Update App.py

App.py

@@ -9,29 +9,42 @@ from pydantic import BaseModel
 torch.set_num_threads(int(os.environ.get("TORCH_THREADS", "2")))
 
 HERE = os.path.dirname(os.path.abspath(__file__))
-SECRET = os.environ.get("TTS_SECRET_KEY", "")          # Space Settings -> Secrets me daalo
-CONFIG = os.environ.get("TTS_CONFIG", "hf://saryps-labs/pocket-tts-hindi/config.yaml")
+SECRET = os.environ.get("TTS_SECRET_KEY", "")           # Space Settings -> Secrets me daalo
+MODEL_REPO = "saryps-labs/pocket-tts-hindi"
+LOCAL_DIR = "/home/user/hindi_model"
 VOICE_FILE = os.path.join(HERE, "voice.wav")
 FALLBACK_VOICE = "hf://kyutai/tts-voices/alba-mackenna/casual.wav"
-QUANTIZE = os.environ.get("TTS_QUANTIZE", "0") == "1"   # int8 (thoda tez), default band
+QUANTIZE = os.environ.get("TTS_QUANTIZE", "0") == "1"
 MAX_CHARS = int(os.environ.get("MAX_CHARS", "6000"))
-GAP_SECONDS = 0.22                                       # vakyon ke beech ka pause
+GAP_SECONDS = 0.22
 
 model = None
 voice_state = None
 SR = 24000
 state_msg = "loading"
+last_error = ""
 gen_lock = threading.Lock()
 
 
+def get_config_path() -> str:
+    cfg = os.environ.get("TTS_CONFIG") or os.path.join(LOCAL_DIR, "config.yaml")
+    if not os.path.exists(cfg):
+        from huggingface_hub import snapshot_download
+        snapshot_download(MODEL_REPO, local_dir=LOCAL_DIR)
+        cfg = os.path.join(LOCAL_DIR, "config.yaml")
+    return cfg
+
+
 def load_everything():
-    global model, voice_state, SR, state_msg
+    global model, voice_state, SR, state_msg, last_error
     try:
         from pocket_tts import TTSModel
+        cfg = get_config_path()
+        print("Config:", cfg)
         try:
-            m = TTSModel.load_model(CONFIG, temp=0.3, lsd_decode_steps=1, eos_threshold=-1.0)
+            m = TTSModel.load_model(cfg, temp=0.3, lsd_decode_steps=1, eos_threshold=-1.0)
         except TypeError:
-            m = TTSModel.load_model(CONFIG)
+            m = TTSModel.load_model(cfg)
         if QUANTIZE:
             try:
                 m = torch.quantization.quantize_dynamic(m, {torch.nn.Linear}, dtype=torch.qint8)
@@ -45,8 +58,9 @@ def load_everything():
         model = m
         state_msg = "ready"
         print("Model ready, sample rate", SR)
-    except Exception:
+    except Exception as e:
         state_msg = "error"
+        last_error = f"{type(e).__name__}: {e}"
         traceback.print_exc()
 
 
@@ -81,12 +95,10 @@ def split_clean(text: str):
     parts = re.split(r"[\u0964\u0965\.\?\!\n;:]+", text)
     out = []
     for p in parts:
-        # sirf Devanagari (danda chhodkar), digits aur space rakho
         p = re.sub(r"[^\u0900-\u0963\u0966-\u097F0-9\s]", " ", p)
         p = re.sub(r"\s+", " ", p).strip()
         if len(p) < 3:
             continue
-        # bahut lambe vakya ko ~200 chars ke tukdon me todo
         while len(p) > 220:
             cut = p.rfind(" ", 0, 200)
             cut = cut if cut > 50 else 200
@@ -113,7 +125,7 @@ def to_pcm16(a) -> bytes:
 
 def pcm_stream(sentences):
     silence = np.zeros(int(GAP_SECONDS * SR), dtype="<i2").tobytes()
-    with gen_lock:                      # ek samay me ek hi generation (CPU bachane ke liye)
+    with gen_lock:
         for s in sentences:
             try:
                 for a in iter_audio(s):
@@ -135,7 +147,7 @@ def root():
 
 @app.get("/health")
 def health():
-    return {"state": state_msg}
+    return {"state": state_msg, "error": last_error}
 
 
 @app.post("/tts")
@@ -145,7 +157,7 @@ def tts(req: TTSReq, request: Request):
     if state_msg == "loading":
         return JSONResponse({"error": "model abhi load ho raha he, 1 minute baad try karo"}, status_code=503)
     if state_msg == "error" or model is None:
-        return JSONResponse({"error": "model load fail hua, Space ke logs dekho"}, status_code=500)
+        return JSONResponse({"error": "model load fail hua: " + last_error}, status_code=500)
     if len(req.text) > MAX_CHARS:
         return JSONResponse({"error": f"text bahut lamba he (max {MAX_CHARS} characters)"}, status_code=413)
     sentences = split_clean(req.text)