Upload app.py with huggingface_hub

app.py

@@ -9,12 +9,13 @@ from fastapi import FastAPI, Request
 from fastapi.middleware.cors import CORSMiddleware
 from fastapi.responses import HTMLResponse
 import uvicorn
+import scipy.io.wavfile
 
-# --- [v146] 🚀 ZEROGPU OPTIMIZED ENGINE ---
-print(f"--- [v146] 📡 BOOTING ZEROGPU ENGINE ---")
+# --- [v147] 🚀 HYBRID ENGINE: GPU-STT + CPU-TTS ---
+print(f"--- [v147] 📡 BOOTING HYBRID ENGINE ---")
 
-from transformers import pipeline, AutoModelForSpeechSeq2Seq, AutoProcessor
-from TTS.api import TTS
+from transformers import pipeline, SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
+from datasets import load_dataset
 from deep_translator import GoogleTranslator
 
 try:
@@ -35,24 +36,31 @@ os.environ["PYTHONWARNINGS"] = "ignore"
 app = FastAPI()
 app.add_middleware(CORSMiddleware, allow_origins=["*"], allow_methods=["*"], allow_headers=["*"])
 
-MODELS = {"stt": None, "tts": None}
+MODELS = {"stt": None, "tts_processor": None, "tts_model": None, "tts_vocoder": None, "tts_speaker": None}
 
-@spaces.GPU(duration=120)
+# Load TTS at startup (CPU)
+print("--- [v147] 📥 LOADING TTS (SpeechT5) ON CPU ---")
+MODELS["tts_processor"] = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
+MODELS["tts_model"] = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts")
+MODELS["tts_vocoder"] = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
+embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
+MODELS["tts_speaker"] = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)
+print("--- [v147] ✅ TTS READY (CPU) ---")
+
+@spaces.GPU(duration=60)
 def process_stt(audio_b64, lang):
-    """Speech-to-Text on GPU"""
+    """Speech-to-Text on GPU (Whisper)"""
     global MODELS
     
-    # Load Whisper if needed
     if MODELS.get("stt") is None:
-        print("--- [v146] 📥 LOADING WHISPER-BASE (STABLE) ---")
-        # Use whisper-base for stability on ZeroGPU
+        print("--- [v147] 📥 LOADING WHISPER ON GPU ---")
         MODELS["stt"] = pipeline(
             "automatic-speech-recognition",
             model="openai/whisper-base",
             device="cuda",
-            torch_dtype=torch.float32  # FP32 for stability
+            torch_dtype=torch.float32
         )
-        print("--- [v146] ✅ WHISPER-BASE LOADED ---")
+        print("--- [v147] ✅ WHISPER READY (GPU) ---")
     
     audio_bytes = base64.b64decode(audio_b64)
     with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
@@ -65,45 +73,19 @@ def process_stt(audio_b64, lang):
         if os.path.exists(temp_path):
             os.unlink(temp_path)
 
-@spaces.GPU(duration=120)
-def process_tts(text, target_lang, speaker_wav_b64=None):
-    """Text-to-Speech on GPU"""
-    global MODELS
-    
-    # Load XTTS if needed
-    if MODELS.get("tts") is None:
-        print("--- [v146] 📥 LOADING XTTS ---")
-        MODELS["tts"] = TTS("tts_models/multilingual/multi-dataset/xtts_v2").to("cuda")
-        print("--- [v146] ✅ XTTS LOADED ---")
-    
-    XTTS_MAP = {"en": "en", "de": "de", "fr": "fr", "es": "es", "it": "it", "pl": "pl", "pt": "pt", "tr": "tr", "ru": "ru", "nl": "nl", "cs": "cs", "ar": "ar", "hu": "hu", "ko": "ko", "hi": "hi", "zh": "zh-cn"}
-    clean_lang = target_lang.split('-')[0].lower()
-    mapped_lang = XTTS_MAP.get(clean_lang) or ("zh-cn" if clean_lang == "zh" else None)
+def process_tts(text):
+    """Text-to-Speech on CPU (SpeechT5)"""
+    inputs = MODELS["tts_processor"](text=text, return_tensors="pt")
+    speech = MODELS["tts_model"].generate_speech(inputs["input_ids"], MODELS["tts_speaker"], vocoder=MODELS["tts_vocoder"])
     
-    if not mapped_lang:
-        return {"error": f"Language {clean_lang} not supported."}
+    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as out_f:
+        out_p = out_f.name
+    scipy.io.wavfile.write(out_p, rate=16000, data=speech.numpy())
     
-    speaker_wav_path = "default_speaker.wav"
-    if speaker_wav_b64:
-        sb = base64.b64decode(speaker_wav_b64)
-        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as f:
-            f.write(sb)
-            speaker_wav_path = f.name
-    elif not os.path.exists(speaker_wav_path):
-        speaker_wav_path = None
-
-    try:
-        with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as out_f:
-            out_p = out_f.name
-        MODELS["tts"].tts_to_file(text=text, language=mapped_lang, file_path=out_p, speaker_wav=speaker_wav_path)
-        with open(out_p, "rb") as f:
-            audio_b64 = base64.b64encode(f.read()).decode()
-        return audio_b64
-    finally:
-        if speaker_wav_b64 and os.path.exists(speaker_wav_path):
-            os.unlink(speaker_wav_path)
-        if 'out_p' in locals() and os.path.exists(out_p):
-            os.unlink(out_p)
+    with open(out_p, "rb") as f:
+        audio_b64 = base64.b64encode(f.read()).decode()
+    os.unlink(out_p)
+    return audio_b64
 
 @app.post("/process")
 async def api_process(request: Request):
@@ -111,9 +93,9 @@ async def api_process(request: Request):
         data = await request.json()
         action = data.get("action")
         if action == "health":
-            return {"status": "awake", "v": "146", "mode": "ZEROGPU"}
+            return {"status": "awake", "v": "147", "mode": "HYBRID_GPU_STT_CPU_TTS"}
         
-        print(f"--- [v146] 🛠️ PROCESSING: {action} ---")
+        print(f"--- [v147] 🛠️ PROCESSING: {action} ---")
         t1 = time.time()
         
         stt_text = None
@@ -133,27 +115,25 @@ async def api_process(request: Request):
             if len(text) < 2:
                 return {"audio": ""} if action == "tts" else {"text": stt_text, "translated": "", "audio": ""}
             
-            audio_res = process_tts(text, data.get("lang") if action == "tts" else target, data.get("speaker_wav"))
+            audio_b64 = process_tts(text)
             
-            if isinstance(audio_res, dict) and "error" in audio_res:
-                return audio_res
             if action == "tts":
-                return {"audio": audio_res}
-            return {"text": stt_text, "translated": trans_text, "audio": audio_res}
+                return {"audio": audio_b64}
+            return {"text": stt_text, "translated": trans_text, "audio": audio_b64}
 
     except Exception as e:
-        print(f"❌ [v146] ERROR: {traceback.format_exc()}")
+        print(f"❌ [v147] ERROR: {traceback.format_exc()}")
         return {"error": str(e)}
     finally:
-        print(f"--- [v146] ✨ DONE ({time.time()-t1:.1f}s) ---")
+        print(f"--- [v147] ✨ DONE ({time.time()-t1:.1f}s) ---")
 
 @app.get("/health")
 def health():
-    return {"status": "ok", "v": "146", "mode": "ZEROGPU", "spaces": HAS_SPACES}
+    return {"status": "ok", "v": "147", "mode": "HYBRID_GPU_STT_CPU_TTS", "spaces": HAS_SPACES}
 
 @app.get("/", response_class=HTMLResponse)
 def root():
-    return f"<html><body><h1>🚀 AI Engine v146 (ZEROGPU)</h1><p>Spaces: {HAS_SPACES}</p></body></html>"
+    return f"<html><body><h1>🚀 AI Engine v147 (HYBRID)</h1><p>STT: GPU Whisper | TTS: CPU SpeechT5</p></body></html>"
 
 if __name__ == "__main__":
     uvicorn.run(app, host="0.0.0.0", port=7860)