Update app.py

app.py

@@ -10,6 +10,8 @@ import os
 from bs4 import BeautifulSoup
 import re
 import numpy as np
+from moviepy.editor import VideoFileClip
+import soundfile as sf
 
 # Load the transcription model
 transcription_pipeline = pipeline("automatic-speech-recognition", model="facebook/wav2vec2-base-960h")
@@ -38,50 +40,38 @@ def download_audio_from_url(url):
         print(f"Error in download_audio_from_url: {str(e)}")
         raise
 
-def transcribe_audio(audio_bytes):
-    audio = AudioSegment.from_file(io.BytesIO(audio_bytes))
-    audio.export("temp_audio.wav", format="wav")
-    waveform, sample_rate = torchaudio.load("temp_audio.wav")
-    os.remove("temp_audio.wav")
-    
-    # Convert torch.Tensor to numpy.ndarray
-    waveform_np = waveform.numpy().squeeze()
-    
-    # Transcribe the audio
-    result = transcription_pipeline(waveform_np, chunk_length_s=30)
-    transcript = result['text']
-    
-    # Split transcript into paragraphs based on silence
-    chunks = split_on_silence(audio, min_silence_len=500, silence_thresh=-40)
-    paragraphs = []
-    current_paragraph = ""
-    
-    for chunk in chunks:
-        chunk.export("temp_chunk.wav", format="wav")
-        waveform_chunk, sample_rate_chunk = torchaudio.load("temp_chunk.wav")
-        os.remove("temp_chunk.wav")
-        
-        # Convert torch.Tensor to numpy.ndarray
-        waveform_chunk_np = waveform_chunk.numpy().squeeze()
-        
-        chunk_result = transcription_pipeline(waveform_chunk_np, chunk_length_s=30)
-        chunk_transcript = chunk_result['text']
-        
-        if chunk_transcript:
-            if current_paragraph:
-                current_paragraph += " " + chunk_transcript
-            else:
-                current_paragraph = chunk_transcript
-        else:
-            if current_paragraph:
-                paragraphs.append(current_paragraph)
-                current_paragraph = ""
-    
-    if current_paragraph:
-        paragraphs.append(current_paragraph)
-    
-    formatted_transcript = "\n\n".join(paragraphs)
-    return formatted_transcript
+def transcribe_audio(video_bytes):
+    try:
+        # Save the video bytes to a temporary file
+        with open("temp_video.mp4", "wb") as f:
+            f.write(video_bytes)
+
+        # Extract audio from video
+        video = VideoFileClip("temp_video.mp4")
+        audio = video.audio
+
+        # Export audio as mono WAV
+        audio.write_audiofile("temp_audio.wav", fps=16000, nbytes=2, codec='pcm_s16le')
+
+        # Load the audio file
+        audio_data, sample_rate = sf.read("temp_audio.wav")
+
+        # Ensure audio is mono
+        if len(audio_data.shape) > 1:
+            audio_data = audio_data.mean(axis=1)
+
+        # Transcribe the audio
+        result = transcription_pipeline(audio_data, sampling_rate=sample_rate)
+        transcript = result['text']
+
+        # Clean up temporary files
+        os.remove("temp_video.mp4")
+        os.remove("temp_audio.wav")
+
+        return transcript
+    except Exception as e:
+        print(f"Error in transcribe_audio: {str(e)}")
+        raise
 
 def transcribe_video(url):
     try: