feat: CosyVoice 한국어 엔진 통합 (Apache-2.0, GPU)

- vendor/CosyVoice: 패치된 CosyVoice 코드 + Matcha-TTS 벤더링
- cosyvoice_worker: 전용 venv에서 상주하는 합성 워커(FastAPI, 모델 1회 로드)
- app/engines/cosyvoice_engine: HTTP 프록시 엔진(속도/피치)
- Dockerfile: 전용 venv(torch cu128) + 모델 베이킹 + 빌드 워밍업
- entrypoint.sh: 워커(8001) + 메인앱(8788) 동시 기동

cosyvoice_worker/requirements.txt

@@ -0,0 +1,35 @@
+# CosyVoice 워커 전용 의존성 (호스트에서 검증된 조합).
+# torch/torchaudio 는 Dockerfile 에서 cu128 휠로 별도 설치.
+# openai-whisper/tiktoken/more-itertools 는 --no-deps 로 별도 설치(토치 덮어쓰기 방지).
+# pyworld/tensorrt/deepspeed/onnxruntime-gpu 는 추론에 불필요하여 제외.
+conformer==0.3.2
+diffusers==0.29.0
+fastapi==0.115.6
+fastapi-cli==0.0.4
+gdown==5.1.0
+gradio==5.4.0
+grpcio==1.57.0
+grpcio-tools==1.57.0
+hydra-core==1.3.2
+HyperPyYAML==1.2.3
+inflect==7.3.1
+librosa==0.10.2
+lightning==2.2.4
+matplotlib==3.7.5
+modelscope==1.20.0
+networkx==3.1
+numpy==1.26.4
+omegaconf==2.3.0
+onnx==1.16.0
+onnxruntime==1.18.0
+protobuf==4.25
+pyarrow==18.1.0
+pydantic==2.7.0
+rich==13.7.1
+soundfile==0.12.1
+tensorboard==2.14.0
+transformers==4.51.3
+x-transformers==2.11.24
+uvicorn==0.30.0
+wetext==0.0.4
+wget==3.2

cosyvoice_worker/worker.py

@@ -0,0 +1,105 @@
+"""CosyVoice 한국어 합성 워커 (전용 venv에서 실행).
+
+모델을 1회만 로드해 상주시키고, 메인 앱이 HTTP로 합성을 요청한다.
+- SFT 프리셋 화자('韩语女' = 한국어)로 참조음성 없이 합성
+- 속도(speed)는 CosyVoice 네이티브, 피치(pitch)는 librosa 후처리
+"""
+from __future__ import annotations
+
+import io
+import os
+import sys
+
+# 벤더링된 CosyVoice 코드 경로
+sys.path.insert(0, "/opt/CosyVoice")
+sys.path.insert(0, "/opt/CosyVoice/third_party/Matcha-TTS")
+
+import numpy as np
+import soundfile as sf
+import torch
+
+try:
+    import librosa
+except Exception:  # pragma: no cover
+    librosa = None
+
+from fastapi import FastAPI
+from fastapi.responses import Response
+from pydantic import BaseModel
+
+from cosyvoice.cli.cosyvoice import CosyVoice
+
+MODEL_DIR = os.environ.get(
+    "COSYVOICE_MODEL_DIR", "/models/cosyvoice/CosyVoice-300M-SFT"
+)
+DEFAULT_SPK = "韩语女"  # 한국어 여성 프리셋
+
+_model = None
+
+
+def get_model() -> CosyVoice:
+    global _model
+    if _model is None:
+        _model = CosyVoice(
+            MODEL_DIR, load_jit=False, load_trt=False, fp16=False
+        )
+    return _model
+
+
+app = FastAPI(title="CosyVoice Worker")
+
+
+class SynthReq(BaseModel):
+    text: str
+    speaker: str = DEFAULT_SPK
+    speed: float = 1.0
+    pitch: float = 0.0
+
+
+@app.on_event("startup")
+def _startup() -> None:
+    get_model()
+
+
+@app.get("/health")
+def health() -> dict:
+    m = get_model()
+    return {
+        "status": "ok",
+        "speakers": m.list_available_spks(),
+        "sr": m.sample_rate,
+        "cuda": torch.cuda.is_available(),
+    }
+
+
+@app.post("/synth")
+def synth(req: SynthReq) -> Response:
+    text = (req.text or "").strip()
+    if not text:
+        return Response(content="empty text", status_code=400)
+
+    m = get_model()
+    spks = m.list_available_spks()
+    spk = req.speaker if req.speaker in spks else (
+        DEFAULT_SPK if DEFAULT_SPK in spks else spks[0]
+    )
+    speed = float(max(0.5, min(2.0, req.speed)))
+    pitch = float(max(-12.0, min(12.0, req.pitch)))
+
+    outs = list(m.inference_sft(text, spk, stream=False, speed=speed))
+    audio = (
+        torch.cat([o["tts_speech"] for o in outs], dim=1)
+        .squeeze(0)
+        .cpu()
+        .numpy()
+        .astype(np.float32)
+    )
+    sr = m.sample_rate
+
+    if abs(pitch) > 1e-3 and librosa is not None:
+        audio = librosa.effects.pitch_shift(audio, sr=sr, n_steps=pitch)
+
+    buf = io.BytesIO()
+    sf.write(buf, audio, sr, format="WAV", subtype="PCM_16")
+    buf.seek(0)
+    return Response(content=buf.read(), media_type="audio/wav")