 1000.0 if samples else 0.0
 
     def _tail_rms(self, sess: StreamSession) -> float:
-        win_bytes = int(sess.sample_rate * STREAM_ENERGY_WINDOW_MS / 1000) * 2
-        if win_bytes <= 0:
+        win = int(sess.sample_rate * STREAM_ENERGY_WINDOW_MS / 1000) * 2
+        if win <= 0:
             return 0.0
-        tail = sess.pcm_buffer[-win_bytes:]
+        tail = sess.pcm_buffer[-win:]
         if len(tail) < 2:
             return 0.0
         arr = pcm_s16le_to_float32(bytes(tail))
-        if arr.size == 0:
-            return 0.0
-        return float(np.sqrt(np.mean(arr * arr)))
+        return float(np.sqrt(np.mean(arr * arr))) if arr.size else 0.0
 
     def _voice_threshold(self, sess: StreamSession) -> float:
         nf = sess.noise_floor
@@ -343,7 +258,8 @@ class SessionManager:
             sess.noise_floor = 0.98 * nf + 0.02 * rms
 
     async def run_endpointer(self) -> None:
-        logger.info("Voxtral-Endpointer gestartet (adaptiver VAD)")
+        logger.info("Voxtral-Endpointer gestartet (adaptiver VAD, interval=%dms)",
+                    STREAM_TRANSCRIBE_INTERVAL_MS)
         while True:
             await asyncio.sleep(0.2)
             now = time.time()
@@ -351,7 +267,7 @@ class SessionManager:
                 try:
                     await self._tick(sess, now)
                 except Exception:
-                    logger.exception("Endpointer-Tick crashed (session=%s)", sid[:8])
+                    logger.exception("Tick crashed (session=%s)", sid[:8])
             for sid, sess in list(self._sessions.items()):
                 if now - sess.last_chunk_at > STREAM_SESSION_TTL_S:
                     logger.info("Stream %s: TTL — drop", sid[:8])
@@ -360,14 +276,13 @@ class SessionManager:
     async def _tick(self, sess: StreamSession, now: float) -> None:
         if sess.endpoint_sent:
             return
-        elapsed_ms = (now - sess.started_at) * 1000.0
-        if elapsed_ms > sess.hard_cap_ms and not sess.closed:
-            await self._finalize(sess, reason="hardcap")
+        if (now - sess.started_at) * 1000.0 > sess.hard_cap_ms and not sess.closed:
+            await self._finalize(sess, "hardcap")
             return
         if sess.closed:
-            await self._finalize(sess, reason="stream_end")
+            await self._finalize(sess, "stream_end")
             return
-        if self._buffer_duration_ms(sess) < STREAM_MIN_AUDIO_MS:
+        if self._buffer_ms(sess) < STREAM_MIN_AUDIO_MS:
             return
         # adaptive akustische Sprach-Aktivitaet
         rms = self._tail_rms(sess)
@@ -375,45 +290,65 @@ class SessionManager:
             sess.last_voice_at = now
         else:
             self._update_noise_floor(sess, rms)
-        # Endpoint: akustisch (Primaer) oder semantisch (Backstop), sobald Text da
+        # Endpoint-Entscheidung, sobald Text erkannt wurde
         if sess.last_growth_at > 0.0:
-            acoustic_silence_ms = (now - sess.last_voice_at) * 1000.0 if sess.last_voice_at > 0 else 0.0
-            semantic_silence_ms = (now - sess.last_growth_at) * 1000.0
-            acoustic_done = sess.last_voice_at > 0 and acoustic_silence_ms >= sess.endpoint_ms
-            semantic_done = semantic_silence_ms >= sess.endpoint_ms * STREAM_SEMANTIC_BACKUP_FACTOR
-            if acoustic_done or semantic_done:
-                await self._finalize(sess, reason="endpoint" if acoustic_done else "endpoint_semantic")
+            ac_sil = (now - sess.last_voice_at) * 1000.0 if sess.last_voice_at > 0 else 0.0
+            se_sil = (now - sess.last_growth_at) * 1000.0
+            ac_done = sess.last_voice_at > 0 and ac_sil >= sess.endpoint_ms
+            se_done = se_sil >= sess.endpoint_ms * STREAM_SEMANTIC_BACKUP_FACTOR
+            if ac_done or se_done:
+                await self._finalize(sess, "endpoint" if ac_done else "endpoint_semantic")
+                return
+        # Partial-Transkription (throttled)
+        if (now - sess.last_transcribe_at) * 1000.0 < STREAM_TRANSCRIBE_INTERVAL_MS:
+            return
+        sess.last_transcribe_at = now
+        audio = pcm_s16le_to_float32(bytes(sess.pcm_buffer))
+        try:
+            text = (await self.runner.transcribe(audio, sess.language)).strip()
+        except Exception:
+            logger.exception("Stream %s: Partial-Transcribe crashed", sess.request_id[:8])
+            return
+        if text and text != sess.last_partial:
+            sess.last_partial = text
+            sess.last_growth_at = now
+            if self._ws is not None:
+                await _send(self._ws, "stt_partial", {
+                    "requestId": sess.request_id,
+                    "audioRequestId": sess.audio_request_id,
+                    "text": text,
+                })
 
     async def _finalize(self, sess: StreamSession, reason: str) -> None:
         if sess.endpoint_sent:
             return
         sess.endpoint_sent = True
-        # vLLM ggf. committen, damit ein letztes transcription.done kommt.
-        if sess.vllm_ws is not None:
-            try:
-                await sess.vllm_ws.send(json.dumps({"type": VLLM_SEND_COMMIT}))
-                await asyncio.sleep(0.15)  # kurz auf finalen Delta warten
-            except Exception:
-                pass
-        final_text = sess.last_partial.strip()
-        duration_s = self._buffer_duration_ms(sess) / 1000.0
-        logger.info("Stream %s: FINAL (reason=%s, %.1fs): %r",
-                    sess.request_id[:8], reason, duration_s, final_text[:120])
+        audio = pcm_s16le_to_float32(bytes(sess.pcm_buffer))
+        t0 = time.time()
+        try:
+            final_text = (await self.runner.transcribe(audio, sess.language)).strip()
+        except Exception:
+            logger.exception("Stream %s: Final-Transcribe crashed", sess.request_id[:8])
+            final_text = sess.last_partial
+        stt_ms = int((time.time() - t0) * 1000)
+        duration_s = audio.size / 16000.0
+        logger.info("Stream %s: FINAL (reason=%s, %.1fs, %dms): %r",
+                    sess.request_id[:8], reason, duration_s, stt_ms, final_text[:120])
         if self._ws is not None:
-            endpoint_payload = {
+            payload = {
                 "requestId": sess.request_id,
                 "audioRequestId": sess.audio_request_id,
                 "text": final_text,
                 "reason": reason,
                 "durationS": duration_s,
-                "sttMs": 0,
+                "sttMs": stt_ms,
                 "voice": sess.voice,
                 "speed": sess.speed,
                 "interrupted": sess.interrupted,
             }
             if sess.location:
-                endpoint_payload["location"] = sess.location
-            await _send(self._ws, "stt_endpoint", endpoint_payload)
+                payload["location"] = sess.location
+            await _send(self._ws, "stt_endpoint", payload)
             await _send(self._ws, "stt_stream_done", {
                 "requestId": sess.request_id,
                 "audioRequestId": sess.audio_request_id,
@@ -447,7 +382,6 @@ async def run_loop(sessions: SessionManager) -> None:
                 sessions.attach_ws(ws)
                 await _broadcast_status(ws, "ready", model=VOXTRAL_MODEL)
                 await _send(ws, "config_request", {"service": "voxtral"})
-
                 async for raw in ws:
                     try:
                         msg = json.loads(raw)
@@ -456,35 +390,32 @@ async def run_loop(sessions: SessionManager) -> None:
                     mtype = msg.get("type", "")
                     payload = msg.get("payload", {}) or {}
                     if mtype == "stt_stream_start":
-                        asyncio.create_task(sessions.start_session(payload))
+                        sessions.start_session(payload)
                     elif mtype == "stt_audio_chunk":
                         sessions.feed_chunk(payload)
                     elif mtype == "stt_stream_end":
                         sessions.end_session(payload.get("requestId", ""))
-                    # stt_request (Legacy One-Shot) macht Voxtral hier NICHT —
-                    # dafuer bleibt die whisper-Bridge (Fallback).
         except Exception as e:
             logger.warning("RVS-Verbindung verloren: %s — retry in %ds", e, retry_s)
             if use_tls and RVS_TLS_FALLBACK and not tls_fallback_tried:
                 use_tls = False
                 tls_fallback_tried = True
-                logger.info("TLS-Fallback: versuche ws:// (kein TLS)")
                 continue
             await asyncio.sleep(retry_s)
             retry_s = min(retry_s * 2, 30)
-            use_tls = RVS_TLS  # fuer den naechsten Zyklus zuruecksetzen
+            use_tls = RVS_TLS
 
 
 async def main() -> None:
     if not RVS_HOST or not RVS_TOKEN:
         logger.error("RVS_HOST/RVS_TOKEN fehlen — .env pruefen. Abbruch.")
         return
-    sessions = SessionManager()
-    logger.info("Voxtral-Bridge startet — vLLM=%s Modell=%s", VOXTRAL_VLLM_URL, VOXTRAL_MODEL)
-    await asyncio.gather(
-        run_loop(sessions),
-        sessions.run_endpointer(),
-    )
+    runner = VoxtralRunner()
+    loop = asyncio.get_running_loop()
+    await loop.run_in_executor(None, runner.load)   # Modell laden (blockierend)
+    sessions = SessionManager(runner)
+    logger.info("Voxtral-Bridge startet — Modell=%s", VOXTRAL_MODEL)
+    await asyncio.gather(run_loop(sessions), sessions.run_endpointer())
 
 
 if __name__ == "__main__":

xtts/voxtral/requirements.txt

@@ -1,5 +1,9 @@
-# Voxtral-Bridge ist reine CPU-Glue (RVS-WS <-> vLLM-Realtime-WS). Das Modell
-# selbst laeuft im separaten voxtral-vllm-Container (GPU). Deshalb hier KEIN
-# torch/vllm — nur der WebSocket-Client + numpy fuer die RMS-Energiemessung.
-websockets>=12.0
+# Voxtral-3B via Transformers. torch/torchaudio kommen cu124-gepinnt aus dem
+# Dockerfile (nicht hier, sonst zieht pip das Default-CUDA-Wheel).
+transformers>=4.54
+mistral-common[audio]>=1.8.1
+accelerate>=0.30
+soundfile>=0.12
+librosa>=0.10          # VoxtralProcessor.load_audio_as nutzt librosa zum WAV-Laden
 numpy>=1.24
+websockets>=12.0