change model

app.py

@@ -1,42 +1,40 @@
 from fastapi import FastAPI
-from transformers import AutoProcessor
-from peft import PeftModel
-from transformers import CsmForConditionalGeneration
+from fastapi.responses import FileResponse
 import torch
 import soundfile as sf
 import aiofiles
 import uvicorn
 import os
+from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
 
 app = FastAPI()
 
-BASE_MODEL = "deep-io/csm-1b"
-FINETUNE_MODEL = "Ellbendls/csm-1b-indonesian-fine-tuned"
+MODEL_TTS = "microsoft/speecht5_tts"
+MODEL_VOCODER = "microsoft/speecht5_hifigan"
 
 device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
 
-# Load base model
-model = CsmForConditionalGeneration.from_pretrained(BASE_MODEL, torch_dtype=torch.float16).to(device)
+# Load processor, model, and vocoder
+processor = SpeechT5Processor.from_pretrained(MODEL_TTS)
+model = SpeechT5ForTextToSpeech.from_pretrained(MODEL_TTS).to(device)
+vocoder = SpeechT5HifiGan.from_pretrained(MODEL_VOCODER).to(device)
 
-# Apply LoRA adapter
-model = PeftModel.from_pretrained(model, FINETUNE_MODEL)
-
-# Load processor
-processor = AutoProcessor.from_pretrained(BASE_MODEL)
+# Generate random speaker embedding
+speaker_embeddings = torch.randn(1, 512).to(device)
 
 @app.post("/tts")
 async def tts_api(payload: dict):
-    text = payload.get("text", "")
+    text = payload.get("text", "").strip()
     if not text:
         return {"error": "Text is required"}
 
-    inputs = processor(f"[0]{text}", return_tensors="pt").to(device)
+    inputs = processor(text=text, return_tensors="pt").to(device)
+
     with torch.no_grad():
-        outputs = model.generate(**inputs, max_new_tokens=200, output_audio=True)
+        speech = model.generate_speech(inputs["input_ids"], speaker_embeddings, vocoder=vocoder)
 
-    audio = outputs[0].detach().cpu().numpy()
     file_path = "output.wav"
-    sf.write(file_path, audio, 24000)
+    sf.write(file_path, speech.cpu().numpy(), samplerate=16000)
 
     async with aiofiles.open(file_path, "rb") as f:
         audio_data = await f.read()
@@ -51,7 +49,6 @@ async def tts_api(payload: dict):
 async def download_file(filename: str):
     fp = os.path.join(os.getcwd(), filename)
     if os.path.exists(fp):
-        from fastapi.responses import FileResponse
         return FileResponse(fp, media_type="audio/wav", filename=filename)
     return {"error": "File not found"}
 

requirements.txt

@@ -5,5 +5,4 @@ pydantic
 transformers>=4.41.0
 torch
 soundfile
-sentencepiece
-peft
+sentencepiece