Update app.py

app.py

@@ -43,7 +43,7 @@ This demo allows you to have a conversation with Sesame CSM 1B, leveraging Whisp
 # --- Constants ---  (Constants can stay outside)
 SPEAKER_ID = 0
 MAX_CONTEXT_SEGMENTS = 5
-MAX_GEMMA_LENGTH = 300
+MAX_GEMMA_LENGTH = 150
 
 # --- Global Conversation History ---
 conversation_history = []
@@ -62,7 +62,7 @@ def transcribe_audio(audio_path: str, whisper_model) -> str:  # Pass whisper_mod
 
 def generate_response(text: str, model_gemma, tokenizer_gemma, device) -> str: # Pass model and tokenizer
     try:
-        input_text = "Here is a response to the user: " + text
+        input_text = "Reapond to the user: " + text
         input = tokenizer_gemma(input_text, return_tensors="pt").to(device)
         generated_output = model_gemma.generate(**input, max_length=MAX_GEMMA_LENGTH, early_stopping=True)
         return tokenizer_gemma.decode(generated_output[0], skip_special_tokens=True)
@@ -110,7 +110,7 @@ def infer(user_audio) -> tuple[int, np.ndarray]:
         whisper_model = whisper.load_model("small.en", device=device)
         logging.info("Whisper model loaded successfully.")
 
-        tokenizer_gemma = AutoTokenizer.from_pretrained("google/gemma-3-1b-pt")
+        tokenizer_gemma = AutoTokenizer.from_pretrained("google/gemma-3-1b-it")
         model_gemma = AutoModelForCausalLM.from_pretrained("google/gemma-3-1b-it").to(device)
         logging.info("Gemma 3 1B pt model loaded successfully.")