Update app.py

app.py

@@ -27,7 +27,7 @@ whisper_model = WhisperForConditionalGeneration.from_pretrained(whisper_model_na
 # Load the Qwen model and tokenizer
 qwen_model_name = "Qwen/Qwen2.5-3B-Instruct"
 qwen_tokenizer = AutoTokenizer.from_pretrained(qwen_model_name, trust_remote_code=True)
-qwen_model = AutoModelForCausalLM.from_pretrained(qwen_model_name, trust_remote_code=True, torch_dtype=torch.float16).to(device)
+qwen_model = AutoModelForCausalLM.from_pretrained(qwen_model_name, trust_remote_code=True).to(device)
 
 def download_audio_from_url(url):
     try:
@@ -86,6 +86,8 @@ def transcribe_audio(audio_file):
         transcription = whisper_processor.batch_decode(predicted_ids, skip_special_tokens=True)
         
         print(f"Transcription complete. Length: {len(transcription[0])} characters")
+        if len(transcription[0]) < 10:
+            raise ValueError(f"Transcription too short: {transcription[0]}")
         return transcription[0]
     except Exception as e:
         print(f"Error in transcribe_audio: {str(e)}")
@@ -95,7 +97,7 @@ def separate_speakers(transcription):
     print("Starting speaker separation...")
     prompt = f"""Analyze the following transcribed text and separate it into different speakers. Identify potential speaker changes based on context, content shifts, or dialogue patterns. Format the output as follows:
 
-1. Label speakers as "Speaker 1", "Speaker 2", etc.  You will have to use dialog context to asume which speaker is saying their dialog as that isn't in the text.
+1. Label speakers as "Speaker 1", "Speaker 2", etc.
 2. Start each speaker's text on a new line beginning with their label.
 3. Separate different speakers' contributions with a blank line.
 4. If the same speaker continues, do not insert a blank line or repeat the speaker label.
@@ -106,7 +108,6 @@ Now, please process the following transcribed text:
 """
     
     inputs = qwen_tokenizer(prompt, return_tensors="pt").to(device)
-    inputs = {k: v.to(torch.float16) for k, v in inputs.items()}  # Convert inputs to float16
     with torch.no_grad():
         outputs = qwen_model.generate(**inputs, max_new_tokens=4000)
     result = qwen_tokenizer.decode(outputs[0], skip_special_tokens=True)
@@ -116,7 +117,7 @@ Now, please process the following transcribed text:
     
     print("Speaker separation complete.")
     return processed_text
-    
+
 def transcribe_video(url):
     try:
         print(f"Attempting to download audio from URL: {url}")
@@ -129,6 +130,9 @@ def transcribe_video(url):
         
         os.unlink(temp_audio.name)
         
+        if len(transcript) < 10:
+            raise ValueError("Transcription too short, possibly failed")
+        
         print("Separating speakers...")
         separated_transcript = separate_speakers(transcript)