Add MMS TTS support and config

Introduce MMS-based Korean TTS, update dependencies and docs, and include environment settings.

.env

@@ -3,3 +3,5 @@ DB_PORT=5432
 DB_NAME=tts
 DB_USER=ncue
 DB_PASSWORD=ncue5004!
+TTS_ENGINE=mms
+MMS_MODEL=facebook/mms-tts-kor

.env.example

@@ -2,4 +2,6 @@ DB_HOST=ncue.net
 DB_PORT=5432
 DB_NAME=tts
 DB_USER=ncue
-DB_PASSWORD=ncue5004!
+DB_PASSWORD=your_db_password
+TTS_ENGINE=mms
+MMS_MODEL=facebook/mms-tts-kor

README.md

@@ -45,3 +45,5 @@ http://localhost:8000
 - 배포 스크립트 기본 포트는 `8019`이며 `PORT`로 변경할 수 있습니다.
 - ffmpeg가 설치되어 있어야 합니다.
 - mp3 파일은 `resources/` 아래에 저장됩니다.
+- 고품질 TTS를 위해 `TTS_ENGINE=mms`를 설정할 수 있습니다.
+- MMS 모델(`facebook/mms-tts-kor`)은 비상업(CC-BY-NC-4.0) 라이선스입니다.

requirements.txt

@@ -4,3 +4,6 @@ python-dotenv
 psycopg2-binary
 pyttsx3
 jinja2
+torch
+transformers
+soundfile

server/tts_service.py

@@ -2,9 +2,48 @@ import os
 import subprocess
 import tempfile
 from pathlib import Path
+from typing import Optional, Tuple
 
 import pyttsx3
 
+_MMS_CACHE: Optional[Tuple[object, object]] = None
+
+
+def _get_mms():
+    global _MMS_CACHE
+    if _MMS_CACHE is not None:
+        return _MMS_CACHE
+
+    try:
+        from transformers import VitsModel, AutoTokenizer
+        import torch
+    except Exception as exc:
+        raise RuntimeError("MMS TTS 사용을 위해 transformers/torch 설치가 필요합니다.") from exc
+
+    model_name = os.getenv("MMS_MODEL", "facebook/mms-tts-kor")
+    tokenizer = AutoTokenizer.from_pretrained(model_name)
+    model = VitsModel.from_pretrained(model_name)
+    model.eval()
+
+    _MMS_CACHE = (model, tokenizer)
+    return _MMS_CACHE
+
+
+def _text_to_wav_mms(text: str, wav_path: str) -> None:
+    try:
+        import torch
+        import soundfile as sf
+    except Exception as exc:
+        raise RuntimeError("MMS TTS 사용을 위해 soundfile 설치가 필요합니다.") from exc
+
+    model, tokenizer = _get_mms()
+    inputs = tokenizer(text, return_tensors="pt")
+    with torch.no_grad():
+        audio = model(**inputs).waveform.squeeze().cpu().numpy()
+
+    sample_rate = getattr(model.config, "sampling_rate", 22050)
+    sf.write(wav_path, audio, sample_rate)
+
 
 def _select_korean_voice(engine: pyttsx3.Engine) -> None:
     try:
@@ -37,21 +76,28 @@ def text_to_mp3(text: str, mp3_path: str) -> None:
     mp3_target = Path(mp3_path)
     mp3_target.parent.mkdir(parents=True, exist_ok=True)
 
-    engine = pyttsx3.init()
+    tts_engine = os.getenv("TTS_ENGINE", "pyttsx3").strip().lower()
-    # 음질 개선: 속도/볼륨 조정 및 한국어 음성 우선 선택
-    try:
-        engine.setProperty("rate", 170)
-        engine.setProperty("volume", 1.0)
-    except Exception:
-        pass
-    _select_korean_voice(engine)
     wav_fd, wav_path = tempfile.mkstemp(suffix=".wav")
     os.close(wav_fd)
 
     try:
+        if tts_engine == "mms":
+            _text_to_wav_mms(text, wav_path)
+            audio_filter = "loudnorm=I=-16:LRA=11:TP=-1.5"
+        else:
+            engine = pyttsx3.init()
+            # 음질 개선: 속도/볼륨 조정 및 한국어 음성 우선 선택
+            try:
+                # 서버 음성이 늘어지는 현상 완화
+                engine.setProperty("rate", 210)
+                engine.setProperty("volume", 1.0)
+            except Exception:
+                pass
+            _select_korean_voice(engine)
             # pyttsx3로 wav 생성 후 ffmpeg로 mp3 변환
             engine.save_to_file(text, wav_path)
             engine.runAndWait()
+            audio_filter = "loudnorm=I=-16:LRA=11:TP=-1.5,atempo=1.15"
 
         subprocess.run(
             [
@@ -66,7 +112,7 @@ def text_to_mp3(text: str, mp3_path: str) -> None:
                 "-b:a",
                 "192k",
                 "-af",
-                "loudnorm=I=-16:LRA=11:TP=-1.5",
+                audio_filter,
                 str(mp3_target),
             ],
             check=True,