Upload app.py

app.py

@@ -22,6 +22,9 @@ VOICE_LIST = [
     "michael", "paul", "vera"
 ]
 
+# Cache voice states so we don't recompute voice prompt embeddings every request
+voice_state_cache = {}
+
 @spaces.GPU
 def generate_tts(text: str, voice: str):
     clean_text = text.strip()
@@ -31,21 +34,15 @@ def generate_tts(text: str, voice: str):
     m = get_model()
     clean_voice = voice.replace("pocket-", "").replace("-", "_")
     
-    # Initialize the model state with the specified voice
-    if hasattr(m, "init_model_state"):
+    # Get or compute voice state for the voice prompt
+    if clean_voice not in voice_state_cache:
         try:
-            model_state = m.init_model_state(voice=clean_voice)
-        except TypeError:
-            model_state = m.init_model_state(clean_voice)
-    elif hasattr(m, "get_voice_state"):
-        model_state = m.get_voice_state(clean_voice)
-    else:
-        # Check all available methods on m for state initialization
-        init_fn = getattr(m, "get_initial_state", getattr(m, "new_state", None))
-        if init_fn:
-            model_state = init_fn(clean_voice)
-        else:
-            raise RuntimeError(f"Available methods on TTSModel: {[fn for fn in dir(m) if not fn.startswith('_')]}")
+            voice_state_cache[clean_voice] = m.get_state_for_audio_prompt(clean_voice)
+        except Exception:
+            # If string voice identifier is passed directly or with fallback
+            voice_state_cache[clean_voice] = m.get_state_for_audio_prompt(f"voices/{clean_voice}.wav") if hasattr(m, "get_state_for_audio_prompt") else {}
+
+    model_state = voice_state_cache[clean_voice]
 
     # Generate audio
     audio = m.generate_audio(model_state, clean_text)