Update app.py

app.py

@@ -114,32 +114,25 @@ def processFile(filePath):
     print("Speakers Detected")
     totalTimeInSeconds = int(waveform_gain_adjusted.shape[-1]/sampleRate)
     print("Time in seconds calculated")
-    return annotations, totalTimeInSeconds
+    return annotations, totalTimeInSeconds, waveform_gain_adjusted, sampleRate
 
-def generate_speaker_clips(annotations, audio_path, file_index):
+def generate_speaker_clips(annotations, waveform, sample_rate, file_index):
     """
     For each unique speaker in `annotations`, find their longest contiguous segment.
     If that segment is >= 5s, clip exactly 5s from its start.
     If < 5s, take the whole segment.
+    Accepts the already-loaded waveform tensor and sample_rate from processFile,
+    so this never needs to re-decode the audio file (avoids torchcodec/FFmpeg issues).
     Saves clips as WAV bytes in st.session_state.speakerClips[file_index].
-    Only works when a real audio file is provided (skips silently otherwise).
     """
-    if audio_path is None or not os.path.exists(audio_path):
-        return
+    import io
 
     # Extend speakerClips list to cover this file index
     while len(st.session_state.speakerClips) <= file_index:
         st.session_state.speakerClips.append({})
 
-    try:
-        waveform, sample_rate = torchaudio.load(audio_path)
-    except Exception as e:
-        print(f"generate_speaker_clips: could not load audio: {e}")
-        return
-
     clips = {}
     for speaker in annotations.labels():
-        # Collect all segments for this speaker
         speaker_segments = [
             segment for segment, _, label in annotations.itertracks(yield_label=True)
             if label == speaker
@@ -147,22 +140,17 @@ def generate_speaker_clips(annotations, audio_path, file_index):
         if not speaker_segments:
             continue
 
-        # Find the longest segment
         longest = max(speaker_segments, key=lambda s: s.duration)
 
-        # Determine clip boundaries
         clip_start = longest.start
         clip_duration = min(longest.duration, 5.0)
         clip_end = clip_start + clip_duration
 
         start_sample = int(clip_start * sample_rate)
-        end_sample = int(clip_end * sample_rate)
-        end_sample = min(end_sample, waveform.shape[-1])
+        end_sample = min(int(clip_end * sample_rate), waveform.shape[-1])
 
         clip_waveform = waveform[:, start_sample:end_sample]
 
-        # Write to an in-memory bytes buffer as WAV
-        import io
         buffer = io.BytesIO()
         torchaudio.save(buffer, clip_waveform, sample_rate, format="wav")
         buffer.seek(0)
@@ -556,7 +544,7 @@ else:
                         st.session_state.unusedSpeakers[i] = speakerNames
                 else:
                     with st.spinner(text=f'Processing File {i+1} of {totalFiles}'):
-                        annotations, totalSeconds = processFile(file_paths[i])
+                        annotations, totalSeconds, waveform, sample_rate = processFile(file_paths[i])
                         print(f"Finished processing {file_paths[i]}")
                         st.session_state.results[i] = (annotations, totalSeconds)
                         print("Results saved")
@@ -566,7 +554,8 @@ else:
                         st.session_state.unusedSpeakers[i] = speakerNames 
                         print("Speakers saved")
                     with st.spinner(text=f'Generating speaker clips for File {i+1} of {totalFiles}'):
-                        generate_speaker_clips(annotations, file_paths[i], i)
+                        generate_speaker_clips(annotations, waveform, sample_rate, i)
+                        del waveform  # free memory after clipping
                         print(f"Speaker clips generated for {file_paths[i]}")
                 with st.spinner(text=f'Analyzing File {i+1} of {totalFiles}'):
                     analyze(file_names[i])