change model

app.py

@@ -1,5 +1,5 @@
 from fastapi import FastAPI
-from transformers import VitsModel, AutoTokenizer
+from transformers import CsmForConditionalGeneration, AutoProcessor
 import torch
 import aiofiles
 import uvicorn
@@ -8,29 +8,20 @@ import soundfile as sf
 
 app = FastAPI()
 
-# Load model once at startup
-model_name = "Ellbendls/csm-1b-indonesian-fine-tuned"
-model = VitsModel.from_pretrained(model_name)
-tokenizer = AutoTokenizer.from_pretrained(model_name)
+# Load model and processor once
+MODEL_NAME = "Ellbendls/csm-1b-indonesian-fine-tuned"
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model = CsmForConditionalGeneration.from_pretrained(MODEL_NAME).to(device)
+processor = AutoProcessor.from_pretrained(MODEL_NAME)
 
 def generate_timestamps(script: str, wpm: int = 150):
     sentences = [s.strip() for s in script.replace("\n", " ").split(".") if s.strip()]
-    timestamps = []
-    current_time = 0.0
-    seconds_per_word = 60.0 / wpm
-
-    for sentence in sentences:
-        word_count = len(sentence.split())
-        duration = round(word_count * seconds_per_word, 2)
-        start_time = round(current_time, 2)
-        end_time = round(current_time + duration, 2)
-        timestamps.append({
-            "sentence": sentence,
-            "start": start_time,
-            "end": end_time
-        })
-        current_time += duration
-
+    timestamps, current = [], 0.0
+    spw = 60.0 / wpm
+    for s in sentences:
+        dur = round(len(s.split()) * spw, 2)
+        timestamps.append({"sentence": s, "start": round(current,2), "end": round(current+dur,2)})
+        current += dur
     return timestamps
 
 @app.get("/")
@@ -38,19 +29,20 @@ async def root():
     return {"message": "API is running!"}
 
 @app.post("/tts")
-async def text_to_speech(payload: dict):
+async def tts_api(payload: dict):
     text = payload.get("text", "")
-
     if not text:
         return {"error": "Text is required"}
 
     file_path = "output.wav"
-    inputs = tokenizer(text, return_tensors="pt")
-    with torch.no_grad():
-        output = model(**inputs).waveform
-    sf.write(file_path, output.squeeze().cpu().numpy(), model.config.sampling_rate)
+    input_text = f"[0]{text}"
+    inputs = processor(input_text, return_tensors="pt").to(device)
+    outputs = model.generate(**inputs, max_new_tokens=200, output_audio=True)
+    audio = outputs[0].detach().cpu().numpy()
+
+    sf.write(file_path, audio, 24000)
 
-    async with aiofiles.open(file_path, mode="rb") as f:
+    async with aiofiles.open(file_path, "rb") as f:
         audio_data = await f.read()
 
     return {
@@ -62,10 +54,10 @@ async def text_to_speech(payload: dict):
 
 @app.get("/download/{filename}")
 async def download_file(filename: str):
-    file_path = os.path.join(os.getcwd(), filename)
-    if os.path.exists(file_path):
+    fp = os.path.join(os.getcwd(), filename)
+    if os.path.exists(fp):
         from fastapi.responses import FileResponse
-        return FileResponse(file_path, media_type="audio/wav", filename=filename)
+        return FileResponse(fp, media_type="audio/wav", filename=filename)
     return {"error": "File not found"}
 
 if __name__ == "__main__":

requirements.txt

@@ -4,6 +4,5 @@ aiofiles
 pydantic
 transformers
 torch
-torchaudio
 soundfile
 sentencepiece