Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -133,7 +133,7 @@ def handle_feedback(feedback):
|
|
133 |
conn.commit()
|
134 |
return "Thank you for your feedback!", None
|
135 |
|
136 |
-
def segment_background_audio(audio_path, background_audio_path="background_segments.wav"):
|
137 |
"""
|
138 |
Uses Demucs to separate audio and extract background (non-vocal) parts.
|
139 |
Merges drums, bass, and other stems into a single background track.
|
@@ -150,6 +150,7 @@ def segment_background_audio(audio_path, background_audio_path="background_segme
|
|
150 |
stem_dir = os.path.join("separated", "htdemucs", filename)
|
151 |
|
152 |
# Step 3: Load and merge background stems
|
|
|
153 |
drums = AudioSegment.from_wav(os.path.join(stem_dir, "drums.wav"))
|
154 |
bass = AudioSegment.from_wav(os.path.join(stem_dir, "bass.wav"))
|
155 |
other = AudioSegment.from_wav(os.path.join(stem_dir, "other.wav"))
|
@@ -158,7 +159,8 @@ def segment_background_audio(audio_path, background_audio_path="background_segme
|
|
158 |
|
159 |
# Step 4: Export the merged background
|
160 |
background.export(background_audio_path, format="wav")
|
161 |
-
|
|
|
162 |
|
163 |
def transcribe_video_with_speakers(video_path):
|
164 |
# Extract audio from video
|
@@ -167,7 +169,7 @@ def transcribe_video_with_speakers(video_path):
|
|
167 |
video.audio.write_audiofile(audio_path)
|
168 |
logger.info(f"Audio extracted from video: {audio_path}")
|
169 |
|
170 |
-
segment_result = segment_background_audio(audio_path)
|
171 |
print(f"Saved non-speech (background) audio to local")
|
172 |
|
173 |
# Set up device
|
@@ -180,7 +182,7 @@ def transcribe_video_with_speakers(video_path):
|
|
180 |
logger.info("WhisperX model loaded")
|
181 |
|
182 |
# Transcribe
|
183 |
-
result = model.transcribe(
|
184 |
logger.info("Audio transcription completed")
|
185 |
|
186 |
# Get the detected language
|
@@ -188,12 +190,12 @@ def transcribe_video_with_speakers(video_path):
|
|
188 |
logger.debug(f"Detected language: {detected_language}")
|
189 |
# Alignment
|
190 |
# model_a, metadata = whisperx.load_align_model(language_code=result["language"], device=device)
|
191 |
-
# result = whisperx.align(result["segments"], model_a, metadata,
|
192 |
# logger.info("Transcription alignment completed")
|
193 |
|
194 |
# Diarization (works independently of Whisper model size)
|
195 |
diarize_model = whisperx.DiarizationPipeline(use_auth_token=hf_api_key, device=device)
|
196 |
-
diarize_segments = diarize_model(
|
197 |
logger.info("Speaker diarization completed")
|
198 |
|
199 |
# Assign speakers
|
@@ -213,31 +215,62 @@ def transcribe_video_with_speakers(video_path):
|
|
213 |
}
|
214 |
for segment in result["segments"]
|
215 |
]
|
216 |
-
|
217 |
# Collect audio for each speaker
|
218 |
speaker_audio = {}
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
224 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
225 |
# Collapse and truncate speaker audio
|
226 |
speaker_sample_paths = {}
|
227 |
-
audio_clip = AudioFileClip(
|
|
|
|
|
|
|
228 |
for speaker, segments in speaker_audio.items():
|
|
|
|
|
229 |
speaker_clips = [audio_clip.subclip(start, end) for start, end in segments]
|
230 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
231 |
truncated_clip = combined_clip.subclip(0, min(30, combined_clip.duration))
|
|
|
|
|
|
|
232 |
sample_path = f"speaker_{speaker}_sample.wav"
|
233 |
truncated_clip.write_audiofile(sample_path)
|
234 |
speaker_sample_paths[speaker] = sample_path
|
235 |
-
logger.info(f"Created sample for {speaker}: {sample_path}")
|
236 |
|
237 |
-
#
|
|
|
238 |
video.close()
|
239 |
audio_clip.close()
|
240 |
-
os.remove(
|
|
|
241 |
|
242 |
return transcript_with_speakers, detected_language
|
243 |
|
@@ -637,6 +670,74 @@ def collapse_ocr_subtitles(ocr_json, text_similarity_threshold=90):
|
|
637 |
|
638 |
return collapsed
|
639 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
640 |
def post_edit_transcribed_segments(transcription_json, video_path,
|
641 |
interval_sec=0.5,
|
642 |
text_similarity_threshold=80,
|
@@ -658,80 +759,12 @@ def post_edit_transcribed_segments(transcription_json, video_path,
|
|
658 |
# Step 2: Collapse repetitive OCR
|
659 |
collapsed_ocr = collapse_ocr_subtitles(ocr_json, text_similarity_threshold=90)
|
660 |
|
661 |
-
# Step 3:
|
662 |
-
|
663 |
-
|
664 |
-
|
665 |
-
for entry_idx, entry in enumerate(transcription_json):
|
666 |
-
start = entry.get("start", 0)
|
667 |
-
end = entry.get("end", 0)
|
668 |
-
base_text = entry.get("text", "")
|
669 |
-
|
670 |
-
best_match_idx = None
|
671 |
-
best_score = -1
|
672 |
-
|
673 |
-
for ocr_idx, ocr in enumerate(collapsed_ocr):
|
674 |
-
time_overlap = not (ocr["end"] < start - time_tolerance or ocr["start"] > end + time_tolerance)
|
675 |
-
if not time_overlap:
|
676 |
-
continue
|
677 |
-
|
678 |
-
sim = fuzz.ratio(ocr["text"], base_text)
|
679 |
-
if sim > best_score:
|
680 |
-
best_score = sim
|
681 |
-
best_match_idx = ocr_idx
|
682 |
-
|
683 |
-
updated_entry = entry.copy()
|
684 |
-
if best_match_idx is not None and best_score >= text_similarity_threshold:
|
685 |
-
updated_entry["text"] = collapsed_ocr[best_match_idx]["text"]
|
686 |
-
updated_entry["ocr_matched"] = True
|
687 |
-
updated_entry["ocr_similarity"] = best_score
|
688 |
-
used_ocr_indices.add(best_match_idx)
|
689 |
-
else:
|
690 |
-
updated_entry["ocr_matched"] = False
|
691 |
-
updated_entry["ocr_similarity"] = best_score if best_score >= 0 else None
|
692 |
-
|
693 |
-
merged_segments.append(updated_entry)
|
694 |
-
|
695 |
-
# Step 4: Insert unused OCR segments (Phase 2)
|
696 |
-
inserted_segments = []
|
697 |
-
for ocr_idx, ocr in enumerate(collapsed_ocr):
|
698 |
-
if ocr_idx in used_ocr_indices:
|
699 |
-
continue
|
700 |
-
|
701 |
-
# Check for fuzzy duplicates in WhisperX
|
702 |
-
duplicate = False
|
703 |
-
for whisper_seg in transcription_json:
|
704 |
-
if abs(ocr["start"] - whisper_seg["start"]) < time_tolerance or abs(ocr["end"] - whisper_seg["end"]) < time_tolerance:
|
705 |
-
sim = fuzz.ratio(ocr["text"], whisper_seg["text"])
|
706 |
-
if sim >= text_similarity_threshold:
|
707 |
-
duplicate = True
|
708 |
-
break
|
709 |
-
|
710 |
-
if duplicate:
|
711 |
-
logger.debug(f"🟡 Skipping near-duplicate OCR: '{ocr['text']}'")
|
712 |
-
continue
|
713 |
-
|
714 |
-
# Infer speaker from nearest WhisperX entry
|
715 |
-
nearby = sorted(transcription_json, key=lambda x: abs(x["start"] - ocr["start"]))
|
716 |
-
speaker_guess = nearby[0].get("speaker", "unknown") if nearby else "unknown"
|
717 |
-
|
718 |
-
inserted_segment = {
|
719 |
-
"start": ocr["start"],
|
720 |
-
"end": ocr["end"],
|
721 |
-
"text": ocr["text"],
|
722 |
-
"speaker": speaker_guess
|
723 |
-
}
|
724 |
-
inserted_segments.append(inserted_segment)
|
725 |
-
|
726 |
-
# Step 5: Combine and sort
|
727 |
-
final_segments = merged_segments + inserted_segments
|
728 |
-
final_segments = sorted(final_segments, key=lambda x: x["start"])
|
729 |
-
|
730 |
-
print(f"✅ Post-editing completed: {len(final_segments)} total segments "
|
731 |
-
f"({len(inserted_segments)} OCR-inserted segments)")
|
732 |
-
|
733 |
-
return final_segments
|
734 |
|
|
|
|
|
735 |
|
736 |
def process_entry(entry, i, tts_model, video_width, video_height, process_mode, target_language, font_path, speaker_sample_paths=None):
|
737 |
logger.debug(f"Processing entry {i}: {entry}")
|
|
|
133 |
conn.commit()
|
134 |
return "Thank you for your feedback!", None
|
135 |
|
136 |
+
def segment_background_audio(audio_path, background_audio_path="background_segments.wav", speech_audio_path="speech_segment.wav"):
|
137 |
"""
|
138 |
Uses Demucs to separate audio and extract background (non-vocal) parts.
|
139 |
Merges drums, bass, and other stems into a single background track.
|
|
|
150 |
stem_dir = os.path.join("separated", "htdemucs", filename)
|
151 |
|
152 |
# Step 3: Load and merge background stems
|
153 |
+
vocals = AudioSegment.from_wav(os.path.join(stem_dir, "vocals.wav"))
|
154 |
drums = AudioSegment.from_wav(os.path.join(stem_dir, "drums.wav"))
|
155 |
bass = AudioSegment.from_wav(os.path.join(stem_dir, "bass.wav"))
|
156 |
other = AudioSegment.from_wav(os.path.join(stem_dir, "other.wav"))
|
|
|
159 |
|
160 |
# Step 4: Export the merged background
|
161 |
background.export(background_audio_path, format="wav")
|
162 |
+
vocals.export(speech_audio_path, format="wav")
|
163 |
+
return background_audio_path, speech_audio_path
|
164 |
|
165 |
def transcribe_video_with_speakers(video_path):
|
166 |
# Extract audio from video
|
|
|
169 |
video.audio.write_audiofile(audio_path)
|
170 |
logger.info(f"Audio extracted from video: {audio_path}")
|
171 |
|
172 |
+
segment_result, speech_audio_path = segment_background_audio(audio_path)
|
173 |
print(f"Saved non-speech (background) audio to local")
|
174 |
|
175 |
# Set up device
|
|
|
182 |
logger.info("WhisperX model loaded")
|
183 |
|
184 |
# Transcribe
|
185 |
+
result = model.transcribe(speech_audio_path, chunk_size=6, print_progress = True)
|
186 |
logger.info("Audio transcription completed")
|
187 |
|
188 |
# Get the detected language
|
|
|
190 |
logger.debug(f"Detected language: {detected_language}")
|
191 |
# Alignment
|
192 |
# model_a, metadata = whisperx.load_align_model(language_code=result["language"], device=device)
|
193 |
+
# result = whisperx.align(result["segments"], model_a, metadata, speech_audio_path, device)
|
194 |
# logger.info("Transcription alignment completed")
|
195 |
|
196 |
# Diarization (works independently of Whisper model size)
|
197 |
diarize_model = whisperx.DiarizationPipeline(use_auth_token=hf_api_key, device=device)
|
198 |
+
diarize_segments = diarize_model(speech_audio_path)
|
199 |
logger.info("Speaker diarization completed")
|
200 |
|
201 |
# Assign speakers
|
|
|
215 |
}
|
216 |
for segment in result["segments"]
|
217 |
]
|
218 |
+
|
219 |
# Collect audio for each speaker
|
220 |
speaker_audio = {}
|
221 |
+
logger.info("🔎 Start collecting valid audio segments per speaker...")
|
222 |
+
|
223 |
+
for idx, segment in enumerate(result["segments"]):
|
224 |
+
speaker = segment.get("speaker", "SPEAKER_00")
|
225 |
+
start = segment["start"]
|
226 |
+
end = segment["end"]
|
227 |
+
|
228 |
+
if end > start and (end - start) > 0.05: # Require >50ms duration
|
229 |
+
if speaker not in speaker_audio:
|
230 |
+
speaker_audio[speaker] = [(start, end)]
|
231 |
+
else:
|
232 |
+
speaker_audio[speaker].append((start, end))
|
233 |
+
|
234 |
+
logger.debug(f"Segment {idx}: Added to speaker {speaker} [{start:.2f}s → {end:.2f}s]")
|
235 |
+
else:
|
236 |
+
logger.warning(f"⚠️ Segment {idx} discarded: invalid duration ({start:.2f}s → {end:.2f}s)")
|
237 |
+
|
238 |
# Collapse and truncate speaker audio
|
239 |
speaker_sample_paths = {}
|
240 |
+
audio_clip = AudioFileClip(speech_audio_path)
|
241 |
+
|
242 |
+
logger.info(f"🔎 Found {len(speaker_audio)} speakers with valid segments. Start creating speaker samples...")
|
243 |
+
|
244 |
for speaker, segments in speaker_audio.items():
|
245 |
+
logger.info(f"🔹 Speaker {speaker}: {len(segments)} valid segments")
|
246 |
+
|
247 |
speaker_clips = [audio_clip.subclip(start, end) for start, end in segments]
|
248 |
+
if not speaker_clips:
|
249 |
+
logger.warning(f"⚠️ No valid audio clips for speaker {speaker}. Skipping sample creation.")
|
250 |
+
continue
|
251 |
+
|
252 |
+
if len(speaker_clips) == 1:
|
253 |
+
logger.debug(f"Speaker {speaker}: Only one clip, skipping concatenation.")
|
254 |
+
combined_clip = speaker_clips[0]
|
255 |
+
else:
|
256 |
+
logger.debug(f"Speaker {speaker}: Concatenating {len(speaker_clips)} clips.")
|
257 |
+
combined_clip = concatenate_audioclips(speaker_clips)
|
258 |
+
|
259 |
truncated_clip = combined_clip.subclip(0, min(30, combined_clip.duration))
|
260 |
+
logger.debug(f"Speaker {speaker}: Truncated to {truncated_clip.duration:.2f} seconds.")
|
261 |
+
|
262 |
+
# Step 4: Save the final result
|
263 |
sample_path = f"speaker_{speaker}_sample.wav"
|
264 |
truncated_clip.write_audiofile(sample_path)
|
265 |
speaker_sample_paths[speaker] = sample_path
|
266 |
+
logger.info(f"✅ Created and saved sample for {speaker}: {sample_path}")
|
267 |
|
268 |
+
# Cleanup
|
269 |
+
logger.info("🧹 Closing audio clip and removing temporary files...")
|
270 |
video.close()
|
271 |
audio_clip.close()
|
272 |
+
os.remove(speech_audio_path)
|
273 |
+
logger.info("✅ Finished processing all speaker samples.")
|
274 |
|
275 |
return transcript_with_speakers, detected_language
|
276 |
|
|
|
670 |
|
671 |
return collapsed
|
672 |
|
673 |
+
def merge_speaker_and_time_from_whisperx(ocr_json, whisperx_json, text_sim_threshold=80, replace_threshold=90):
|
674 |
+
"""
|
675 |
+
Given OCR and WhisperX segments, merge speaker ID and optionally replace time.
|
676 |
+
"""
|
677 |
+
merged = []
|
678 |
+
|
679 |
+
for ocr in ocr_json:
|
680 |
+
ocr_start = ocr["start"]
|
681 |
+
ocr_end = ocr["end"]
|
682 |
+
ocr_text = ocr["text"]
|
683 |
+
|
684 |
+
best_match = None
|
685 |
+
best_score = -1
|
686 |
+
|
687 |
+
for wx in whisperx_json:
|
688 |
+
wx_start, wx_end = wx["start"], wx["end"]
|
689 |
+
wx_text = wx["text"]
|
690 |
+
|
691 |
+
# Time overlap (soft constraint)
|
692 |
+
time_center_diff = abs((ocr_start + ocr_end)/2 - (wx_start + wx_end)/2)
|
693 |
+
if time_center_diff > 3: # skip if too far
|
694 |
+
continue
|
695 |
+
|
696 |
+
# Text similarity
|
697 |
+
sim = fuzz.ratio(ocr_text, wx_text)
|
698 |
+
if sim > best_score:
|
699 |
+
best_score = sim
|
700 |
+
best_match = wx
|
701 |
+
|
702 |
+
new_entry = copy.deepcopy(ocr)
|
703 |
+
if best_match:
|
704 |
+
new_entry["speaker"] = best_match.get("speaker", "UNKNOWN")
|
705 |
+
new_entry["ocr_similarity"] = best_score
|
706 |
+
|
707 |
+
if best_score >= replace_threshold:
|
708 |
+
new_entry["start"] = best_match["start"]
|
709 |
+
new_entry["end"] = best_match["end"]
|
710 |
+
|
711 |
+
else:
|
712 |
+
new_entry["speaker"] = "UNKNOWN"
|
713 |
+
new_entry["ocr_similarity"] = None
|
714 |
+
|
715 |
+
merged.append(new_entry)
|
716 |
+
|
717 |
+
return merged
|
718 |
+
|
719 |
+
|
720 |
+
def realign_ocr_segments(merged_ocr_json, min_gap=0.2):
|
721 |
+
"""
|
722 |
+
Realign OCR segments to avoid overlaps using midpoint-based adjustment.
|
723 |
+
"""
|
724 |
+
merged_ocr_json = sorted(merged_ocr_json, key=lambda x: x["start"])
|
725 |
+
|
726 |
+
for i in range(1, len(merged_ocr_json)):
|
727 |
+
prev = merged_ocr_json[i - 1]
|
728 |
+
curr = merged_ocr_json[i]
|
729 |
+
|
730 |
+
# If current overlaps with previous, adjust
|
731 |
+
if curr["start"] < prev["end"] + min_gap:
|
732 |
+
midpoint = (prev["end"] + curr["start"]) / 2
|
733 |
+
prev["end"] = round(midpoint - min_gap / 2, 3)
|
734 |
+
curr["start"] = round(midpoint + min_gap / 2, 3)
|
735 |
+
|
736 |
+
# Prevent negative durations
|
737 |
+
if curr["start"] >= curr["end"]:
|
738 |
+
curr["end"] = round(curr["start"] + 0.3, 3)
|
739 |
+
|
740 |
+
return merged_ocr_json
|
741 |
def post_edit_transcribed_segments(transcription_json, video_path,
|
742 |
interval_sec=0.5,
|
743 |
text_similarity_threshold=80,
|
|
|
759 |
# Step 2: Collapse repetitive OCR
|
760 |
collapsed_ocr = collapse_ocr_subtitles(ocr_json, text_similarity_threshold=90)
|
761 |
|
762 |
+
# Step 3: Merge and realign OCR segments.
|
763 |
+
ocr_merged = merge_speaker_and_time_from_whisperx(ocr_json, whisperx_json)
|
764 |
+
ocr_realigned = realign_ocr_segments(ocr_merged)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
765 |
|
766 |
+
logger.info(f"✅ Final merged and realigned OCR: {len(ocr_realigned)} segments")
|
767 |
+
return ocr_realigned
|
768 |
|
769 |
def process_entry(entry, i, tts_model, video_width, video_height, process_mode, target_language, font_path, speaker_sample_paths=None):
|
770 |
logger.debug(f"Processing entry {i}: {entry}")
|