feat: OuteTTS 엔진 추가(개인용/비상업 CC-BY-NC, 한국어)

- Qwen2 500M + WavTokenizer, torch cu128 GPU, 전용 venv 워커(8003) - 한국어 프리셋 4종(male/female), 속도·피치 librosa 후처리 - transformers 4.49로 torchvision 강제 import 회피, soundfile 저장 - scenema-audio는 VRAM 24GB+/Gemma12B 게이트로 8GB GPU에선 불가 → 미포함
2026-08-25 10:02:24 +09:00
parent ed09c6a08f
commit 1aa0bbf3e6
6 changed files with 230 additions and 1 deletions
--- a/16
+++ b/16
@@ -72,11 +72,25 @@ RUN HF_HOME=/models /opt/supertonic-venv/bin/python -c "from supertonic import T

 ENV SUPERTONIC_URL=http://127.0.0.1:8002

+# ===== OuteTTS 엔진 (개인용/비상업, CC-BY-NC-4.0) =====
+# Qwen2 500M LM + WavTokenizer 코덱. torch cu128 GPU. transformers 4.49(토치비전 회피).
+COPY outetts_worker /opt/outetts_worker
+RUN python -m venv /opt/oute-venv \
+ && /opt/oute-venv/bin/pip install --no-cache-dir torch==2.11.0 torchaudio==2.11.0 \
+        --index-url https://download.pytorch.org/whl/cu128 \
+ && /opt/oute-venv/bin/pip install --no-cache-dir -r /opt/outetts_worker/requirements.txt
+RUN /opt/oute-venv/bin/pip uninstall -y torchvision 2>/dev/null || true
+
+# 모델+코덱 사전 다운로드(베이킹) — 빌드 시엔 GPU 없이 CPU 로 로드
+RUN HF_HOME=/models /opt/oute-venv/bin/python -c "import outetts; cfg=outetts.HFModelConfig_v1(model_path='OuteAI/OuteTTS-0.2-500M', language='ko'); outetts.InterfaceHF(model_version='0.2', cfg=cfg); print('oute warmup ok')"
+
+ENV OUTETTS_URL=http://127.0.0.1:8003
+
 COPY app ./app
 COPY frontend ./frontend
 COPY entrypoint.sh /entrypoint.sh
 RUN chmod +x /entrypoint.sh

-# MeloTTS(8788, 메인) + Supertonic 워커(8002) 동시 기동
+# MeloTTS(8788) + Supertonic(8002) + OuteTTS(8003) 동시 기동
 EXPOSE 8788
 CMD ["/entrypoint.sh"]
--- a/app/engines/init.py
+++ b/app/engines/init.py
@@ -41,6 +41,16 @@ def _build() -> None:
        except Exception as exc:  # pragma: no cover
            print(f"[registry] Supertonic 엔진 스킵: {exc}")

+        # OuteTTS (개인용/비상업, 전용 워커 URL 이 설정된 경우에만)
+        try:
+            from .outetts_engine import OuteTTSEngine
+
+            if OuteTTSEngine.available():
+                e = OuteTTSEngine()
+                _ENGINES[e.id] = e
+        except Exception as exc:  # pragma: no cover
+            print(f"[registry] OuteTTS 엔진 스킵: {exc}")
+
        _INITED = True


--- a/app/engines/outetts_engine.py
+++ b/app/engines/outetts_engine.py
@@ -0,0 +1,84 @@
+"""OuteTTS 엔진 (프록시, 개인용/비상업).
+
+CC-BY-NC-4.0 라이선스라 개인/비상업 용도로만 사용. 워커 URL(OUTETTS_URL)이
+설정된 경우에만 활성화된다.
+"""
+from __future__ import annotations
+
+import json
+import os
+import urllib.error
+import urllib.request
+
+from .base import BaseEngine
+
+WORKER_URL = os.environ.get("OUTETTS_URL", "").rstrip("/")
+
+_VOICES = [
+    {"id": "female_1", "label": "여성 1"},
+    {"id": "female_2", "label": "여성 2"},
+    {"id": "male_1", "label": "남성 1"},
+    {"id": "male_2", "label": "남성 2"},
+]
+
+
+class OuteTTSEngine(BaseEngine):
+    id = "outetts"
+    label = "OuteTTS (개인용·비상업)"
+    license = "CC-BY-NC-4.0"
+    uses_gpu = True
+    notes = "Qwen2 기반 한국어 TTS. 비상업 라이선스라 개인 사용 전용. 속도·피치 후처리."
+
+    @staticmethod
+    def available() -> bool:
+        return bool(WORKER_URL)
+
+    def describe(self) -> dict:
+        return {
+            "id": self.id,
+            "label": self.label,
+            "license": self.license,
+            "uses_gpu": self.uses_gpu,
+            "notes": self.notes,
+            "languages": [
+                {"code": "KO", "label": "한국어", "speakers": _VOICES},
+            ],
+            "default_language": "KO",
+            "supports": {"speed": True, "pitch": True},
+        }
+
+    def synth_wav(
+        self,
+        text: str,
+        language: str,
+        speaker: str | None = None,
+        speed: float = 1.0,
+        pitch: float = 0.0,
+    ) -> bytes:
+        text = (text or "").strip()
+        if not text:
+            raise ValueError("텍스트가 비어 있습니다.")
+        if not WORKER_URL:
+            raise ValueError("OuteTTS 워커가 설정되지 않았습니다.")
+
+        payload = json.dumps(
+            {
+                "text": text,
+                "voice": speaker or "female_1",
+                "speed": speed,
+                "pitch": pitch,
+            }
+        ).encode("utf-8")
+        request = urllib.request.Request(
+            WORKER_URL + "/synth",
+            data=payload,
+            headers={"Content-Type": "application/json"},
+        )
+        try:
+            with urllib.request.urlopen(request, timeout=180) as resp:
+                return resp.read()
+        except urllib.error.HTTPError as exc:
+            detail = exc.read().decode("utf-8", errors="ignore")
+            raise ValueError(f"OuteTTS 합성 실패: {detail}")
+        except Exception as exc:  # pragma: no cover
+            raise ValueError(f"OuteTTS 워커 연결 실패: {exc}")
--- a/entrypoint.sh
+++ b/entrypoint.sh
@@ -10,6 +10,15 @@ if [ -n "$SUPERTONIC_URL" ] && [ -x /opt/supertonic-venv/bin/python ]; then
      > /tmp/supertonic_worker.log 2>&1 &
 fi

+# OuteTTS 워커(전용 venv, 개인용) — 모델을 1회 로드해 상주
+if [ -n "$OUTETTS_URL" ] && [ -x /opt/oute-venv/bin/python ]; then
+  echo "[entrypoint] starting OuteTTS worker on :8003"
+  /opt/oute-venv/bin/python -m uvicorn worker:app \
+      --app-dir /opt/outetts_worker \
+      --host 127.0.0.1 --port 8003 \
+      > /tmp/outetts_worker.log 2>&1 &
+fi
+
 # 메인 앱(글로벌 파이썬, MeloTTS 인프로세스)
 echo "[entrypoint] starting main app on :8788"
 exec uvicorn app.server:app --host 0.0.0.0 --port 8788
--- a/outetts_worker/requirements.txt
+++ b/outetts_worker/requirements.txt
@@ -0,0 +1,7 @@
+# OuteTTS 워커 (torch cu128 은 Dockerfile 에서 별도 설치)
+outetts==0.3.3
+transformers==4.49.0
+soundfile
+librosa
+fastapi==0.115.6
+uvicorn==0.30.0
--- a/outetts_worker/worker.py
+++ b/outetts_worker/worker.py
@@ -0,0 +1,105 @@
+"""OuteTTS 0.2-500M 한국어 합성 워커 (전용 venv, 개인용/비상업 CC-BY-NC).
+
+- Qwen2 기반 500M LM + WavTokenizer 코덱, torch cu128 GPU 사용
+- 한국어 프리셋 화자 4종(male_1/2, female_1/2)
+- 속도/피치는 librosa 후처리(OuteTTS 자체 speed 파라미터 없음)
+"""
+from __future__ import annotations
+
+import io
+
+import numpy as np
+import soundfile as sf
+
+try:
+    import librosa
+except Exception:  # pragma: no cover
+    librosa = None
+
+from fastapi import FastAPI
+from fastapi.responses import Response
+from pydantic import BaseModel
+
+import outetts
+
+VOICES = ["female_1", "female_2", "male_1", "male_2"]
+
+_interface = None
+_speakers: dict = {}
+
+
+def get_interface():
+    global _interface
+    if _interface is None:
+        cfg = outetts.HFModelConfig_v1(
+            model_path="OuteAI/OuteTTS-0.2-500M", language="ko"
+        )
+        _interface = outetts.InterfaceHF(model_version="0.2", cfg=cfg)
+    return _interface
+
+
+def get_speaker(name: str):
+    if name not in _speakers:
+        _speakers[name] = get_interface().load_default_speaker(name=name)
+    return _speakers[name]
+
+
+app = FastAPI(title="OuteTTS Worker")
+
+
+class SynthReq(BaseModel):
+    text: str
+    voice: str = "female_1"
+    speed: float = 1.0
+    pitch: float = 0.0
+
+
+@app.on_event("startup")
+def _startup() -> None:
+    get_speaker("female_1")
+
+
+@app.get("/health")
+def health() -> dict:
+    get_interface()
+    return {"status": "ok", "voices": VOICES}
+
+
+@app.post("/synth")
+def synth(req: SynthReq) -> Response:
+    text = (req.text or "").strip()
+    if not text:
+        return Response(content="empty text", status_code=400)
+
+    it = get_interface()
+    voice = req.voice if req.voice in VOICES else "female_1"
+    speaker = get_speaker(voice)
+    speed = float(max(0.5, min(2.0, req.speed)))
+    pitch = float(max(-12.0, min(12.0, req.pitch)))
+
+    out = it.generate(
+        text=text,
+        temperature=0.1,
+        repetition_penalty=1.1,
+        max_length=4096,
+        speaker=speaker,
+    )
+    audio = out.audio.cpu().numpy().reshape(-1).astype(np.float32)
+    sr = int(out.sr)
+
+    if librosa is not None:
+        if abs(speed - 1.0) > 1e-3:
+            try:
+                audio = librosa.effects.time_stretch(audio, rate=speed).astype(np.float32)
+            except Exception:
+                try:
+                    audio = librosa.effects.time_stretch(audio, speed).astype(np.float32)
+                except Exception:
+                    pass
+        if abs(pitch) > 1e-3:
+            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")