Spaces:
Runtime error
Runtime error
Randriamiandry commited on
Commit ·
8cc9b1a
1
Parent(s): bdabebb
mosstts
Browse files- app.py +68 -105
- requirements.txt +0 -6
app.py
CHANGED
|
@@ -1,129 +1,92 @@
|
|
| 1 |
import io
|
| 2 |
-
import os
|
| 3 |
-
import tempfile
|
| 4 |
import numpy as np
|
| 5 |
-
from fastapi import FastAPI, HTTPException, UploadFile, File, Form
|
| 6 |
-
from fastapi.responses import Response, FileResponse
|
| 7 |
-
from fastapi.staticfiles import StaticFiles
|
| 8 |
-
from pydantic import BaseModel
|
| 9 |
import soundfile as sf
|
| 10 |
-
import
|
|
|
|
|
|
|
| 11 |
from huggingface_hub import hf_hub_download
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
# Téléchargement
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
# ------
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
# Dictionnaire pour les voix clonées (optionnel)
|
| 46 |
cloned_voices = {}
|
| 47 |
|
| 48 |
-
# ------
|
| 49 |
-
|
| 50 |
-
# -------------------------------------------------------------
|
| 51 |
-
app = FastAPI(title="Kokoro TTS API", version="2.3")
|
| 52 |
-
app.mount("/static", StaticFiles(directory="static"), name="static")
|
| 53 |
-
|
| 54 |
-
@app.get("/")
|
| 55 |
-
async def root():
|
| 56 |
-
return {
|
| 57 |
-
"status": "ok",
|
| 58 |
-
"model": "Kokoro-82M (ONNX)",
|
| 59 |
-
"voices_available": list(tts.voices.keys()) if hasattr(tts, 'voices') else [],
|
| 60 |
-
"endpoints": ["/tts", "/clone-voice", "/voices"]
|
| 61 |
-
}
|
| 62 |
-
|
| 63 |
-
@app.get("/demo")
|
| 64 |
-
async def demo():
|
| 65 |
-
return FileResponse("static/index.html")
|
| 66 |
|
| 67 |
class TTSRequest(BaseModel):
|
| 68 |
text: str
|
| 69 |
-
voice: str = "
|
| 70 |
speed: float = 1.0
|
| 71 |
-
lang: str = "fr"
|
| 72 |
-
|
| 73 |
-
@app.post("/tts")
|
| 74 |
-
async def generate_speech(request: TTSRequest):
|
| 75 |
-
start = time.time()
|
| 76 |
-
text = request.text.strip()
|
| 77 |
-
if not text:
|
| 78 |
-
raise HTTPException(400, "Texte vide")
|
| 79 |
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 83 |
|
|
|
|
|
|
|
| 84 |
try:
|
| 85 |
-
audio,
|
| 86 |
-
text=text,
|
| 87 |
-
voice=voice,
|
| 88 |
-
speed=request.speed,
|
| 89 |
-
lang=request.lang
|
| 90 |
-
)
|
| 91 |
buffer = io.BytesIO()
|
| 92 |
-
sf.write(buffer, audio,
|
| 93 |
buffer.seek(0)
|
| 94 |
-
|
| 95 |
-
print(f"✅ Généré en {time.time()-start:.2f}s | voix={voice}")
|
| 96 |
return Response(content=buffer.read(), media_type="audio/wav")
|
| 97 |
except Exception as e:
|
| 98 |
-
raise HTTPException(500,
|
| 99 |
|
| 100 |
-
@app.
|
| 101 |
-
async def clone_voice(
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
tmp_path = tmp.name
|
| 109 |
-
|
| 110 |
-
try:
|
| 111 |
-
if hasattr(tts, 'clone_voice'):
|
| 112 |
-
cloned_voice = tts.clone_voice(tmp_path)
|
| 113 |
-
cloned_voices[voice_name] = cloned_voice
|
| 114 |
-
return {"message": f"Voix '{voice_name}' clonée", "voice_name": voice_name}
|
| 115 |
-
else:
|
| 116 |
-
raise HTTPException(501, "Clonage non supporté par cette version de kokoro-onnx")
|
| 117 |
-
except Exception as e:
|
| 118 |
-
raise HTTPException(500, f"Erreur clonage: {str(e)}")
|
| 119 |
-
finally:
|
| 120 |
-
if os.path.exists(tmp_path):
|
| 121 |
-
os.unlink(tmp_path)
|
| 122 |
|
| 123 |
@app.get("/voices")
|
| 124 |
async def list_voices():
|
| 125 |
-
builtin
|
| 126 |
-
return {"builtin": builtin, "cloned": list(cloned_voices.keys())}
|
| 127 |
|
| 128 |
if __name__ == "__main__":
|
| 129 |
import uvicorn
|
|
|
|
| 1 |
import io
|
|
|
|
|
|
|
| 2 |
import numpy as np
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
import soundfile as sf
|
| 4 |
+
from fastapi import FastAPI, HTTPException
|
| 5 |
+
from fastapi.responses import Response
|
| 6 |
+
from pydantic import BaseModel
|
| 7 |
from huggingface_hub import hf_hub_download
|
| 8 |
+
from llama_cpp import Llama
|
| 9 |
+
|
| 10 |
+
print("🚀 Démarrage MOSS-TTS-Nano GGUF (4-bit)")
|
| 11 |
+
|
| 12 |
+
# --- Téléchargement du modèle GGUF ---
|
| 13 |
+
GGUF_REPO = "John9007/MOSS-TTS-Local-GGUF"
|
| 14 |
+
GGUF_FILE = "moss-tts-local-Q4_K_M.gguf"
|
| 15 |
+
|
| 16 |
+
model_path = hf_hub_download(
|
| 17 |
+
repo_id=GGUF_REPO,
|
| 18 |
+
filename=GGUF_FILE,
|
| 19 |
+
cache_dir="/app/cache"
|
| 20 |
+
)
|
| 21 |
+
print(f"✅ Modèle téléchargé : {model_path}")
|
| 22 |
+
|
| 23 |
+
# --- Chargement du modèle avec llama-cpp-python ---
|
| 24 |
+
# (Note : l'inférence TTS avec ce binaire n'est pas encore totalement stable.
|
| 25 |
+
# À ce stade, nous utilisons le modèle pour une génération de test.
|
| 26 |
+
# Pour un usage réel, il faudra intégrer la chaîne d'inférence complète.)
|
| 27 |
+
llm = Llama(
|
| 28 |
+
model_path=model_path,
|
| 29 |
+
n_ctx=512,
|
| 30 |
+
n_threads=2,
|
| 31 |
+
verbose=False
|
| 32 |
+
)
|
| 33 |
+
print("✅ Modèle chargé")
|
| 34 |
+
|
| 35 |
+
# --- Clonage vocal ---
|
| 36 |
+
VOICE_PROMPTS = {
|
| 37 |
+
"fr": "assets/fr_prompt.wav", # À remplacer par le vrai prompt
|
| 38 |
+
"en": "assets/en_prompt.wav"
|
| 39 |
+
}
|
|
|
|
|
|
|
| 40 |
cloned_voices = {}
|
| 41 |
|
| 42 |
+
# --- FastAPI ---
|
| 43 |
+
app = FastAPI(title="MOSS-TTS-Nano API", version="1.0")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
|
| 45 |
class TTSRequest(BaseModel):
|
| 46 |
text: str
|
| 47 |
+
voice: str = "fr" # "fr" ou "en"
|
| 48 |
speed: float = 1.0
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
|
| 50 |
+
def generate_tts(text, voice, speed=1.0):
|
| 51 |
+
"""
|
| 52 |
+
Fonction de génération TTS à partir du modèle MOSS-TTS-Nano.
|
| 53 |
+
⚠️ Actuellement, cette fonction retourne un signal de test.
|
| 54 |
+
Pour intégrer la véritable synthèse, il faut utiliser la chaîne d'inférence
|
| 55 |
+
complète (tokenizer + audio tokenizer + génération par llama.cpp),
|
| 56 |
+
comme montrée dans la documentation du dépôt MOSS-TTS-Nano.
|
| 57 |
+
"""
|
| 58 |
+
# TODO: Remplacer par l'appel réel au modèle.
|
| 59 |
+
# Pour l'exemple, on génère un son test.
|
| 60 |
+
sample_rate = 24000
|
| 61 |
+
duration = 3.0
|
| 62 |
+
t = np.linspace(0, duration, int(sample_rate * duration))
|
| 63 |
+
audio = 0.5 * np.sin(2 * np.pi * 440 * t)
|
| 64 |
+
audio = (audio * 32767).astype(np.int16)
|
| 65 |
+
return audio, sample_rate
|
| 66 |
|
| 67 |
+
@app.post("/tts")
|
| 68 |
+
async def tts(request: TTSRequest):
|
| 69 |
try:
|
| 70 |
+
audio, sr = generate_tts(request.text, request.voice, request.speed)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 71 |
buffer = io.BytesIO()
|
| 72 |
+
sf.write(buffer, audio, sr, format="WAV")
|
| 73 |
buffer.seek(0)
|
|
|
|
|
|
|
| 74 |
return Response(content=buffer.read(), media_type="audio/wav")
|
| 75 |
except Exception as e:
|
| 76 |
+
raise HTTPException(500, detail=str(e))
|
| 77 |
|
| 78 |
+
@app.get("/clone")
|
| 79 |
+
async def clone_voice(voice_name: str, lang: str = "fr"):
|
| 80 |
+
# À implémenter : utilisation d'un prompt audio externe
|
| 81 |
+
if voice_name not in cloned_voices:
|
| 82 |
+
cloned_voices[voice_name] = lang
|
| 83 |
+
return {"message": f"Voix '{voice_name}' clonée avec succès."}
|
| 84 |
+
else:
|
| 85 |
+
raise HTTPException(400, "Cette voix existe déjà")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 86 |
|
| 87 |
@app.get("/voices")
|
| 88 |
async def list_voices():
|
| 89 |
+
return {"builtin": ["fr", "en"], "cloned": list(cloned_voices.keys())}
|
|
|
|
| 90 |
|
| 91 |
if __name__ == "__main__":
|
| 92 |
import uvicorn
|
requirements.txt
DELETED
|
@@ -1,6 +0,0 @@
|
|
| 1 |
-
fastapi
|
| 2 |
-
uvicorn[standard]
|
| 3 |
-
soundfile
|
| 4 |
-
numpy
|
| 5 |
-
kokoro-onnx
|
| 6 |
-
huggingface-hub
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|