Update app.py

app.py

@@ -29,7 +29,118 @@ 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).to(device)
 
-# ... (keep all the existing functions as they are)
+def download_audio_from_url(url):
+    try:
+        if "youtube.com" in url or "youtu.be" in url:
+            print("Processing YouTube URL...")
+            yt = YouTube(url)
+            audio_stream = yt.streams.filter(only_audio=True).first()
+            with tempfile.NamedTemporaryFile(delete=False, suffix=".mp4") as temp_file:
+                audio_stream.download(output_path=temp_file.name)
+                audio_bytes = open(temp_file.name, "rb").read()
+            os.unlink(temp_file.name)
+        elif "share" in url:
+            print("Processing shareable link...")
+            response = requests.get(url)
+            soup = BeautifulSoup(response.content, 'html.parser')
+            video_tag = soup.find('video')
+            if video_tag and 'src' in video_tag.attrs:
+                video_url = video_tag['src']
+                print(f"Extracted video URL: {video_url}")
+            else:
+                raise ValueError("Direct video URL not found in the shareable link.")
+            response = requests.get(video_url)
+            audio_bytes = response.content
+        else:
+            print(f"Downloading video from URL: {url}")
+            response = requests.get(url)
+            audio_bytes = response.content
+        
+        print(f"Successfully downloaded {len(audio_bytes)} bytes of data")
+        return audio_bytes
+    except Exception as e:
+        print(f"Error in download_audio_from_url: {str(e)}")
+        raise
+
+def transcribe_audio(audio_file):
+    try:
+        print("Loading audio file...")
+        audio = AudioSegment.from_file(audio_file)
+        audio = audio.set_channels(1).set_frame_rate(16000)
+        audio_array = torch.tensor(audio.get_array_of_samples()).float()
+        
+        print(f"Audio duration: {len(audio) / 1000:.2f} seconds")
+        print("Starting transcription...")
+        input_features = whisper_processor(audio_array, sampling_rate=16000, return_tensors="pt").input_features.to(device)
+        
+        # Create attention mask
+        attention_mask = torch.ones_like(input_features)
+        
+        # Generate with specific parameters
+        predicted_ids = whisper_model.generate(
+            input_features,
+            attention_mask=attention_mask,
+            language='en',
+            task='translate'
+        )
+        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)}")
+        raise
+
+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.
+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.
+
+Now, please process the following transcribed text:
+
+{transcription}
+"""
+    
+    inputs = qwen_tokenizer(prompt, return_tensors="pt").to(device)
+    with torch.no_grad():
+        outputs = qwen_model.generate(**inputs, max_new_tokens=4000)
+    result = qwen_tokenizer.decode(outputs[0], skip_special_tokens=True)
+    
+    # Extract the processed text (remove the instruction part)
+    processed_text = result.split("Now, please process the following transcribed text:")[-1].strip()
+    
+    print("Speaker separation complete.")
+    return processed_text
+
+def transcribe_video(url):
+    try:
+        print(f"Attempting to download audio from URL: {url}")
+        audio_bytes = download_audio_from_url(url)
+        print(f"Successfully downloaded {len(audio_bytes)} bytes of audio data")
+        
+        with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_audio:
+            AudioSegment.from_file(io.BytesIO(audio_bytes)).export(temp_audio.name, format="wav")
+            transcript = transcribe_audio(temp_audio.name)
+        
+        os.unlink(temp_audio.name)
+        
+        if len(transcript) < 10:
+            raise ValueError("Transcription too short, possibly failed")
+        
+        print("Separating speakers...")
+        separated_transcript = separate_speakers(transcript)
+        
+        return separated_transcript
+    except Exception as e:
+        error_message = f"An error occurred: {str(e)}"
+        print(error_message)
+        return error_message
 
 app = dash.Dash(__name__, external_stylesheets=[dbc.themes.BOOTSTRAP])