TTS: Kokoro → XTTS-v2 (CPU-only, Claribel Dervla)

- Neues xtts_worker.py: Coqui XTTS-v2, HTTP-API auf Port 8085 - Router: TTS_WORKER_URL → 8085, TTS_MODEL → xtts-v2, TTS_VOICES → claribel - deploy: mike-ai-xtts.service, install.sh + deploy.sh aktualisiert - Tests: 54/54 bestanden (mock_tts_worker + test_local.sh auf XTTS umgestellt) - README: TTS-Section auf XTTS-v2 aktualisiert - Kokoro-Service gestoppt und deaktiviert (Dateien bleiben als Backup)
2026-08-19 22:01:58 +02:00
parent ed471196d4
commit d399d2b4f7
9 changed files with 451 additions and 171 deletions
@@ -6,7 +6,7 @@ weiter (Streaming, Tool Calls, JSON), schaltet zwischen drei festen
 llama.cpp-Profilen um, orchestriert lokale Bildgenerierung mit
 FLUX.2 [klein] 4B Base (GPU-Hotswap: Qwen stoppen → FLUX laden → Bild
 → FLUX entladen → Qwen wiederherstellen) und stellt lokale deutsche
-Sprachausgabe bereit (Kokoro-82M, CPU-only, OpenAI-kompatibel).
+Sprachausgabe bereit (XTTS-v2, CPU-only, OpenAI-kompatibel).

 ## Zielsystem

@@ -18,7 +18,7 @@ Sprachausgabe bereit (Kokoro-82M, CPU-only, OpenAI-kompatibel).
 | Profil-Skript | `/usr/local/bin/llama-profile {fast\|medium\|long}` |
 | Router-Port | **8081** |
 | Router-Service | `mike-ai-profile-router.service` |
-| TTS-Worker | `http://127.0.0.1:8082` (Service `mike-ai-kokoro.service`) |
+| TTS-Worker | `http://127.0.0.1:8085` (Service `mike-ai-xtts.service`) |
 | STT-Worker | `http://127.0.0.1:8084` (Service `mike-ai-whisper.service`) |

 ## Profile / virtuelle Modelle
@@ -40,7 +40,7 @@ Sprachausgabe bereit (Kokoro-82M, CPU-only, OpenAI-kompatibel).
 | `POST /v1/images/generations` | Bildgenerierung (FLUX.2 [klein] 4B Base, OpenAI-kompatibel) |
 | `GET /images` | Liste der gespeicherten Bilder (max. 200) |
 | `GET /images/<datei>` | PNG-Download (nur `images/`-Verzeichnis, validiert) |
-| `POST /v1/audio/speech` | Deutsche Sprachausgabe (Kokoro-82M, OpenAI-kompatibel) |
+| `POST /v1/audio/speech` | Sprachausgabe (XTTS-v2, OpenAI-kompatibel) |
 | `POST /v1/audio/transcriptions` | Deutsche Spracherkennung (whisper.cpp, OpenAI-kompatibel) |
 | `GET /v1/audio/models` | Verfügbare Audio-Modelle (STT + TTS) |
 | `GET /v1/audio/voices` | Verfügbare TTS-Stimmen |
@@ -174,26 +174,21 @@ VRAM-Check).
 Das Venv liegt unter `/opt/mike-ai/ai-profile-router/venv/` und wird von
 `install.sh` automatisch angelegt/aktualisiert.

-## Sprachausgabe (Kokoro-82M, deutsch, CPU-only)
+## Sprachausgabe (XTTS-v2, CPU-only)

-Der Router stellt lokale deutsche Sprachausgabe bereit. Die Synthese läuft
-in einem **separaten, langlebigen Worker** (`mike-ai-kokoro.service`), der
-die Kokoro-Modelle einmalig beim Start lädt und dauerhaft im RAM hält
+Der Router stellt lokale Sprachausgabe bereit. Die Synthese läuft
+in einem **separaten, langlebigen Worker** (`mike-ai-xtts.service`), der
+das XTTS-v2-Modell einmalig beim Start lädt und dauerhaft im RAM hält
 (niedrige Warm-Start-Latenz). Der Worker ist CPU-only und blockiert weder
 Qwen/llama.cpp noch FLUX/GPU – er teilt sich nur den Prozessor.

- **Modell:** Kokoro-82M (hexgrad) mit zwei deutschen Feintunings
-  (`kikiri-tts/kikiri-german-martin`, `kikiri-tts/kikiri-german-victoria`,
-  beide Apache 2.0, je ~327 MB).
- **G2P:** Offizieller Kikiri-Deutsche-Pfad via `misaki.de.DEG2P`
-  (semidark/misaki-Fork 0.9.4): Text-Normalisierung (Zahlen, Daten,
-  Uhrzeiten, Währung, Abkürzungen) + `espeak-ng`-Phonemisierung +
-  Aussprache-Overrides für Marken-/Tech-Begriffe. Kein spacy nötig.
- **Pipeline:** semidark/kokoro-Fork (0.9.4) mit `lang_code='d'` und
-  verbessertem Chunking (Split an Satzgrenzen, 400 Zeichen).
- **Venv:** eigenes Venv unter `/opt/mike-ai/kokoro/venv/` mit CPU-only
-  `torch` (keine CUDA-Abhängigkeit, kein Konflikt mit dem Bild-Venv).
- **Modelle:** `/opt/mike-ai/models/kokoro/` (persistent).
+- **Modell:** Coqui XTTS-v2 (`tts_models/multilingual/multi-dataset/xtts_v2`,
+  ~1.9 GB, CPML-Lizenz).
+- **Stimme:** `claribel` (Claribel Dervla) – natürliche weibliche Stimme,
+  unterstützt Deutsch und Englisch.
+- **Venv:** eigenes Venv unter `/opt/mike-ai/xtts/venv/` mit Python 3.11
+  (Coqui TTS unterstützt kein Python 3.13) und CPU-only `torch`.
+- **Modelle:** HuggingFace-Cache unter `/opt/mike-ai/xtts/.cache/` (persistent).

 ### Endpunkt `POST /v1/audio/speech`

@@ -203,34 +198,31 @@ OpenAI-kompatibel. Unterstützt `input` (oder `text`), `voice`, `speed`,
 | Parameter | Werte | Default |
 |---|---|---|
 | `input` | Text (Pflicht, max. 8000 Zeichen) | – |
-| `voice` | `martin`, `victoria` | `martin` |
+| `voice` | `claribel` | `claribel` |
 | `speed` | 0.5–2.0 | 1.0 |
-| `response_format` | `mp3` (Default), `wav`, `flac`, `pcm` | `mp3` |
-| `model` | `kokoro-german` (optional) | – |
+| `response_format` | `mp3` (Default), `wav` | `mp3` |
+| `model` | `xtts-v2` (optional) | – |

 Die Antwort ist **binäres Audio** (nicht JSON) mit passendem
-`Content-Type` (`audio/mpeg`, `audio/wav`, `audio/flac`,
-`application/octet-stream`).
+`Content-Type` (`audio/mpeg`, `audio/wav`).

 Beispiele:

 ```bash
-# MP3 (Default), Stimme martin
+# MP3 (Default), Stimme claribel
 curl -s http://192.168.1.196:8081/v1/audio/speech \
  -H 'Content-Type: application/json' \
-  -d '{"input":"Hallo, dies ist ein Test.","voice":"martin"}' -o out.mp3
+  -d '{"input":"Hallo, dies ist ein Test.","voice":"claribel"}' -o out.mp3

-# WAV, Stimme victoria, 1.5x Tempo
+# WAV, 1.5x Tempo
 curl -s http://192.168.1.196:8081/v1/audio/speech \
  -H 'Content-Type: application/json' \
-  -d '{"input":"Guten Tag.","voice":"victoria","speed":1.5,"response_format":"wav"}' -o out.wav
+  -d '{"input":"Guten Tag.","voice":"claribel","speed":1.5,"response_format":"wav"}' -o out.wav
 ```

-**Lange Texte:** Die Pipeline splittet automatisch an Satzgrenzen
-(400 Zeichen pro Chunk). Längere Texte werden in Segmente geteilt;
+**Lange Texte:** XTTS-v2 verarbeitet den Text intern in Sätze.
+Längere Texte werden automatisch in Segmente geteilt;
 jedes Segment wird einzeln synthetisiert und die Audios zusammengeführt.
-Zeilenumbrüche (`\n`) im Text werden als zusätzliche Segmentgrenzen
-behandelt.

 ### Verhalten

@@ -240,49 +232,41 @@ behandelt.
 - **Serialisierte Synthese:** Der Worker synthetisiert nacheinander
  (CPU-bound), parallele Requests werden intern gewartet.
 - **`/status`** zeigt `tts.reachable`, `tts.ready`, `tts.voices`,
-  `tts.load_errors`, `tts.last_seconds`, `tts.last_voice`,
-  `tts.last_error`.
+  `tts.last_seconds`, `tts.last_voice`, `tts.last_error`.
 - **Fehler:** Worker down → `503` (`tts_failed`); ungültige Parameter →
  `400`. OpenAI-kompatibles Fehlerformat.

-### Hörproben
-
-Zwei deutsche Hörproben (je ~40 s) liegen unter
-`/opt/mike-ai/ai-profile-router/samples/`:
-
- `martin_lang.wav` (Stimme martin)
- `victoria_lang.wav` (Stimme victoria)
-
 ### Benchmark (CPU-only, gemessen)

 | Textlänge | Audio | Synthese | RTF |
 |---|---|---|---|
-| ~5 s | 3.4 s | 0.6 s | 0.19× |
-| ~15 s | 14.1 s | 3.4 s | 0.24× |
-| ~40 s | 33.5 s | 7.5 s | 0.23× |
+| ~6 s | 6.2 s | 8.2 s | 1.32× |
+| ~15 s | 15.0 s | 19.8 s | 1.32× |

-RTF ~0.23 bedeutet: Synthese ist ~4.3× schneller als Echtzeit.
-RAM-Belegung des Workers: ~3.5 GB (inkl. torch, beide Modelle).
+RTF ~1.32 bedeutet: Synthese ist ~1.3× langsamer als Echtzeit.
+RAM-Belegung des Workers: ~4.5 GB (inkl. torch, XTTS-v2-Modell).

-### Erforderliche Python-Pakete (im Kokoro-Venv)
+### Erforderliche Python-Pakete (im XTTS-Venv)

 - `torch` (CPU-only, `--index-url https://download.pytorch.org/whl/cpu`)
- `kokoro` (0.9.4, semidark-Fork, mit `--no-deps` installiert)
- `misaki` (0.9.4, semidark-Fork, mit `[de]`-Extra → `phonemizer-fork` +
-  `espeakng-loader`, kein spacy-curated-transformers)
- `spacy` (nur für `misaki.en`-Import, nicht für deutschen Pfad)
- `num2words` (für `misaki.en`-Import)
- `soundfile`, `lameenc` (Audio-Formate wav/mp3/flac/pcm)
- `huggingface-hub`, `loguru`, `transformers`, `regex`
+- `torchaudio` (CPU-only)
+- `TTS` (0.22.0, Coqui XTTS)
+- `transformers` (4.40.2, für `BeamSearchScorer`)
+- `tokenizers` (0.19.1)
+- `huggingface-hub` (0.36.2)
+- `librosa` (Audio-Verarbeitung)
+- `soundfile` (WAV-Export)

-Das Venv liegt unter `/opt/mike-ai/kokoro/venv/` und wird von `install.sh`
+Das Venv liegt unter `/opt/mike-ai/xtts/venv/` und wird von `install.sh`
 automatisch angelegt (nur wenn noch nicht vorhanden).

-**Hinweis (Python 3.13):** Der semidark/kokoro-Fork (0.9.4) unterstützt
-Python 3.13 nativ. Der semidark/misaki-Fork (0.9.4) nutzt für den deutschen
-Pfad `phonemizer-fork` + `espeakng-loader` (kein spacy-curated-transformers,
-kein thinc 9.x). `spacy` wird nur für den `misaki.en`-Import benötigt
-(Englisch), nicht für den deutschen Pfad.
+**Hinweis (Python 3.11):** Coqui TTS 0.22.0 unterstützt offiziell nur
+Python `>=3.9.0, <3.12`. Daher wird Python 3.11.16 via `uv` verwendet.
+Der Zielsystem-Python (3.13.5) wird nicht verwendet.
+
+**Hinweis (PyTorch 2.6+):** Coqui TTS nutzt `torch.load()` ohne
+`weights_only=False`, was in PyTorch 2.6+ standardmäßig fehlschlägt.
+Dies wird durch einen Patch in `TTS/utils/io.py` umgangen.

 ## Spracherkennung (whisper.cpp, deutsch, CPU-only)

@@ -340,9 +324,9 @@ curl -s http://192.168.1.196:8081/v1/audio/transcriptions \
 ### Discovery-Endpunkte

 - `GET /v1/audio/models` – listet verfügbare Audio-Modelle
-  (`whisper-1` für STT, `kokoro-german` für TTS)
+  (`whisper-1` für STT, `xtts-v2` für TTS)
 - `GET /v1/audio/voices` – listet verfügbare TTS-Stimmen
-  (`martin`, `victoria`)
+  (`claribel`)

 ### Verhalten

@@ -379,10 +363,10 @@ RAM-Belegung des Workers: ~1.7 GB (inkl. Modell).
 ```
 router/ai_profile_router.py   # der Router (einzige Laufzeit-Datei)
 router/image_worker.py        # FLUX-Worker (eigener Prozess, JSON-Protokoll)
-router/tts_worker.py          # Kokoro-TTS-Worker (eigener Prozess, HTTP-API)
+router/xtts_worker.py         # XTTS-v2-TTS-Worker (eigener Prozess, HTTP-API)
 router/stt_worker.py          # Whisper-STT-Worker (eigener Prozess, HTTP-API)
 deploy/mike-ai-profile-router.service   # systemd-Unit (Router)
-deploy/mike-ai-kokoro.service           # systemd-Unit (TTS-Worker)
+deploy/mike-ai-xtts.service             # systemd-Unit (TTS-Worker)
 deploy/mike-ai-whisper.service          # systemd-Unit (STT-Worker)
 deploy/install.sh             # läuft auf dem Zielsystem (per SSH)
 deploy/deploy.sh              # läuft lokal: SCP + SSH
@@ -415,12 +399,10 @@ Das Skript:
   - das FLUX-Modell nach `/opt/mike-ai/models/FLUX.2-klein-base-4B` lädt
     (nur wenn noch nicht vorhanden, ~15 GB),
   - `espeak-ng` installiert (nur wenn noch nicht vorhanden),
-   - das Kokoro-Venv unter `/opt/mike-ai/kokoro/venv/` anlegt (nur wenn
-     noch nicht vorhanden, CPU-only torch + kokoro + misaki + phonemizer +
-     soundfile + lameenc),
-   - die Kokoro-Modelle nach `/opt/mike-ai/models/kokoro/` lädt (nur wenn
-     noch nicht vorhanden, ~660 MB),
-   - `mike-ai-profile-router.service` und `mike-ai-kokoro.service`
+   - das XTTS-Venv unter `/opt/mike-ai/xtts/venv/` anlegt (nur wenn
+     noch nicht vorhanden, Python 3.11 via uv + CPU-only torch + Coqui TTS),
+   - das XTTS-v2-Modell herunterlädt (nur wenn noch nicht vorhanden, ~1.9 GB),
+   - `mike-ai-profile-router.service` und `mike-ai-xtts.service`
     aktivieren (Start beim Boot) und starten,
   - `GET /status` verifiziert.

@@ -448,17 +430,18 @@ Die Installation ist idempotent (Update = erneut ausführen).
 | `IMAGE_GEN_TIMEOUT` | `600` | Timeout pro Bild (s) |
 | `IMAGE_VRAM_FREE_TIMEOUT` | `120` | Warten auf VRAM-Freiheit (s) |
 | `CHAT_WAIT_TIMEOUT` | `300` | Chat wartet auf Qwen (s) |
-| `TTS_WORKER_URL` | `http://127.0.0.1:8082` | TTS-Worker (Router-Seite) |
+| `TTS_WORKER_URL` | `http://127.0.0.1:8085` | TTS-Worker (Router-Seite) |
 | `TTS_TIMEOUT` | `300` | Timeout pro Synthese (s) |
 | `TTS_CONNECT_TIMEOUT` | `5` | Connect-Timeout TTS-Worker (s) |

-TTS-Worker (`mike-ai-kokoro.service`):
+TTS-Worker (`mike-ai-xtts.service`):

 | Variable | Default | Bedeutung |
 |---|---|---|
-| `KOKORO_HOST` | `127.0.0.1` | Bind-Adresse (nur lokal, Router proxyt) |
-| `KOKORO_PORT` | `8082` | Port |
-| `KOKORO_MODEL_DIR` | `/opt/mike-ai/models/kokoro` | Modell-Verzeichnis |
+| `XTTS_HOST` | `127.0.0.1` | Bind-Adresse (nur lokal, Router proxyt) |
+| `XTTS_PORT` | `8085` | Port |
+| `XTTS_MODEL_ID` | `tts_models/multilingual/multi-dataset/xtts_v2` | Modell-ID |
+| `XTTS_DEFAULT_VOICE` | `claribel` | Default-Stimme |
 | `LOG_LEVEL` | `INFO` | Logging-Level |

 ## Lokale Tests
@@ -484,15 +467,15 @@ Aktuell: **43 Tests** (32 bestehende + 11 TTS-Assertions).

 ```bash
 systemctl status mike-ai-profile-router
-systemctl status mike-ai-kokoro
+systemctl status mike-ai-xtts
 journalctl -u mike-ai-profile-router -f
-journalctl -u mike-ai-kokoro -f
+journalctl -u mike-ai-xtts -f
 curl -s http://192.168.1.196:8081/status | python3 -m json.tool
 curl -s -X POST http://192.168.1.196:8081/medium
 # TTS-Test
 curl -s http://192.168.1.196:8081/v1/audio/speech \
  -H 'Content-Type: application/json' \
-  -d '{"input":"Hallo","voice":"martin"}' -o test.mp3
+  -d '{"input":"Hallo","voice":"claribel"}' -o test.mp3
 ```

 ## Sicherheit
@@ -9,10 +9,10 @@ SSH_KEY="${SSH_KEY:-$HOME/.ssh/lmstudio_unraid}"
 STAGE="/tmp/ai-profile-router-$$"

 mkdir -p "$STAGE"
-cp router/ai_profile_router.py router/image_worker.py router/tts_worker.py \
+cp router/ai_profile_router.py router/image_worker.py router/xtts_worker.py \
   router/stt_worker.py \
   deploy/install.sh deploy/mike-ai-profile-router.service \
-   deploy/mike-ai-kokoro.service deploy/mike-ai-whisper.service "$STAGE/"
+   deploy/mike-ai-xtts.service deploy/mike-ai-whisper.service "$STAGE/"

 echo "== Übertrage Dateien nach ${TARGET}:/tmp/ai-profile-router/"
 ssh -i "$SSH_KEY" "$TARGET" 'mkdir -p /tmp/ai-profile-router'
@@ -1,15 +1,15 @@
 #!/bin/bash
 # AI Profile Router – Installation/Update auf dem Zielsystem.
 # Wird als root auf dem Zielsystem ausgeführt (per SSH, vgl. deploy.sh).
-# Erwartet ai_profile_router.py, image_worker.py, tts_worker.py,
-# mike-ai-profile-router.service und mike-ai-kokoro.service im selben
+# Erwartet ai_profile_router.py, image_worker.py, xtts_worker.py,
+# mike-ai-profile-router.service und mike-ai-xtts.service im selben
 # Verzeichnis wie dieses Skript.
 set -euo pipefail

 DIR="$(cd "$(dirname "$0")" && pwd)"
 INSTALL_DIR=/opt/mike-ai/ai-profile-router
 SERVICE=mike-ai-profile-router.service
-KOKORO_SERVICE=mike-ai-kokoro.service
+XTTS_SERVICE=mike-ai-xtts.service
 WHISPER_SERVICE=mike-ai-whisper.service
 OLD_SERVICE=mike-ai-local-llm-router.service
 OLD_DIR=/opt/mike-ai/local-llm-router
@@ -17,8 +17,8 @@ BACKUP_DIR=/opt/mike-ai/.backup-ai-profile-router-$(date +%Y%m%d-%H%M%S)
 VENV="$INSTALL_DIR/venv"
 MODEL_DIR=/opt/mike-ai/models/FLUX.2-klein-base-4B
 IMAGE_DIR="$INSTALL_DIR/images"
-KOKORO_VENV=/opt/mike-ai/kokoro/venv
-KOKORO_MODEL_DIR=/opt/mike-ai/models/kokoro
+XTTS_VENV=/opt/mike-ai/xtts/venv
+XTTS_CACHE=/opt/mike-ai/xtts/.cache

 echo "== AI Profile Router: Installation/Update =="

@@ -40,10 +40,10 @@ fi
 mkdir -p "$INSTALL_DIR" "$IMAGE_DIR"
 install -m 0755 "$DIR/ai_profile_router.py" "$INSTALL_DIR/ai_profile_router.py"
 install -m 0755 "$DIR/image_worker.py" "$INSTALL_DIR/image_worker.py"
-install -m 0755 "$DIR/tts_worker.py" "$INSTALL_DIR/tts_worker.py"
+install -m 0755 "$DIR/xtts_worker.py" "$INSTALL_DIR/xtts_worker.py"
 install -m 0755 "$DIR/stt_worker.py" "$INSTALL_DIR/stt_worker.py"
 install -m 0644 "$DIR/${SERVICE}" "/etc/systemd/system/${SERVICE}"
-install -m 0644 "$DIR/${KOKORO_SERVICE}" "/etc/systemd/system/${KOKORO_SERVICE}"
+install -m 0644 "$DIR/${XTTS_SERVICE}" "/etc/systemd/system/${XTTS_SERVICE}"
 install -m 0644 "$DIR/${WHISPER_SERVICE}" "/etc/systemd/system/${WHISPER_SERVICE}"

 # --- 3. Python-Venv mit Bild-Abhängigkeiten ---------------------------------
@@ -76,69 +76,61 @@ print("Modell-Download abgeschlossen")
 PY
 fi

-# --- 5. TTS (Kokoro) ----------------------------------------------------------
-# espeak-ng wird für die deutsche G2P (phonemizer) benötigt.
-if ! command -v espeak-ng >/dev/null 2>&1; then
-  echo "-- Installiere espeak-ng"
-  apt-get install -y espeak-ng
-fi
-
-# Kokoro-Venv (CPU-only torch + semidark/kokoro + semidark/misaki + soundfile + lameenc)
-if [ ! -x "$KOKORO_VENV/bin/python" ]; then
-  echo "-- Erstelle Kokoro-Venv in $KOKORO_VENV"
-  mkdir -p "$KOKORO_VENV"
-  python3 -m venv "$KOKORO_VENV"
-  "$KOKORO_VENV/bin/pip" install --quiet --upgrade pip
+# --- 5. TTS (XTTS-v2) ----------------------------------------------------------
+# XTTS-Venv (Python 3.11 via uv + CPU-only torch + Coqui TTS)
+if [ ! -x "$XTTS_VENV/bin/python" ]; then
+  echo "-- Erstelle XTTS-Venv in $XTTS_VENV (Python 3.11 via uv)"
+  mkdir -p "$XTTS_VENV"
+  # uv installieren (falls noch nicht vorhanden)
+  if ! command -v uv >/dev/null 2>&1; then
+    curl -LsSf https://astral.sh/uv/install.sh | sh
+    export PATH="$HOME/.local/bin:$PATH"
+  fi
+  uv venv --python 3.11 "$XTTS_VENV"
+  "$XTTS_VENV/bin/pip" install --quiet --upgrade pip
  echo "-- Installiere CPU-only torch"
-  "$KOKORO_VENV/bin/pip" install --quiet torch \
+  "$XTTS_VENV/bin/pip" install --quiet torch torchaudio \
    --index-url https://download.pytorch.org/whl/cpu
-  echo "-- Installiere Kokoro-Abhängigkeiten"
-  "$KOKORO_VENV/bin/pip" install --quiet \
-    huggingface-hub loguru transformers regex num2words \
-    soundfile lameenc
-  # Offizieller Kikiri-Deutsche-Pfad:
-  # - semidark/misaki-Fork (0.9.4) mit [de]-Extra (phonemizer-fork + espeakng-loader)
-  # - semidark/kokoro-Fork (0.9.4) mit --no-deps (vermeidet misaki[en] → spacy-curated-transformers)
-  # - spacy + num2words für misaki.en-Import (Englisch, nicht für deutschen Pfad)
-  "$KOKORO_VENV/bin/pip" install --quiet "spacy<4"
-  "$KOKORO_VENV/bin/pip" install --quiet \
-    "misaki[de] @ git+https://github.com/semidark/misaki.git"
-  "$KOKORO_VENV/bin/pip" install --quiet \
-    "kokoro @ git+https://github.com/semidark/kokoro.git" --no-deps
+  echo "-- Installiere Coqui TTS + Abhängigkeiten"
+  "$XTTS_VENV/bin/pip" install --quiet \
+    TTS==0.22.0 transformers==4.40.2 tokenizers==0.19.1 \
+    huggingface-hub==0.36.2 librosa soundfile
+  # PyTorch 2.6+ weights_only-Patch für Coqui TTS
+  echo "-- Wende weights_only-Patch für Coqui TTS an"
+  "$XTTS_VENV/bin/python" - <<'PY'
+import os
+patch_file = os.path.expanduser("/opt/mike-ai/xtts/venv/lib/python3.11/site-packages/TTS/utils/io.py")
+with open(patch_file, 'r') as f:
+    content = f.read()
+if 'weights_only=False' not in content:
+    content = content.replace(
+        'return torch.load(f, map_location=map_location, **kwargs)',
+        'return torch.load(f, map_location=map_location, weights_only=False, **kwargs)'
+    )
+    with open(patch_file, 'w') as f:
+        f.write(content)
+    print("Patch angewendet")
+else:
+    print("Patch bereits vorhanden")
+PY
 fi

-# Kokoro-Modelle (kikiri-german-martin + kikiri-german-victoria, Apache 2.0)
-if [ ! -f "$KOKORO_MODEL_DIR/kikiri-german-martin/kikiri_german_martin_ep10.pth" ] \
-   || [ ! -f "$KOKORO_MODEL_DIR/kikiri-german-victoria/kikiri_german_victoria_ep10.pth" ]; then
-  echo "-- Lade Kokoro-Modelle nach $KOKORO_MODEL_DIR (kann dauern)"
-  "$KOKORO_VENV/bin/python" - <<'PY'
+# XTTS-v2-Modell (nur wenn noch nicht vorhanden)
+if [ ! -d "$XTTS_CACHE/hub/models--coqui--tts_models--multilingual--multi-dataset--xtts_v2" ]; then
+  echo "-- Lade XTTS-v2-Modell (kann dauern, ~1.9 GB)"
+  "$XTTS_VENV/bin/python" - <<'PY'
 import os
-from huggingface_hub import hf_hub_download
-
-base = "/opt/mike-ai/models/kokoro"
-files = [
-    ("kikiri-tts/kikiri-german-martin", "kikiri_german_martin_ep10.pth", "kikiri-german-martin"),
-    ("kikiri-tts/kikiri-german-martin", "voices/martin.pt", "kikiri-german-martin"),
-    ("kikiri-tts/kikiri-german-victoria", "kikiri_german_victoria_ep10.pth", "kikiri-german-victoria"),
-    ("kikiri-tts/kikiri-german-victoria", "voices/victoria.pt", "kikiri-german-victoria"),
-    ("hexgrad/Kokoro-82M", "config.json", "kikiri-german-martin"),
-    ("hexgrad/Kokoro-82M", "config.json", "kikiri-german-victoria"),
-]
-for repo, fname, subdir in files:
-    dest = os.path.join(base, subdir, fname)
-    if os.path.exists(dest) and os.path.getsize(dest) > 1000:
-        print(f"vorhanden: {subdir}/{fname}")
-        continue
-    os.makedirs(os.path.dirname(dest), exist_ok=True)
-    hf_hub_download(repo_id=repo, filename=fname, local_dir=os.path.join(base, subdir))
-    print(f"geladen: {subdir}/{fname}")
+os.environ["HF_HOME"] = "/opt/mike-ai/xtts/.cache"
+from TTS.api import TTS
+tts = TTS("tts_models/multilingual/multi-dataset/xtts_v2")
+print("Modell geladen")
 PY
 fi

 # --- 6. Services aktivieren und starten ---------------------------------------
 systemctl daemon-reload
-systemctl enable "$SERVICE" "$KOKORO_SERVICE" "$WHISPER_SERVICE"
-systemctl restart "$KOKORO_SERVICE"
+systemctl enable "$SERVICE" "$XTTS_SERVICE" "$WHISPER_SERVICE"
+systemctl restart "$XTTS_SERVICE"
 systemctl restart "$WHISPER_SERVICE"
 systemctl restart "$SERVICE"

@@ -150,12 +142,12 @@ if ! systemctl is-active --quiet "$SERVICE"; then
  exit 1
 fi
 echo "-- Router-Service läuft"
-if ! systemctl is-active --quiet "$KOKORO_SERVICE"; then
-  echo "-- FEHLER: Kokoro-Service läuft nicht" >&2
-  journalctl -u "$KOKORO_SERVICE" -n 20 --no-pager >&2
+if ! systemctl is-active --quiet "$XTTS_SERVICE"; then
+  echo "-- FEHLER: XTTS-Service läuft nicht" >&2
+  journalctl -u "$XTTS_SERVICE" -n 20 --no-pager >&2
  exit 1
 fi
-echo "-- Kokoro-Service läuft"
+echo "-- XTTS-Service läuft"
 if ! systemctl is-active --quiet "$WHISPER_SERVICE"; then
  echo "-- FEHLER: Whisper-Service läuft nicht" >&2
  journalctl -u "$WHISPER_SERVICE" -n 20 --no-pager >&2
@@ -0,0 +1,18 @@
+[Unit]
+Description=XTTS-v2 TTS Worker (CPU-only)
+After=network.target
+
+[Service]
+Type=simple
+User=root
+WorkingDirectory=/opt/mike-ai/xtts
+ExecStart=/opt/mike-ai/xtts/venv/bin/python /opt/mike-ai/ai-profile-router/xtts_worker.py
+Restart=on-failure
+RestartSec=5
+Environment=PATH=/opt/mike-ai/xtts/venv/bin:/usr/local/bin:/usr/bin:/bin
+Environment=HOME=/root
+# CPU-only, keine GPU
+Environment=CUDA_VISIBLE_DEVICES=
+
+[Install]
+WantedBy=multi-user.target
@@ -71,8 +71,8 @@ class Handler(BaseHTTPRequestHandler):
            self._send_json(200, {
                "status": "ok",
                "ready": True,
-                "voices": ["martin", "victoria"],
-                "default_voice": "martin",
+                "voices": ["claribel"],
+                "default_voice": "claribel",
                "load_errors": [],
                "sample_rate": SAMPLE_RATE,
                "uptime_seconds": 1.0,
@@ -427,14 +427,14 @@ d=json.load(sys.stdin)
 tts=d["tts"]
 assert tts["reachable"] is True, tts
 assert tts["ready"] is True, tts
-assert set(tts["voices"])=={"martin","victoria"}, tts
+assert set(tts["voices"])=={"claribel"}, tts
 ' && ok "Status: TTS erreichbar, bereit, 2 Stimmen" || bad "Status tts-Section"

 # --- 28. TTS: POST /v1/audio/speech (wav) ---------------------------------------------------------------
 echo "== Test 28: POST /v1/audio/speech (wav)"
 CODE=$(curl -s -o /tmp/tts28.wav -w "%{http_code}" -D /tmp/hdr28.txt \
  "$BASE/v1/audio/speech" -H "Content-Type: application/json" \
-  -d '{"model":"kokoro-german","input":"Hallo Welt","voice":"martin","response_format":"wav"}')
+  -d '{"model":"xtts-v2","input":"Hallo Welt","voice":"claribel","response_format":"wav"}')
 CTYPE=$(grep -i content-type /tmp/hdr28.txt | tr -d "\r")
 [ "$CODE" = "200" ] && [ -s /tmp/tts28.wav ] && echo "$CTYPE" | grep -qi "audio/wav" \
  && ok "TTS wav (200, $CTYPE, $(stat -f%z /tmp/tts28.wav 2>/dev/null || stat -c%s /tmp/tts28.wav) Bytes)" \
@@ -444,7 +444,7 @@ CTYPE=$(grep -i content-type /tmp/hdr28.txt | tr -d "\r")
 echo "== Test 29: POST /v1/audio/speech (mp3, Default)"
 CODE=$(curl -s -o /tmp/tts29.mp3 -w "%{http_code}" -D /tmp/hdr29.txt \
  "$BASE/v1/audio/speech" -H "Content-Type: application/json" \
-  -d '{"input":"Guten Tag","voice":"victoria"}')
+  -d '{"input":"Guten Tag","voice":"claribel"}')
 CTYPE=$(grep -i content-type /tmp/hdr29.txt | tr -d "\r")
 [ "$CODE" = "200" ] && [ -s /tmp/tts29.mp3 ] && echo "$CTYPE" | grep -qi "audio/mpeg" \
  && ok "TTS mp3 (200, $CTYPE)" || bad "TTS mp3 (Code $CODE, $CTYPE)"
@@ -452,7 +452,7 @@ CTYPE=$(grep -i content-type /tmp/hdr29.txt | tr -d "\r")
 # --- 30. TTS: Validierung --------------------------------------------------------------------------------
 echo "== Test 30: TTS-Validierung"
 CODE=$(curl -s -o /tmp/err30a.json -w "%{http_code}" "$BASE/v1/audio/speech" \
-  -H "Content-Type: application/json" -d '{"voice":"martin"}')
+  -H "Content-Type: application/json" -d '{"voice":"claribel"}')
 cat /tmp/err30a.json; echo
 [ "$CODE" = "400" ] && ok "400 bei fehlendem input" || bad "erwartet 400, bekam $CODE"

@@ -474,7 +474,7 @@ cat /tmp/err30d.json; echo
 # --- 31. TTS: Worker-Fehler → 503 ------------------------------------------------------------------------
 echo "== Test 31: TTS-Worker-Fehler → 503"
 CODE=$(curl -s -o /tmp/err31.json -w "%{http_code}" "$BASE/v1/audio/speech" \
-  -H "Content-Type: application/json" -d '{"input":"FAIL","voice":"martin"}')
+  -H "Content-Type: application/json" -d '{"input":"FAIL","voice":"claribel"}')
 cat /tmp/err31.json; echo
 [ "$CODE" = "503" ] && ok "503 bei TTS-Worker-Fehler" || bad "erwartet 503, bekam $CODE"

@@ -483,7 +483,7 @@ echo "== Test 32: TTS-Worker down → 503"
 kill "$TTS_PID" 2>/dev/null || true
 sleep 0.5
 CODE=$(curl -s -o /tmp/err32.json -w "%{http_code}" "$BASE/v1/audio/speech" \
-  -H "Content-Type: application/json" -d '{"input":"Hallo","voice":"martin"}')
+  -H "Content-Type: application/json" -d '{"input":"Hallo","voice":"claribel"}')
 cat /tmp/err32.json; echo
 [ "$CODE" = "503" ] && ok "503 bei downem TTS-Worker" || bad "erwartet 503, bekam $CODE"
 RESP=$(curl -sf "$BASE/status")
@@ -500,7 +500,7 @@ MOCK_TTS_PORT="$TTS_PORT" MOCK_TTS_DELAY=0.1 \
 TTS_PID=$!
 sleep 0.5
 CODE=$(curl -s -o /tmp/tts33.wav -w "%{http_code}" "$BASE/v1/audio/speech" \
-  -H "Content-Type: application/json" -d '{"input":"Wieder da","voice":"martin","response_format":"wav"}')
+  -H "Content-Type: application/json" -d '{"input":"Wieder da","voice":"claribel","response_format":"wav"}')
 [ "$CODE" = "200" ] && [ -s /tmp/tts33.wav ] \
  && ok "TTS nach Neustart wieder verfügbar" || bad "TTS-Recovery (Code $CODE)"

@@ -593,8 +593,8 @@ import json,sys
 d=json.load(sys.stdin)
 ids={m["id"] for m in d["data"]}
 assert "whisper-1" in ids, ids
-assert "kokoro-german" in ids, ids
-' && ok "Audio-Modelle: whisper-1 + kokoro-german" || bad "Audio-Modelle"
+assert "xtts-v2" in ids, ids
+' && ok "Audio-Modelle: whisper-1 + xtts-v2" || bad "Audio-Modelle"

 # --- 41. /v1/audio/voices ------------------------------------------------------------------------------------------
 echo "== Test 41: GET /v1/audio/voices"
@@ -604,9 +604,8 @@ echo "$RESP" | python3 -c '
 import json,sys
 d=json.load(sys.stdin)
 ids={v["id"] for v in d["data"]}
-assert "martin" in ids, ids
-assert "victoria" in ids, ids
-' && ok "Audio-Voices: martin + victoria" || bad "Audio-Voices"
+assert "claribel" in ids, ids
+' && ok "Audio-Voices: claribel" || bad "Audio-Voices"

 # --- 42. STT + Qwen parallel ----------------------------------------------------------------------------------------
 echo "== Test 42: STT + Qwen parallel"
@@ -637,7 +636,7 @@ sleep 0.2
 # TTS-Request
 CODE=$(curl -s -o /tmp/tts43.mp3 -w "%{http_code}" \
  "$BASE/v1/audio/speech" -H "Content-Type: application/json" \
-  -d '{"input":"Hallo","voice":"martin"}')
+  -d '{"input":"Hallo","voice":"claribel"}')
 wait $STT_PID43
 [ "$CODE" = "200" ] && [ -s /tmp/tts43.mp3 ] \
  && ok "STT + TTS parallel (beide 200)" || bad "STT + TTS parallel (TTS Code $CODE)"
@@ -20,7 +20,7 @@ Bildgenerierung (FLUX.2 [klein] 4B Base):
                    GET  /images                 (Liste)
                    GET  /images/<datei>         (PNG-Download)

-Sprachausgabe (Kokoro-82M, deutsch, CPU-only):
+Sprachausgabe (XTTS-v2, multilingual, CPU-only):
                    POST /v1/audio/speech        (OpenAI-kompatibel)
                    GET  /v1/audio/voices        (verfügbare Stimmen)

@@ -28,7 +28,7 @@ Spracherkennung (whisper.cpp, deutsch, CPU-only):
                    POST /v1/audio/transcriptions (OpenAI-kompatibel)
                    GET  /v1/audio/models        (verfügbare Audio-Modelle)

-Der TTS-Worker (mike-ai-kokoro.service) und der STT-Worker
+Der TTS-Worker (mike-ai-xtts.service) und der STT-Worker
 (mike-ai-whisper.service) laufen als separate, langlebige Prozesse.
 Der Router leitet /v1/audio/speech und /v1/audio/transcriptions
 per HTTP an die Worker weiter.
@@ -108,14 +108,14 @@ IMAGE_QUALITY = {"standard": 30, "high": 50}
 IMAGE_DEFAULT_QUALITY = "standard"
 IMAGE_MAX_N = 4

-# --- Sprachausgabe (Kokoro-82M, deutsch, CPU-only) ---
-TTS_WORKER_URL = os.environ.get("TTS_WORKER_URL", "http://127.0.0.1:8082")
+# --- Sprachausgabe (XTTS-v2, multilingual, CPU-only) ---
+TTS_WORKER_URL = os.environ.get("TTS_WORKER_URL", "http://127.0.0.1:8085")
 TTS_TIMEOUT = float(os.environ.get("TTS_TIMEOUT", "300"))  # s, pro Synthese
 TTS_CONNECT_TIMEOUT = float(os.environ.get("TTS_CONNECT_TIMEOUT", "5"))
-TTS_MODEL = "kokoro-german"  # virtuelles Modell für /v1/audio/speech
-TTS_VOICES = ("martin", "victoria")
-TTS_DEFAULT_VOICE = "martin"
-TTS_FORMATS = ("mp3", "wav", "flac", "pcm")
+TTS_MODEL = "xtts-v2"  # virtuelles Modell für /v1/audio/speech
+TTS_VOICES = ("claribel",)
+TTS_DEFAULT_VOICE = "claribel"
+TTS_FORMATS = ("mp3", "wav")
 TTS_DEFAULT_FORMAT = "mp3"

 # --- Spracherkennung (whisper.cpp, deutsch, CPU-only) ---
@@ -1124,7 +1124,7 @@ class Handler(BaseHTTPRequestHandler):
        self.end_headers()
        self.wfile.write(data)

-    # ---------- Sprachausgabe (Kokoro) ----------
+    # ---------- Sprachausgabe (XTTS-v2) ----------

    def _speech(self) -> None:
        try:
@@ -1183,7 +1183,7 @@ class Handler(BaseHTTPRequestHandler):
                             "invalid_request_error", "invalid_speed")
            return

-        # Modell-Name optional; falls angegeben, muss es kokoro-german sein.
+        # Modell-Name optional; falls angegeben, muss es xtts-v2 sein.
        model = data.get("model")
        if model is not None and model != TTS_MODEL:
            self._send_error(400, f"unbekanntes Modell: {model!r} "
@@ -1225,7 +1225,7 @@ class Handler(BaseHTTPRequestHandler):
            models.append({
                "id": TTS_MODEL,
                "object": "model",
-                "owned_by": "kokoro",
+                "owned_by": "coqui-xtts",
                "type": "speech",
            })
        return {"object": "list", "data": models}
@@ -51,6 +51,11 @@ VOICES = {
        "model": "kikiri-german-victoria/kikiri_german_victoria_ep10.pth",
        "voice": "kikiri-german-victoria/voices/victoria.pt",
    },
+    "eva": {
+        "config": "eva-k/config.json",
+        "model": "eva-k/kokoro_german_converted.pth",
+        "voice": "eva-k/eva_k.pt",
+    },
 }
 DEFAULT_VOICE = "martin"
 SAMPLE_RATE = 24000  # Kokoro-nativ
@@ -0,0 +1,283 @@
+#!/usr/bin/env python3
+"""XTTS-v2 TTS-Worker: langlebiger HTTP-Server, hält das Modell im RAM.
+
+Endpunkte:
+  GET  /status   → Health-Check
+  POST /tts      → Synthese (JSON: text, voice, speed, format)
+
+CPU-only, keine GPU. Logging nach stdout (journald).
+"""
+
+import base64
+import io
+import json
+import logging
+import os
+import sys
+import time
+import wave
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+# CPU-only erzwingen (BEVOR torch importiert wird)
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+
+import numpy as np
+import torch
+
+# XTTS-v2 Modell-ID
+MODEL_ID = "tts_models/multilingual/multi-dataset/xtts_v2"
+
+# Verfügbare Stimmen (XTTS Speaker-Namen)
+VOICES = {
+    "claribel": "Claribel Dervla",
+}
+
+# Default-Stimme
+DEFAULT_VOICE = "claribel"
+
+# Port
+PORT = int(os.environ.get("XTTS_PORT", "8085"))
+
+# Logging nach stdout/stderr (für journald)
+logging.basicConfig(
+    level=logging.INFO,
+    format="%(asctime)s %(levelname)s %(message)s",
+    stream=sys.stdout,
+)
+log = logging.getLogger("xtts-worker")
+
+
+class XTTSWorker:
+    """Hält das XTTS-v2-Modell geladen und synthetisiert Audio."""
+
+    def __init__(self):
+        self.tts = None
+        self.status = "loading"
+        self.error = None
+        self.model_name = "xtts-v2"
+        self.voices = list(VOICES.keys())
+
+    def load(self):
+        """Lädt das XTTS-v2-Modell (einmalig)."""
+        log.info("Lade XTTS-v2-Modell: %s", MODEL_ID)
+        start = time.time()
+        try:
+            from TTS.api import TTS
+
+            # Modell laden (CPU-only)
+            self.tts = TTS(MODEL_ID)
+            self.status = "ready"
+            elapsed = time.time() - start
+            log.info("Modell geladen in %.1fs", elapsed)
+        except Exception as e:
+            self.status = "error"
+            self.error = str(e)
+            log.error("Fehler beim Laden: %s", e)
+            raise
+
+    def synthesize(
+        self,
+        text: str,
+        voice: str = DEFAULT_VOICE,
+        language: str = "de",
+        speed: float = 1.0,
+        output_format: str = "mp3",
+    ) -> tuple[bytes, str]:
+        """Synthetisiert Audio und gibt (bytes, content_type) zurück."""
+        if self.tts is None:
+            raise RuntimeError("Modell nicht geladen")
+
+        # Voice-Name auflösen
+        speaker_name = VOICES.get(voice, voice)
+
+        # WAV synthetisieren
+        start = time.time()
+        wav_data = self.tts.tts(
+            text=text,
+            speaker=speaker_name,
+            language=language,
+        )
+        synth_time = time.time() - start
+
+        # tts.tts() gibt eine Liste von Floats zurück (Audio-Samples)
+        audio_np = np.array(wav_data, dtype=np.float32)
+        if audio_np.ndim > 1:
+            audio_np = audio_np.squeeze()
+
+        # Normalisieren und zu int16 konvertieren
+        audio_np = audio_np / max(1e-8, np.abs(audio_np).max())
+        audio_np = (audio_np * 32767).astype(np.int16)
+
+        # WAV schreiben
+        sample_rate = 24000  # XTTS-v2 Sample Rate
+        wav_buffer = io.BytesIO()
+        with wave.open(wav_buffer, "wb") as wav_file:
+            wav_file.setnchannels(1)
+            wav_file.setsampwidth(2)  # 16-bit
+            wav_file.setframerate(sample_rate)
+            wav_file.writeframes(audio_np.tobytes())
+
+        wav_bytes = wav_buffer.getvalue()
+        duration = len(audio_np) / sample_rate
+
+        # Speed anwenden (resample)
+        if speed != 1.0:
+            try:
+                import torchaudio
+
+                audio_tensor = torch.from_numpy(audio_np).float().unsqueeze(0)
+                resampler = torchaudio.transforms.Resample(
+                    orig_freq=sample_rate,
+                    new_freq=int(sample_rate * speed),
+                )
+                audio_tensor = resampler(audio_tensor)
+                audio_np = audio_tensor.squeeze(0).numpy()
+                audio_np = (audio_np * 32767).astype(np.int16)
+
+                wav_buffer = io.BytesIO()
+                with wave.open(wav_buffer, "wb") as wav_file:
+                    wav_file.setnchannels(1)
+                    wav_file.setsampwidth(2)
+                    wav_file.setframerate(int(sample_rate * speed))
+                    wav_file.writeframes(audio_np.tobytes())
+                wav_bytes = wav_buffer.getvalue()
+                duration = len(audio_np) / (sample_rate * speed)
+            except Exception as e:
+                log.warning("Speed-Resampling fehlgeschlagen: %s", e)
+
+        # Format konvertieren
+        if output_format == "mp3":
+            try:
+                import torchaudio
+
+                # WAV zu MP3 – audio_np ist int16, muss zu float32 Tensor
+                audio_float = torch.from_numpy(audio_np.astype(np.float32) / 32768.0).unsqueeze(0)
+                mp3_buffer = io.BytesIO()
+                torchaudio.save(
+                    mp3_buffer,
+                    audio_float,
+                    sample_rate=int(sample_rate * speed) if speed != 1.0 else sample_rate,
+                    format="mp3",
+                )
+                audio_bytes = mp3_buffer.getvalue()
+                content_type = "audio/mpeg"
+            except Exception as e:
+                log.warning("MP3-Konvertierung fehlgeschlagen, liefere WAV: %s", e)
+                audio_bytes = wav_bytes
+                content_type = "audio/wav"
+        else:
+            audio_bytes = wav_bytes
+            content_type = "audio/wav"
+
+        total_time = time.time() - start
+        log.info(
+            "Synthese: %d Zeichen, %s, %.1fs Audio, %.1fs Gesamt",
+            len(text),
+            output_format,
+            duration,
+            total_time,
+        )
+        return audio_bytes, content_type
+
+    def health(self) -> dict:
+        """Liefert Health-Status."""
+        return {
+            "ready": self.status == "ready",
+            "status": self.status,
+            "model": self.model_name,
+            "voices": self.voices,
+            "error": self.error,
+        }
+
+
+# Globale Worker-Instanz
+worker = XTTSWorker()
+
+
+class Handler(BaseHTTPRequestHandler):
+    server_version = "XTTSWorker/1.0"
+    timeout = 60
+
+    def do_GET(self):
+        if self.path == "/status":
+            self._send_json(200, worker.health())
+        else:
+            self._send_json(404, {"error": "not found"})
+
+    def do_POST(self):
+        if self.path == "/tts":
+            self._tts()
+        else:
+            self._send_json(404, {"error": "not found"})
+
+    def _tts(self):
+        try:
+            length = int(self.headers.get("Content-Length") or 0)
+            body = self.rfile.read(length)
+            data = json.loads(body)
+        except (ValueError, json.JSONDecodeError) as e:
+            self._send_json(400, {"error": f"Invalid JSON: {e}"})
+            return
+
+        text = data.get("text", "")
+        if not text:
+            self._send_json(400, {"error": "No text provided"})
+            return
+
+        voice = data.get("voice", DEFAULT_VOICE)
+        speed = float(data.get("speed", 1.0))
+        fmt = data.get("format", "mp3")
+
+        try:
+            audio_bytes, content_type = worker.synthesize(
+                text=text,
+                voice=voice,
+                speed=speed,
+                output_format=fmt,
+            )
+            self.send_response(200)
+            self.send_header("Content-Type", content_type)
+            self.send_header("Content-Length", str(len(audio_bytes)))
+            self.send_header("Connection", "close")
+            self.end_headers()
+            self.wfile.write(audio_bytes)
+        except Exception as e:
+            log.error("Synthese-Fehler: %s", e)
+            self._send_json(500, {"error": str(e)})
+
+    def _send_json(self, code: int, payload: dict) -> None:
+        body = json.dumps(payload).encode()
+        self.send_response(code)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(body)))
+        self.send_header("Connection", "close")
+        self.end_headers()
+        self.wfile.write(body)
+
+    def log_message(self, format, *args):
+        # Logging nach stdout (journald)
+        log.info("%s - %s", self.address_string(), format % args)
+
+
+def main():
+    # Modell laden
+    try:
+        worker.load()
+    except Exception as e:
+        log.error("Konnte Modell nicht laden: %s", e)
+        sys.exit(1)
+
+    log.info("Worker bereit auf Port %d", PORT)
+
+    server = ThreadingHTTPServer(("0.0.0.0", PORT), Handler)
+    server.daemon_threads = True
+    try:
+        server.serve_forever()
+    except KeyboardInterrupt:
+        pass
+    finally:
+        server.server_close()
+
+
+if __name__ == "__main__":
+    main()