Update app.py

app.py

@@ -145,8 +145,8 @@ def response_gen(user_review):
     inputs = tokenizer(prompt, return_tensors="pt")  # Tokenize the input prompt
     outputs = model.generate(
         **inputs,
-        max_new_tokens=300,
-        min_length=75,  # Ensure concise and complete responses
+        max_new_tokens=200,
+        min_length=50,  # Ensure concise and complete responses
         no_repeat_ngram_size=2,  # Avoid repetitive phrases
         temperature=0.7  # Add randomness for more natural responses
     )
@@ -168,9 +168,19 @@ def sound_gen(response):
     embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")  # Load speaker embeddings
     speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)  # Use a default embedding
 
-    inputs = processor(text=response, return_tensors="pt")  # Process text for spectrogram generation
-    inputs["input_ids"] = inputs["input_ids"].to(torch.int32)  # Match tensor format (fix runtime error)
-    spectrogram = model.generate_speech(inputs["input_ids"], speaker_embeddings)  # Generate the spectrogram
+    # Ensure the response is not too long for the model's capacity
+    max_tokens = 300  # Limit the input text tokens to a maximum of 300
+    truncated_response = response[:max_tokens]  # Truncate the response to fit within the limit
+
+    # Process the truncated text for spectrogram generation
+    inputs = processor(text=truncated_response, return_tensors="pt")# Process text for spectrogram generation
+    inputs_embeds_size = inputs["input_ids"].size(1)
+
+    # Ensure tensor dimensions align between input IDs and speaker embeddings
+    speaker_embeddings = speaker_embeddings[:, :inputs_embeds_size]  # Match dimensions with input IDs
+
+    # Generate the spectrogram using the SpeechT5 model
+    spectrogram = model.generate_speech(inputs["input_ids"], speaker_embeddings) # Generate the spectrogram
 
     with torch.no_grad():
         speech = vocoder(spectrogram)  # Convert spectrogram to waveform