Update app.py

app.py

@@ -11,22 +11,60 @@ from typing import Optional
 torch.set_num_threads(int(os.environ.get("TORCH_THREADS", "2")))
 
 HERE = os.path.dirname(os.path.abspath(__file__))
+VOICES_DIR = os.path.join(HERE, "voices")
+AUDIO_EXTS = (".wav", ".mp3", ".flac", ".ogg")
 SECRET = os.environ.get("TTS_SECRET_KEY", "")           # khali he to koi key nahi lagegi
+DEFAULT_VOICE = os.environ.get("DEFAULT_VOICE", "")     # chahe to default awaaz ka naam yaha daalo
 MODEL_REPO = "saryps-labs/pocket-tts-hindi"
 LOCAL_DIR = "/home/user/hindi_model"
-VOICE_FILE = os.path.join(HERE, "voice.wav")            # apni (permission wali) awaaz ka 10-20 sec sample
 FALLBACK_VOICE = "hf://kyutai/tts-voices/alba-mackenna/casual.wav"
 MAX_CHARS = int(os.environ.get("MAX_CHARS", "6000"))
 GAP_SECONDS = 0.22
 
 model = None
-voice_state = None
 SR = 24000
 state_msg = "loading"
 last_error = ""
 gen_lock = threading.Lock()
+voice_cache = {}
 
 
+# ---------------- voices ----------------
+def discover_voices() -> dict:
+    """voices/ folder aur root me padi saari audio files -> {naam: path}"""
+    found = {}
+    for folder in (VOICES_DIR, HERE):
+        if not os.path.isdir(folder):
+            continue
+        for fn in sorted(os.listdir(folder)):
+            if fn.lower().endswith(AUDIO_EXTS):
+                base = os.path.splitext(fn)[0]
+                name, i = base, 2
+                while name in found:
+                    name = f"{base}_{i}"
+                    i += 1
+                found[name] = os.path.join(folder, fn)
+    return found
+
+
+def resolve_voice(name: Optional[str]):
+    voices = discover_voices()
+    if not voices:
+        return "default", FALLBACK_VOICE
+    if not name or name not in voices:
+        name = DEFAULT_VOICE if DEFAULT_VOICE in voices else next(iter(voices))
+    return name, voices[name]
+
+
+def get_voice_state(path: str):
+    key = path if path.startswith("hf://") else f"{path}:{os.path.getmtime(path)}"
+    if key not in voice_cache:
+        print("Voice process ho rahi he:", path)
+        voice_cache[key] = model.get_state_for_audio_prompt(path)
+    return voice_cache[key]
+
+
+# ---------------- model ----------------
 def get_config_path() -> str:
     cfg = os.path.join(LOCAL_DIR, "config.yaml")
     if not os.path.exists(cfg):
@@ -43,13 +81,13 @@ def try_load(TTSModel, arg):
 
 
 def load_everything():
-    global model, voice_state, SR, state_msg, last_error
+    global model, SR, state_msg, last_error
     try:
         from pocket_tts import TTSModel
         cfg = get_config_path()
         print("Config:", cfg)
         m, err = None, None
-        for arg in (Path(cfg), cfg):            # pehle Path object, phir string
+        for arg in (Path(cfg), cfg):
             try:
                 m = try_load(TTSModel, arg)
                 break
@@ -58,11 +96,10 @@ def load_everything():
                 traceback.print_exc()
         if m is None:
             raise err
-        voice_path = VOICE_FILE if os.path.exists(VOICE_FILE) else FALLBACK_VOICE
-        print("Voice prompt:", voice_path)
-        voice_state = m.get_state_for_audio_prompt(voice_path)
-        SR = int(getattr(m, "sample_rate", 24000))
         model = m
+        SR = int(getattr(m, "sample_rate", 24000))
+        _, path = resolve_voice(DEFAULT_VOICE)
+        get_voice_state(path)                 # default awaaz pehle se taiyar
         state_msg = "ready"
         print("Pocket TTS Hindi ready, sample rate", SR)
     except Exception as e:
@@ -85,7 +122,7 @@ app.add_middleware(
 
 def authorized(request: Request) -> bool:
     if not SECRET:
-        return True          # koi secret set nahi he -> sab ko allow
+        return True
     cands = [
         request.headers.get("x-secret-key", ""),
         request.query_params.get("key", ""),
@@ -96,8 +133,8 @@ def authorized(request: Request) -> bool:
     return any(hmac.compare_digest(c.encode(), SECRET.encode()) for c in cands if c)
 
 
+# ---------------- text / audio ----------------
 def split_clean(text: str):
-    """Is model ko punctuation nahi chahiye: vakyon me todo, sirf Devanagari + digits rakho."""
     text = text.replace("\r", "\n")
     parts = re.split(r"[\u0964\u0965\.\?\!\n;:]+", text)
     out = []
@@ -116,12 +153,12 @@ def split_clean(text: str):
     return out
 
 
-def iter_audio(text: str):
+def iter_audio(text: str, vstate):
     if hasattr(model, "generate_audio_stream"):
-        for a in model.generate_audio_stream(voice_state, text):
+        for a in model.generate_audio_stream(vstate, text):
             yield a
     else:
-        yield model.generate_audio(voice_state, text)
+        yield model.generate_audio(vstate, text)
 
 
 def to_pcm16(a) -> bytes:
@@ -130,12 +167,12 @@ def to_pcm16(a) -> bytes:
     return (arr * 32767.0).astype("<i2").tobytes()
 
 
-def pcm_stream(sentences):
+def pcm_stream(sentences, vstate):
     silence = np.zeros(int(GAP_SECONDS * SR), dtype="<i2").tobytes()
     with gen_lock:                        # ek samay me ek hi generation
         for s in sentences:
             try:
-                for a in iter_audio(s):
+                for a in iter_audio(s, vstate):
                     yield to_pcm16(a)
                 yield silence
             except Exception:
@@ -143,9 +180,10 @@ def pcm_stream(sentences):
                 traceback.print_exc()
 
 
+# ---------------- API ----------------
 class TTSReq(BaseModel):
     text: str
-    voice: Optional[str] = None      # HTML bhejta he, par is model me ignore hota he
+    voice: Optional[str] = None
     speed: Optional[float] = None
 
 
@@ -159,6 +197,15 @@ def health():
     return {"state": state_msg, "error": last_error}
 
 
+@app.get("/voices")
+def voices(request: Request):
+    if not authorized(request):
+        return JSONResponse({"error": "unauthorized"}, status_code=401)
+    names = list(discover_voices().keys()) or ["default"]
+    default, _ = resolve_voice(DEFAULT_VOICE)
+    return {"voices": names, "default": default}
+
+
 @app.post("/tts")
 def tts(req: TTSReq, request: Request):
     if not authorized(request):
@@ -172,8 +219,15 @@ def tts(req: TTSReq, request: Request):
     sentences = split_clean(req.text)
     if not sentences:
         return JSONResponse({"error": "text me Hindi (देवनागरी) shabd nahi mile"}, status_code=400)
+    name, path = resolve_voice(req.voice)
+    try:
+        with gen_lock:                    # nayi awaaz pehli baar process hogi (kuch second lagenge)
+            vstate = get_voice_state(path)
+    except Exception as e:
+        traceback.print_exc()
+        return JSONResponse({"error": f"awaaz '{name}' load nahi hui: {e}"}, status_code=500)
     return StreamingResponse(
-        pcm_stream(sentences),
+        pcm_stream(sentences, vstate),
         media_type="application/octet-stream",
         headers={"X-Sample-Rate": str(SR), "Cache-Control": "no-store"},
     )