febb3f3ee4
- Supertonic/OuteTTS도 자연 속도 합성 후 Rubberband로 배속 → 고배속에서 목소리 안 뭉개짐 - Typecast 등 AI보이스가 쓰는 방식(pitch-preserving TSM)과 동일 - pyrubberband 워커 의존성 추가(rubberband-cli는 이미 이미지에 포함)
131 lines
3.4 KiB
Python
131 lines
3.4 KiB
Python
"""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 shutil
|
|
|
|
import numpy as np
|
|
import soundfile as sf
|
|
|
|
try:
|
|
import librosa
|
|
except Exception: # pragma: no cover
|
|
librosa = None
|
|
|
|
try:
|
|
import pyrubberband as prb
|
|
except Exception: # pragma: no cover
|
|
prb = None
|
|
|
|
|
|
def time_stretch(audio: np.ndarray, sr: int, rate: float) -> np.ndarray:
|
|
"""피치 보존 배속(Rubberband R3 → librosa 폴백). rate>1 = 빠르게."""
|
|
if abs(rate - 1.0) < 1e-3:
|
|
return audio
|
|
audio = np.asarray(audio, dtype=np.float32)
|
|
if prb is not None and shutil.which("rubberband"):
|
|
try:
|
|
return np.asarray(
|
|
prb.time_stretch(audio, sr, rate, rbargs={"--engine": "finer"}),
|
|
dtype=np.float32,
|
|
)
|
|
except Exception:
|
|
try:
|
|
return np.asarray(prb.time_stretch(audio, sr, rate), dtype=np.float32)
|
|
except Exception:
|
|
pass
|
|
if librosa is not None:
|
|
try:
|
|
return librosa.effects.time_stretch(audio, rate=rate).astype(np.float32)
|
|
except Exception:
|
|
pass
|
|
return audio
|
|
|
|
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)
|
|
|
|
# 피치 보존 배속(고배속에서도 목소리 안 뭉개짐)
|
|
if abs(speed - 1.0) > 1e-3:
|
|
audio = time_stretch(audio, sr, speed)
|
|
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")
|