debug emb cache

.gitignore

@@ -0,0 +1,5 @@
+.DS_Store/
+venv/
+env/
+__Pycache__/
+__pycache__/

app.py

@@ -59,7 +59,14 @@ def generate_speech_gpu(text, model_choice, mode, speaker_choice, t, top_p, rp):
         selected_model = models[model_choice]
 
         # Get speaker embedding based on mode
+        print(f"[generate_speech_gpu] Mode: {mode}, Speaker choice: {speaker_choice}")
         speaker_emb = speaker_manager.get_speaker_emb(mode, speaker_choice)
+        print(f"[generate_speech_gpu] Speaker emb type: {type(speaker_emb)}")
+        if speaker_emb is not None:
+            if isinstance(speaker_emb, str):
+                print(f"[generate_speech_gpu] Speaker emb is path: {speaker_emb}")
+            elif torch.is_tensor(speaker_emb):
+                print(f"[generate_speech_gpu] Speaker emb is tensor: shape={speaker_emb.shape}, device={speaker_emb.device}")
 
         print(f"Generating speech with {model_choice}...")
         audio, _ = selected_model(
@@ -114,7 +121,6 @@ with gr.Blocks(title="😻 KaniTTS - Text to Speech", theme=gr.themes.Ocean()) a
                     type="numpy",
                     sources=["upload", "microphone"],
                     format="wav",
-                    waveform_options={"sample_rate": 16000}
                 )
 
                 with gr.Row():

util.py

@@ -130,13 +130,20 @@ class SpeakerManager:
         """
         if mode == "select":
             if speaker_name and speaker_name in self.speaker_map:
-                return self.speaker_map[speaker_name]
+                path = self.speaker_map[speaker_name]
+                print(f"[SpeakerManager] Returning speaker path: {path}")
+                return path
             return None
         elif mode == "generate":
+            print(f"[SpeakerManager] Cached embedding: {self.cached_embedding}")
+            print(f"[SpeakerManager] Cached embedding type: {type(self.cached_embedding)}")
+            if self.cached_embedding is not None:
+                print(f"[SpeakerManager] Cached embedding shape: {self.cached_embedding.shape}")
+                print(f"[SpeakerManager] Cached embedding device: {self.cached_embedding.device}")
             return self.cached_embedding
         return None
 
-    def generate_embedding(self, audio_data, sample_rate: int = 16000):
+    def generate_embedding(self, audio_data, sample_rate: int):
         """
         Generate speaker embedding from audio data.
 
@@ -154,19 +161,25 @@ class SpeakerManager:
         """
         # Initialize embedder lazily
         if self.embedder is None:
+            print("[SpeakerManager] Initializing SpeakerEmbedder...")
             self.embedder = SpeakerEmbedder()
 
         # Handle Gradio audio format (sr, audio) tuple
         if isinstance(audio_data, tuple):
             sample_rate, audio_array = audio_data
+            print(f"[SpeakerManager] Audio tuple: sr={sample_rate}, shape={audio_array.shape}")
         else:
             audio_array = audio_data
+            print(f"[SpeakerManager] Audio array shape: {audio_array.shape}")
 
         # Generate embedding
+        print("[SpeakerManager] Generating embedding...")
         embedding = self.embedder.embed_audio(audio_array, sample_rate=sample_rate)
+        print(f"[SpeakerManager] Generated embedding shape: {embedding.shape}, device: {embedding.device}")
 
         # Cache the result
         self.cached_embedding = embedding
+        print(f"[SpeakerManager] Cached embedding (id={id(self.cached_embedding)})")
 
         return embedding