Add XTTS-v2 as first TTS backend, extensible for more

Mirrors the ASR backend pattern exactly: a BACKENDS dict in tts/router.py maps a backend name to its worker URL, so adding the next backend is just a new worker file + venv + supervisord entry + one dict line. XTTS-v2 runs in its own venv (--system-site-packages, inherits base image torch/CUDA) as a new supervisord program on port 8005. XTTS is zero-shot voice cloning, so a reference-voice library was added (/tts/voices CRUD, stored under /srv/tts/voices) — synthesis requires picking a previously uploaded voice. Results and model cache live under /srv/tts/{results,models-cache}, new quadlet volumes, owned by the same 983:983 user as the existing /srv/asr dirs. Fixed two environment issues uncovered while getting XTTS to actually run inside the container (non-root user, root-built venvs): - coqui-tts only pins transformers>=4.57 with no ceiling, so pip installed an incompatible 5.x; pinned to the last 4.x release. - HOME defaults to /app (owned by root) for the container's runtime user, so numba/matplotlib/torch cache writes failed; HOME is now forced to /tmp in all three workers (faster_whisper, qwen3, xtts) before any of those libraries get imported. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-06-18 18:46:45 +09:00
parent ac49bfc7f9
commit acdf098c41
13 changed files with 761 additions and 24 deletions
--- a/app/tts/config.py
+++ b/app/tts/config.py
@@ -6,7 +6,15 @@ from core.config import env_str

 TTS_BASE_DIR = Path(env_str("TTS_BASE_DIR", "/srv/tts"))
 TTS_MODEL_CACHE = Path(env_str("TTS_MODEL_CACHE_DIR", str(TTS_BASE_DIR / "models-cache")))
-TTS_UPLOAD_DIR = TTS_BASE_DIR / "uploads"
+TTS_VOICE_DIR = TTS_BASE_DIR / "voices"
 TTS_RESULT_DIR = TTS_BASE_DIR / "results"

-DEFAULT_TTS_BACKEND = env_str("DEFAULT_TTS_BACKEND", "")
+DEFAULT_TTS_BACKEND = env_str("DEFAULT_TTS_BACKEND", "xtts")
+DEFAULT_TTS_LANGUAGE = env_str("DEFAULT_TTS_LANGUAGE", "ko")
+
+XTTS_URL = env_str("XTTS_URL", "http://127.0.0.1:8005")
+
+
+def ensure_runtime_dirs() -> None:
+    for p in [TTS_BASE_DIR, TTS_MODEL_CACHE, TTS_VOICE_DIR, TTS_RESULT_DIR]:
+        p.mkdir(parents=True, exist_ok=True)
--- a/app/tts/router.py
+++ b/app/tts/router.py
@@ -1,29 +1,223 @@
 from __future__ import annotations

-from typing import Any, Dict, List
+import json
+import re
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Dict, List, Optional

-from fastapi import APIRouter, HTTPException
+import httpx
+from fastapi import APIRouter, File, Form, HTTPException, UploadFile
+from fastapi.responses import FileResponse

-from tts.config import DEFAULT_TTS_BACKEND
+from tts.config import (
+    DEFAULT_TTS_BACKEND,
+    DEFAULT_TTS_LANGUAGE,
+    TTS_RESULT_DIR,
+    TTS_VOICE_DIR,
+    XTTS_URL,
+    ensure_runtime_dirs,
+)

 router = APIRouter()

+BACKENDS = {
+    "xtts": XTTS_URL,
+}
+
+# XTTS-v2가 지원하는 언어 (Coqui 공식 문서 기준)
+XTTS_LANGUAGES = [
+    "ko", "en", "es", "fr", "de", "it", "pt", "pl", "tr", "ru",
+    "nl", "cs", "ar", "zh-cn", "ja", "hu", "hi",
+]
+
+_SAFE_NAME = re.compile(r"[^a-zA-Z0-9_\-가-힣]+")
+
+
+def _sanitize_voice_name(name: str) -> str:
+    name = name.strip()
+    if not name:
+        raise HTTPException(status_code=400, detail="보이스 이름이 비어 있습니다.")
+    return _SAFE_NAME.sub("_", name)
+
+
+def _find_voice_path(name: str) -> Optional[Path]:
+    safe = _sanitize_voice_name(name)
+    matches = list(TTS_VOICE_DIR.glob(f"{safe}.*"))
+    return matches[0] if matches else None
+
+
+# ─── Config ────────────────────────────────────────────────────────────────────

@router.get("/config")
 def config() -> Dict[str, Any]:
    return {
-        "status": "not_implemented",
        "default_backend": DEFAULT_TTS_BACKEND,
-        "backends": {},
-        "message": "TTS 백엔드가 아직 연동되지 않았습니다.",
+        "default_language": DEFAULT_TTS_LANGUAGE,
+        "backends": {
+            "xtts": {"languages": XTTS_LANGUAGES},
+        },
    }


+# ─── Voices (레퍼런스 보이스 관리) ───────────────────────────────────────────────
+
+@router.get("/voices")
+def list_voices() -> List[Dict[str, str]]:
+    ensure_runtime_dirs()
+    voices = []
+    for p in sorted(TTS_VOICE_DIR.iterdir()):
+        if p.is_file():
+            voices.append({"name": p.stem, "filename": p.name})
+    return voices
+
+
+@router.post("/voices")
+async def upload_voice(file: UploadFile = File(...), name: str = Form(...)) -> Dict[str, str]:
+    ensure_runtime_dirs()
+    safe_name = _sanitize_voice_name(name)
+    suffix = Path(file.filename or "voice.wav").suffix or ".wav"
+
+    # 같은 이름의 기존 파일(다른 확장자 포함) 제거 후 새로 저장
+    for old in TTS_VOICE_DIR.glob(f"{safe_name}.*"):
+        old.unlink(missing_ok=True)
+
+    dest = TTS_VOICE_DIR / f"{safe_name}{suffix}"
+    dest.write_bytes(await file.read())
+    return {"name": safe_name, "filename": dest.name}
+
+
+@router.delete("/voices/{name}")
+def delete_voice(name: str) -> Dict[str, str]:
+    path = _find_voice_path(name)
+    if not path:
+        raise HTTPException(status_code=404, detail="보이스를 찾을 수 없습니다.")
+    path.unlink(missing_ok=True)
+    return {"status": "deleted", "name": name}
+
+
+@router.get("/voices/{name}/audio")
+def serve_voice_audio(name: str) -> FileResponse:
+    path = _find_voice_path(name)
+    if not path:
+        raise HTTPException(status_code=404, detail="보이스를 찾을 수 없습니다.")
+    return FileResponse(path)
+
+
+# ─── History ──────────────────────────────────────────────────────────────────
+
@router.get("/history")
 def list_history() -> List[Dict[str, Any]]:
-    return []
+    ensure_runtime_dirs()
+    records = []
+    for p in sorted(TTS_RESULT_DIR.glob("*.json"), reverse=True):
+        try:
+            data = json.loads(p.read_text(encoding="utf-8"))
+            records.append({
+                "id": p.stem,
+                "backend": data.get("backend", ""),
+                "language": data.get("language", ""),
+                "voice": data.get("voice", ""),
+                "text_preview": (data.get("text", "") or "")[:120],
+                "duration": data.get("duration"),
+            })
+        except Exception:
+            continue
+    return records


+@router.get("/history/{record_id}")
+def get_history(record_id: str) -> Dict[str, Any]:
+    safe_id = Path(record_id).name
+    p = TTS_RESULT_DIR / f"{safe_id}.json"
+    if not p.exists():
+        raise HTTPException(status_code=404, detail="Record not found")
+    return json.loads(p.read_text(encoding="utf-8"))
+
+
+@router.delete("/history/{record_id}")
+def delete_history(record_id: str) -> Dict[str, str]:
+    safe_id = Path(record_id).name
+    p = TTS_RESULT_DIR / f"{safe_id}.json"
+    if not p.exists():
+        raise HTTPException(status_code=404, detail="Record not found")
+
+    try:
+        data = json.loads(p.read_text(encoding="utf-8"))
+        audio_path = data.get("audio_file", "")
+        if audio_path:
+            ap = Path(audio_path)
+            if ap.exists() and ap.is_relative_to(TTS_RESULT_DIR):
+                ap.unlink(missing_ok=True)
+    except Exception:
+        pass
+
+    p.unlink(missing_ok=True)
+    return {"status": "deleted", "id": safe_id}
+
+
+@router.get("/audio/{filename}")
+def serve_audio(filename: str) -> FileResponse:
+    safe_name = Path(filename).name
+    p = TTS_RESULT_DIR / safe_name
+    if not p.exists():
+        raise HTTPException(status_code=404, detail="File not found")
+    return FileResponse(p)
+
+
+# ─── Synthesize ───────────────────────────────────────────────────────────────
+
@router.post("/synthesize")
-def synthesize() -> None:
-    raise HTTPException(status_code=501, detail="TTS 합성 기능은 아직 구현되지 않았습니다.")
+async def synthesize(
+    text: str = Form(...),
+    backend: str = Form(DEFAULT_TTS_BACKEND),
+    language: str = Form(DEFAULT_TTS_LANGUAGE),
+    voice: str = Form(...),
+) -> Dict[str, Any]:
+    if backend not in BACKENDS:
+        raise HTTPException(status_code=400, detail=f"Unsupported backend: {backend}")
+    if not text.strip():
+        raise HTTPException(status_code=400, detail="text가 비어 있습니다.")
+
+    voice_path = _find_voice_path(voice)
+    if not voice_path:
+        raise HTTPException(status_code=404, detail=f"보이스 '{voice}'를 찾을 수 없습니다. 먼저 보이스를 업로드하세요.")
+
+    ensure_runtime_dirs()
+    worker_url = BACKENDS[backend]
+
+    if backend == "xtts":
+        data = {
+            "text": text,
+            "language": language,
+            "speaker_wav": str(voice_path),
+        }
+    else:
+        data = {"text": text, "language": language}
+
+    async with httpx.AsyncClient(timeout=httpx.Timeout(600.0, connect=60.0)) as client:
+        response = await client.post(f"{worker_url}/synthesize", data=data)
+
+    if response.status_code >= 400:
+        raise HTTPException(status_code=response.status_code, detail=f"Backend error: {response.text}")
+
+    timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f")
+    audio_path = TTS_RESULT_DIR / f"{timestamp}.wav"
+    audio_path.write_bytes(response.content)
+
+    duration = response.headers.get("x-audio-duration")
+    payload = {
+        "backend": backend,
+        "language": language,
+        "voice": voice,
+        "text": text,
+        "duration": float(duration) if duration else None,
+        "audio_file": str(audio_path),
+        "_timestamp": timestamp,
+    }
+    result_path = TTS_RESULT_DIR / f"{timestamp}.json"
+    result_path.write_text(json.dumps(payload, ensure_ascii=False, indent=2), encoding="utf-8")
+
+    payload["id"] = timestamp
+    payload["audio_url"] = f"/tts/audio/{audio_path.name}"
+    return payload
--- a/app/tts/workers/xtts_worker.py
+++ b/app/tts/workers/xtts_worker.py
@@ -0,0 +1,117 @@
+from __future__ import annotations
+
+import argparse
+import os
+import tempfile
+from pathlib import Path
+from typing import Any, Dict
+
+import sys
+sys.path.insert(0, "/app")
+from tts.config import TTS_MODEL_CACHE, ensure_runtime_dirs
+
+# TTS.api를 import하기 전에 설정해야 적용된다.
+os.environ.setdefault("TTS_HOME", str(TTS_MODEL_CACHE))
+os.environ.setdefault("COQUI_TOS_AGREED", "1")  # XTTS(CPML 라이선스) 동의 프롬프트를 비대화식으로 통과
+# 컨테이너는 983 유저로 실행되는데 HOME(/app)이 root 소유라 numba/matplotlib/torch가
+# 각자 ~/.cache, ~/.config 아래에 쓰려다 실패한다. HOME을 쓰기 가능한 곳으로 돌린다.
+os.environ["HOME"] = "/tmp"  # 컨테이너 기본 HOME=/app은 983 유저가 쓰기 불가 (setdefault로는 덮어쓰기 안 됨)
+os.environ.setdefault("NUMBA_CACHE_DIR", "/tmp/numba_cache")
+
+import soundfile as sf
+import uvicorn
+from fastapi import FastAPI, Form, HTTPException
+from fastapi.responses import Response
+
+app = FastAPI(title="TTS XTTS Worker")
+
+MODEL_NAME = "tts_models/multilingual/multi-dataset/xtts_v2"
+_MODEL_CACHE: Dict[str, Any] = {}
+
+
+def _log(msg: str) -> None:
+    print(f"[xtts] {msg}", flush=True)
+
+
+def _device() -> str:
+    try:
+        import torch
+        return "cuda" if torch.cuda.is_available() else "cpu"
+    except Exception:
+        return "cpu"
+
+
+def _load_model() -> Any:
+    if MODEL_NAME in _MODEL_CACHE:
+        return _MODEL_CACHE[MODEL_NAME]
+
+    from TTS.api import TTS
+
+    _log(f"loading model {MODEL_NAME}")
+    tts = TTS(MODEL_NAME).to(_device())
+    _MODEL_CACHE[MODEL_NAME] = tts
+    _log("model loaded")
+    return tts
+
+
+@app.on_event("startup")
+def startup() -> None:
+    ensure_runtime_dirs()
+
+
+@app.get("/health")
+def health() -> Dict[str, Any]:
+    return {"status": "ok", "device": _device(), "loaded": MODEL_NAME in _MODEL_CACHE}
+
+
+@app.post("/synthesize")
+def synthesize(
+    text: str = Form(...),
+    language: str = Form("ko"),
+    speaker_wav: str = Form(...),
+) -> Response:
+    speaker_path = Path(speaker_wav)
+    if not speaker_path.exists():
+        raise HTTPException(status_code=400, detail=f"speaker_wav not found: {speaker_wav}")
+
+    try:
+        tts = _load_model()
+    except Exception as e:
+        _log(f"model load error: {type(e).__name__}: {e}")
+        raise HTTPException(status_code=500, detail=f"모델 로드 실패: {type(e).__name__}: {e}")
+
+    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as tmp:
+        out_path = Path(tmp.name)
+
+    try:
+        _log(f"synthesizing language={language} chars={len(text)}")
+        tts.tts_to_file(
+            text=text,
+            speaker_wav=str(speaker_path),
+            language=language,
+            file_path=str(out_path),
+        )
+        audio_bytes = out_path.read_bytes()
+        duration = sf.info(str(out_path)).duration
+        return Response(
+            content=audio_bytes,
+            media_type="audio/wav",
+            headers={"x-audio-duration": str(round(duration, 3))},
+        )
+    except HTTPException:
+        raise
+    except Exception as e:
+        _log(f"synthesize error: {type(e).__name__}: {e}")
+        raise HTTPException(status_code=500, detail=f"{type(e).__name__}: {e}")
+    finally:
+        out_path.unlink(missing_ok=True)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--host", default="0.0.0.0")
+    parser.add_argument("--port", type=int, default=8005)
+    args = parser.parse_args()
+    ensure_runtime_dirs()
+    _log(f"starting host={args.host} port={args.port} device={_device()}")
+    uvicorn.run(app, host=args.host, port=args.port)