fix: 전 엔진 속도 조절을 피치보존 타임스트레치(Rubberband R3)로 통일

- 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)

outetts_worker/requirements.txt

@@ -5,3 +5,4 @@ soundfile
 librosa
 fastapi==0.115.6
 uvicorn==0.30.0
+pyrubberband==0.4.0

outetts_worker/worker.py

@@ -8,6 +8,8 @@ from __future__ import annotations
 
 import io
 
+import shutil
+
 import numpy as np
 import soundfile as sf
 
@@ -16,6 +18,35 @@ try:
 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
@@ -87,17 +118,11 @@ def synth(req: SynthReq) -> Response:
     audio = out.audio.cpu().numpy().reshape(-1).astype(np.float32)
     sr = int(out.sr)
 
-    if librosa is not None:
+    # 피치 보존 배속(고배속에서도 목소리 안 뭉개짐)
-        if abs(speed - 1.0) > 1e-3:
+    if abs(speed - 1.0) > 1e-3:
-            try:
+        audio = time_stretch(audio, sr, speed)
-                audio = librosa.effects.time_stretch(audio, rate=speed).astype(np.float32)
+    if abs(pitch) > 1e-3 and librosa is not None:
-            except Exception:
+        audio = librosa.effects.pitch_shift(audio, sr=sr, n_steps=pitch)
-                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")

supertonic_worker/requirements.txt

@@ -2,3 +2,4 @@
 supertonic==1.3.1
 fastapi==0.115.6
 uvicorn==0.30.0
+pyrubberband==0.4.0

supertonic_worker/worker.py

@@ -9,6 +9,8 @@ from __future__ import annotations
 
 import io
 
+import shutil
+
 import numpy as np
 import soundfile as sf
 
@@ -17,6 +19,35 @@ try:
 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
@@ -79,9 +110,12 @@ def synth(req: SynthReq) -> Response:
     speed = float(max(0.5, min(2.0, req.speed)))
     pitch = float(max(-12.0, min(12.0, req.pitch)))
 
-    wav, _dur = tts.synthesize(text, voice_style=style, lang=lang, speed=speed)
+    # 자연 속도로 합성 후 피치 보존 배속(고배속에서도 목소리 안 뭉개짐)
+    wav, _dur = tts.synthesize(text, voice_style=style, lang=lang, speed=1.0)
     audio = np.asarray(wav).reshape(-1).astype(np.float32)
 
+    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)