Spaces:
Sleeping
Sleeping
Commit
•
84a138e
1
Parent(s):
a96698f
Update diarization.py
Browse files- diarization.py +81 -83
diarization.py
CHANGED
@@ -1,83 +1,81 @@
|
|
1 |
-
from pyannote.audio import Pipeline
|
2 |
-
from pydub import AudioSegment
|
3 |
-
import os
|
4 |
-
import re
|
5 |
-
import torch
|
6 |
-
|
7 |
-
def perform_diarization(audio_file_path, translated_file_path, output_dir='./audio/diarization'):
|
8 |
-
|
9 |
-
# Initialize diarization pipeline
|
10 |
-
accesstoken = os.environ['Diarization']
|
11 |
-
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", use_auth_token=accesstoken )
|
12 |
-
|
13 |
-
# Send pipeline to GPU (when available)
|
14 |
-
pipeline.to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
|
15 |
-
|
16 |
-
# Load audio file
|
17 |
-
audio = AudioSegment.from_wav(audio_file_path)
|
18 |
-
|
19 |
-
# Apply pretrained pipeline
|
20 |
-
diarization = pipeline(audio_file_path)
|
21 |
-
|
22 |
-
os.makedirs(output_dir, exist_ok=True)
|
23 |
-
|
24 |
-
# Process and save each speaker's audio segments
|
25 |
-
speaker_segments_audio = {}
|
26 |
-
for turn, _, speaker in diarization.itertracks(yield_label=True):
|
27 |
-
start_ms = int(turn.start * 1000) # Convert to milliseconds
|
28 |
-
end_ms = int(turn.end * 1000) # Convert to milliseconds
|
29 |
-
segment = audio[start_ms:end_ms]
|
30 |
-
|
31 |
-
if speaker in speaker_segments_audio:
|
32 |
-
speaker_segments_audio[speaker] += segment
|
33 |
-
else:
|
34 |
-
speaker_segments_audio[speaker] = segment
|
35 |
-
|
36 |
-
# Save audio segments
|
37 |
-
for speaker, segment in speaker_segments_audio.items():
|
38 |
-
output_path = os.path.join(output_dir, f"{speaker}.wav")
|
39 |
-
segment.export(output_path, format="wav")
|
40 |
-
print(f"Combined audio for speaker {speaker} saved in {output_path}")
|
41 |
-
|
42 |
-
# Load translated text
|
43 |
-
with open(translated_file_path, "r") as file:
|
44 |
-
translated_lines = file.readlines()
|
45 |
-
|
46 |
-
# Process and align translated text with diarization data
|
47 |
-
last_speaker = None
|
48 |
-
aligned_text = []
|
49 |
-
timestamp_pattern = re.compile(r'\[(\d+\.\d+)\-(\d+\.\d+)\]')
|
50 |
-
for line in translated_lines:
|
51 |
-
match = timestamp_pattern.match(line)
|
52 |
-
|
53 |
-
if match:
|
54 |
-
start_time = float(match.group(1))
|
55 |
-
end_time = float(match.group(2))
|
56 |
-
text = line[match.end():].strip() # Extract text part
|
57 |
-
|
58 |
-
speaker_found = False
|
59 |
-
# Find corresponding speaker
|
60 |
-
for turn, _, speaker in diarization.itertracks(yield_label=True):
|
61 |
-
speaker_start = turn.start
|
62 |
-
speaker_end = turn.end
|
63 |
-
# Check for overlap between speaker segment and line timestamp
|
64 |
-
if max(speaker_start, start_time) < min(speaker_end, end_time):
|
65 |
-
aligned_text.append(f"[{speaker}] [{start_time}-{end_time}] {text}")
|
66 |
-
speaker_found = True
|
67 |
-
last_speaker = speaker
|
68 |
-
break
|
69 |
-
|
70 |
-
# If no speaker found, use the last speaker
|
71 |
-
if not speaker_found:
|
72 |
-
if last_speaker is not None:
|
73 |
-
aligned_text.append(f"[{last_speaker}] [{start_time}-{end_time}] {text}")
|
74 |
-
else:
|
75 |
-
aligned_text.append(f"[Unknown Speaker] [{start_time}-{end_time}] {text}")
|
76 |
-
|
77 |
-
# Save aligned text to a single file
|
78 |
-
aligned_text_output_path = os.path.join(output_dir, "aligned_text.txt")
|
79 |
-
with open(aligned_text_output_path, "w") as aligned_text_file:
|
80 |
-
aligned_text_file.write('\n'.join(aligned_text))
|
81 |
-
print(f"Aligned text saved in {aligned_text_output_path}")
|
82 |
-
|
83 |
-
# The rest of your script, if any
|
|
|
1 |
+
from pyannote.audio import Pipeline
|
2 |
+
from pydub import AudioSegment
|
3 |
+
import os
|
4 |
+
import re
|
5 |
+
import torch
|
6 |
+
|
7 |
+
def perform_diarization(audio_file_path, translated_file_path, output_dir='./audio/diarization'):
|
8 |
+
|
9 |
+
# Initialize diarization pipeline
|
10 |
+
accesstoken = os.environ['Diarization']
|
11 |
+
pipeline = Pipeline.from_pretrained("pyannote/speaker-diarization-3.1", use_auth_token=accesstoken )
|
12 |
+
|
13 |
+
# Send pipeline to GPU (when available)
|
14 |
+
pipeline.to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
|
15 |
+
|
16 |
+
# Load audio file
|
17 |
+
audio = AudioSegment.from_wav(audio_file_path)
|
18 |
+
|
19 |
+
# Apply pretrained pipeline
|
20 |
+
diarization = pipeline(audio_file_path)
|
21 |
+
|
22 |
+
os.makedirs(output_dir, exist_ok=True)
|
23 |
+
|
24 |
+
# Process and save each speaker's audio segments
|
25 |
+
speaker_segments_audio = {}
|
26 |
+
for turn, _, speaker in diarization.itertracks(yield_label=True):
|
27 |
+
start_ms = int(turn.start * 1000) # Convert to milliseconds
|
28 |
+
end_ms = int(turn.end * 1000) # Convert to milliseconds
|
29 |
+
segment = audio[start_ms:end_ms]
|
30 |
+
|
31 |
+
if speaker in speaker_segments_audio:
|
32 |
+
speaker_segments_audio[speaker] += segment
|
33 |
+
else:
|
34 |
+
speaker_segments_audio[speaker] = segment
|
35 |
+
|
36 |
+
# Save audio segments
|
37 |
+
for speaker, segment in speaker_segments_audio.items():
|
38 |
+
output_path = os.path.join(output_dir, f"{speaker}.wav")
|
39 |
+
segment.export(output_path, format="wav")
|
40 |
+
print(f"Combined audio for speaker {speaker} saved in {output_path}")
|
41 |
+
|
42 |
+
# Load translated text
|
43 |
+
with open(translated_file_path, "r") as file:
|
44 |
+
translated_lines = file.readlines()
|
45 |
+
|
46 |
+
# Process and align translated text with diarization data
|
47 |
+
last_speaker = None
|
48 |
+
aligned_text = []
|
49 |
+
timestamp_pattern = re.compile(r'\[(\d+\.\d+)\-(\d+\.\d+)\]')
|
50 |
+
for line in translated_lines:
|
51 |
+
match = timestamp_pattern.match(line)
|
52 |
+
|
53 |
+
if match:
|
54 |
+
start_time = float(match.group(1))
|
55 |
+
end_time = float(match.group(2))
|
56 |
+
text = line[match.end():].strip() # Extract text part
|
57 |
+
|
58 |
+
speaker_found = False
|
59 |
+
# Find corresponding speaker
|
60 |
+
for turn, _, speaker in diarization.itertracks(yield_label=True):
|
61 |
+
speaker_start = turn.start
|
62 |
+
speaker_end = turn.end
|
63 |
+
# Check for overlap between speaker segment and line timestamp
|
64 |
+
if max(speaker_start, start_time) < min(speaker_end, end_time):
|
65 |
+
aligned_text.append(f"[{speaker}] [{start_time}-{end_time}] {text}")
|
66 |
+
speaker_found = True
|
67 |
+
last_speaker = speaker
|
68 |
+
break
|
69 |
+
|
70 |
+
# If no speaker found, use the last speaker
|
71 |
+
if not speaker_found:
|
72 |
+
if last_speaker is not None:
|
73 |
+
aligned_text.append(f"[{last_speaker}] [{start_time}-{end_time}] {text}")
|
74 |
+
else:
|
75 |
+
aligned_text.append(f"[Unknown Speaker] [{start_time}-{end_time}] {text}")
|
76 |
+
|
77 |
+
# Save aligned text to a single file
|
78 |
+
aligned_text_output_path = os.path.join(output_dir, "aligned_text.txt")
|
79 |
+
with open(aligned_text_output_path, "w") as aligned_text_file:
|
80 |
+
aligned_text_file.write('\n'.join(aligned_text))
|
81 |
+
print(f"Aligned text saved in {aligned_text_output_path}")
|
|
|
|