def hyp(input_path, output_path, device, batch_size, html, ext, sample_rate,
max_duration):
os.makedirs(output_path, exist_ok=True)
audio_source = ([
(input_path, audio_name) for audio_name in os.listdir(input_path)
] if os.path.isdir(input_path) else [(os.path.dirname(input_path),
os.path.basename(input_path))])
model = PyannoteDiarizationModel(device=device, batch_size=batch_size)
for i, (input_path, audio_name) in enumerate(audio_source):
print(i, '/', len(audio_source), audio_name)
audio_path = os.path.join(input_path, audio_name)
noextname = audio_name[:-len(ext)]
transcript_path = os.path.join(output_path, noextname + '.json')
rttm_path = os.path.join(output_path, noextname + '.rttm')
signal, sample_rate = audio.read_audio(audio_path,
sample_rate=sample_rate,
mono=True,
dtype='float32',
duration=max_duration)
transcript = model(signal,
sample_rate=sample_rate,
extra=dict(audio_path=audio_path))
transcripts.collect_speaker_names(transcript, set_speaker_data=True)
transcripts.save(transcript_path, transcript)
print(transcript_path)
transcripts.save(rttm_path, transcript)
print(rttm_path)
if html:
html_path = os.path.join(output_path, audio_name + '.html')
vis.transcript(html_path,
sample_rate=sample_rate,
mono=True,
transcript=transcript,
duration=max_duration)
def main(args, ext_json=['.json', '.json.gz']):
utils.enable_jit_fusion()
assert args.output_json or args.output_html or args.output_txt or args.output_csv, \
'at least one of the output formats must be provided'
os.makedirs(args.output_path, exist_ok=True)
audio_data_paths = set(
p for f in args.input_path
for p in ([os.path.join(f, g)
for g in os.listdir(f)] if os.path.isdir(f) else [f])
if os.path.isfile(p) and any(map(p.endswith, args.ext)))
json_data_paths = set(
p for p in args.input_path if any(map(p.endswith, ext_json))
and not utils.strip_suffixes(p, ext_json) in audio_data_paths)
data_paths = list(audio_data_paths | json_data_paths)
exclude = set([
os.path.splitext(basename)[0]
for basename in os.listdir(args.output_path)
if basename.endswith('.json')
]) if args.skip_processed else None
data_paths = [
path for path in data_paths
if exclude is None or os.path.basename(path) not in exclude
]
text_pipeline, frontend, model, generator = setup(args)
val_dataset = datasets.AudioTextDataset(
data_paths, [text_pipeline],
args.sample_rate,
frontend=frontend if not args.frontend_in_model else None,
mono=args.mono,
time_padding_multiple=args.batch_time_padding_multiple,
audio_backend=args.audio_backend,
exclude=exclude,
max_duration=args.transcribe_first_n_sec,
mode='batched_channels'
if args.join_transcript else 'batched_transcript',
string_array_encoding=args.dataset_string_array_encoding,
debug_short_long_records_features_from_whole_normalized_signal=args.
debug_short_long_records_features_from_whole_normalized_signal)
print('Examples count: ', len(val_dataset))
val_meta = val_dataset.pop_meta()
val_data_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=None,
collate_fn=val_dataset.collate_fn,
num_workers=args.num_workers)
csv_sep = dict(tab='\t', comma=',')[args.csv_sep]
csv_lines = [] # only used if args.output_csv is True
oom_handler = utils.OomHandler(max_retries=args.oom_retries)
for i, (meta, s, x, xlen, y, ylen) in enumerate(val_data_loader):
print(f'Processing: {i}/{len(val_dataset)}')
meta = [val_meta[t['example_id']] for t in meta]
audio_path = meta[0]['audio_path']
audio_name = transcripts.audio_name(audio_path)
begin_end = [dict(begin=t['begin'], end=t['end']) for t in meta]
begin = torch.tensor([t['begin'] for t in begin_end],
dtype=torch.float)
end = torch.tensor([t['end'] for t in begin_end], dtype=torch.float)
#TODO WARNING assumes frontend not in dataset
if not args.frontend_in_model:
print('\n' * 10 + 'WARNING\n' * 5)
print(
'transcribe.py assumes frontend in model, in other case time alignment was incorrect'
)
print('WARNING\n' * 5 + '\n')
duration = x.shape[-1] / args.sample_rate
channel = [t['channel'] for t in meta]
speaker = [t['speaker'] for t in meta]
speaker_name = [t['speaker_name'] for t in meta]
if x.numel() == 0:
print(f'Skipping empty [{audio_path}].')
continue
try:
tic = time.time()
y, ylen = y.to(args.device), ylen.to(args.device)
log_probs, logits, olen = model(
x.squeeze(1).to(args.device), xlen.to(args.device))
print('Input:', audio_name)
print('Input time steps:', log_probs.shape[-1],
'| target time steps:', y.shape[-1])
print(
'Time: audio {audio:.02f} sec | processing {processing:.02f} sec'
.format(audio=sum(map(transcripts.compute_duration, meta)),
processing=time.time() - tic))
ts: shaping.Bt = duration * torch.linspace(
0, 1, steps=log_probs.shape[-1],
device=log_probs.device).unsqueeze(0).expand(x.shape[0], -1)
ref_segments = [[
dict(channel=channel[i],
begin=begin_end[i]['begin'],
end=begin_end[i]['end'],
ref=text_pipeline.postprocess(
text_pipeline.preprocess(meta[i]['ref'])))
] for i in range(len(meta))]
hyp_segments = [
alternatives[0] for alternatives in generator.generate(
tokenizer=text_pipeline.tokenizer,
log_probs=log_probs,
begin=begin,
end=end,
output_lengths=olen,
time_stamps=ts,
segment_text_key='hyp',
segment_extra_info=[
dict(speaker=s, speaker_name=sn, channel=c)
for s, sn, c in zip(speaker, speaker_name, channel)
])
]
hyp_segments = [
transcripts.map_text(text_pipeline.postprocess, hyp=hyp)
for hyp in hyp_segments
]
hyp, ref = '\n'.join(
transcripts.join(hyp=h)
for h in hyp_segments).strip(), '\n'.join(
transcripts.join(ref=r) for r in ref_segments).strip()
if args.verbose:
print('HYP:', hyp)
print('CER: {cer:.02%}'.format(cer=metrics.cer(hyp=hyp, ref=ref)))
tic_alignment = time.time()
if args.align and y.numel() > 0:
alignment: shaping.BY = ctc.alignment(
log_probs.permute(2, 0, 1),
y[:, 0, :], # assumed that 0 channel is char labels
olen,
ylen[:, 0],
blank=text_pipeline.tokenizer.eps_id,
pack_backpointers=args.pack_backpointers)
aligned_ts: shaping.Bt = ts.gather(1, alignment)
ref_segments = [
alternatives[0] for alternatives in generator.generate(
tokenizer=text_pipeline.tokenizer,
log_probs=torch.nn.functional.one_hot(
y[:,
0, :], num_classes=log_probs.shape[1]).permute(
0, 2, 1),
begin=begin,
end=end,
output_lengths=ylen,
time_stamps=aligned_ts,
segment_text_key='ref',
segment_extra_info=[
dict(speaker=s, speaker_name=sn, channel=c)
for s, sn, c in zip(speaker, speaker_name, channel)
])
]
ref_segments = [
transcripts.map_text(text_pipeline.postprocess, hyp=ref)
for ref in ref_segments
]
oom_handler.reset()
except:
if oom_handler.try_recover(model.parameters()):
print(f'Skipping {i} / {len(val_dataset)}')
continue
else:
raise
print('Alignment time: {:.02f} sec'.format(time.time() -
tic_alignment))
ref_transcript, hyp_transcript = [
sorted(transcripts.flatten(segments), key=transcripts.sort_key)
for segments in [ref_segments, hyp_segments]
]
if args.max_segment_duration:
if ref:
ref_segments = list(
transcripts.segment_by_time(ref_transcript,
args.max_segment_duration))
hyp_segments = list(
transcripts.segment_by_ref(hyp_transcript, ref_segments))
else:
hyp_segments = list(
transcripts.segment_by_time(hyp_transcript,
args.max_segment_duration))
ref_segments = [[] for _ in hyp_segments]
#### HACK for diarization
elif args.ref_transcript_path and args.join_transcript:
audio_name_hack = audio_name.split('.')[0]
#TODO: normalize ref field
ref_segments = [[t] for t in sorted(transcripts.load(
os.path.join(args.ref_transcript_path, audio_name_hack +
'.json')),
key=transcripts.sort_key)]
hyp_segments = list(
transcripts.segment_by_ref(hyp_transcript,
ref_segments,
set_speaker=True,
soft=False))
#### END OF HACK
has_ref = bool(transcripts.join(ref=transcripts.flatten(ref_segments)))
transcript = []
for hyp_transcript, ref_transcript in zip(hyp_segments, ref_segments):
hyp, ref = transcripts.join(hyp=hyp_transcript), transcripts.join(
ref=ref_transcript)
transcript.append(
dict(audio_path=audio_path,
ref=ref,
hyp=hyp,
speaker_name=transcripts.speaker_name(ref=ref_transcript,
hyp=hyp_transcript),
words=metrics.align_words(
*metrics.align_strings(hyp=hyp, ref=ref))
if args.align_words else [],
words_ref=ref_transcript,
words_hyp=hyp_transcript,
**transcripts.summary(hyp_transcript),
**(dict(cer=metrics.cer(hyp=hyp, ref=ref))
if has_ref else {})))
transcripts.collect_speaker_names(transcript,
set_speaker_data=True,
num_speakers=2)
filtered_transcript = list(
transcripts.prune(transcript,
align_boundary_words=args.align_boundary_words,
cer=args.prune_cer,
duration=args.prune_duration,
gap=args.prune_gap,
allowed_unk_count=args.prune_unk,
num_speakers=args.prune_num_speakers))
print('Filtered segments:', len(filtered_transcript), 'out of',
len(transcript))
if args.output_json:
transcript_path = os.path.join(args.output_path,
audio_name + '.json')
print(transcripts.save(transcript_path, filtered_transcript))
if args.output_html:
transcript_path = os.path.join(args.output_path,
audio_name + '.html')
print(
vis.transcript(transcript_path, args.sample_rate, args.mono,
transcript, filtered_transcript))
if args.output_txt:
transcript_path = os.path.join(args.output_path,
audio_name + '.txt')
with open(transcript_path, 'w') as f:
f.write(' '.join(t['hyp'].strip()
for t in filtered_transcript))
print(transcript_path)
if args.output_csv:
assert len({t['audio_path'] for t in filtered_transcript}) == 1
audio_path = filtered_transcript[0]['audio_path']
hyp = ' '.join(t['hyp'].strip() for t in filtered_transcript)
begin = min(t['begin'] for t in filtered_transcript)
end = max(t['end'] for t in filtered_transcript)
csv_lines.append(
csv_sep.join([audio_path, hyp,
str(begin),
str(end)]))
if args.logits:
logits_file_path = os.path.join(args.output_path,
audio_name + '.pt')
if args.logits_crop:
begin_end = [
dict(
zip(['begin', 'end'], [
t['begin'] + c / float(o) * (t['end'] - t['begin'])
for c in args.logits_crop
])) for o, t in zip(olen, begin_end)
]
logits_crop = [slice(*args.logits_crop) for o in olen]
else:
logits_crop = [slice(int(o)) for o in olen]
# TODO: filter ref / hyp by channel?
torch.save([
dict(audio_path=audio_path,
logits=l[..., logits_crop[i]],
**begin_end[i],
ref=ref,
hyp=hyp) for i, l in enumerate(logits.cpu())
], logits_file_path)
print(logits_file_path)
print('Done: {:.02f} sec\n'.format(time.time() - tic))
if args.output_csv:
transcript_path = os.path.join(args.output_path, 'transcripts.csv')
with open(transcript_path, 'w') as f:
f.write('\n'.join(csv_lines))
print(transcript_path)
def main(args):
utils.enable_jit_fusion()
assert args.output_json or args.output_html or args.output_txt or args.output_csv, \
'at least one of the output formats must be provided'
os.makedirs(args.output_path, exist_ok=True)
data_paths = [
p for f in args.input_path
for p in ([os.path.join(f, g)
for g in os.listdir(f)] if os.path.isdir(f) else [f])
if os.path.isfile(p) and any(map(p.endswith, args.ext))
] + [
p
for p in args.input_path if any(map(p.endswith, ['.json', '.json.gz']))
]
exclude = set([
os.path.splitext(basename)[0]
for basename in os.listdir(args.output_path)
if basename.endswith('.json')
] if args.skip_processed else [])
data_paths = [
path for path in data_paths if os.path.basename(path) not in exclude
]
labels, frontend, model, decoder = setup(args)
val_dataset = datasets.AudioTextDataset(
data_paths, [labels],
args.sample_rate,
frontend=None,
segmented=True,
mono=args.mono,
time_padding_multiple=args.batch_time_padding_multiple,
audio_backend=args.audio_backend,
exclude=exclude,
max_duration=args.transcribe_first_n_sec,
join_transcript=args.join_transcript,
string_array_encoding=args.dataset_string_array_encoding)
num_examples = len(val_dataset)
print('Examples count: ', num_examples)
val_data_loader = torch.utils.data.DataLoader(
val_dataset,
batch_size=None,
collate_fn=val_dataset.collate_fn,
num_workers=args.num_workers)
csv_sep = dict(tab='\t', comma=',')[args.csv_sep]
output_lines = [] # only used if args.output_csv is True
oom_handler = utils.OomHandler(max_retries=args.oom_retries)
for i, (meta, s, x, xlen, y, ylen) in enumerate(val_data_loader):
print(f'Processing: {i}/{num_examples}')
meta = [val_dataset.meta.get(m['example_id']) for m in meta]
audio_path = meta[0]['audio_path']
if x.numel() == 0:
print(f'Skipping empty [{audio_path}].')
continue
begin = meta[0]['begin']
end = meta[0]['end']
audio_name = transcripts.audio_name(audio_path)
try:
tic = time.time()
y, ylen = y.to(args.device), ylen.to(args.device)
log_probs, olen = model(
x.squeeze(1).to(args.device), xlen.to(args.device))
#speech = vad.detect_speech(x.squeeze(1), args.sample_rate, args.window_size, aggressiveness = args.vad, window_size_dilate = args.window_size_dilate)
#speech = vad.upsample(speech, log_probs)
#log_probs.masked_fill_(models.silence_space_mask(log_probs, speech, space_idx = labels.space_idx, blank_idx = labels.blank_idx), float('-inf'))
decoded = decoder.decode(log_probs, olen)
print('Input:', audio_name)
print('Input time steps:', log_probs.shape[-1],
'| target time steps:', y.shape[-1])
print(
'Time: audio {audio:.02f} sec | processing {processing:.02f} sec'
.format(audio=sum(
transcripts.compute_duration(t) for t in meta),
processing=time.time() - tic))
ts = (x.shape[-1] / args.sample_rate) * torch.linspace(
0, 1,
steps=log_probs.shape[-1]).unsqueeze(0) + torch.FloatTensor(
[t['begin'] for t in meta]).unsqueeze(1)
channel = [t['channel'] for t in meta]
speaker = [t['speaker'] for t in meta]
ref_segments = [[
dict(channel=channel[i],
begin=meta[i]['begin'],
end=meta[i]['end'],
ref=labels.decode(y[i, 0, :ylen[i]].tolist()))
] for i in range(len(decoded))]
hyp_segments = [
labels.decode(decoded[i],
ts[i],
channel=channel[i],
replace_blank=True,
replace_blank_series=args.replace_blank_series,
replace_repeat=True,
replace_space=False,
speaker=speaker[i] if isinstance(
speaker[i], str) else None)
for i in range(len(decoded))
]
ref, hyp = '\n'.join(
transcripts.join(ref=r)
for r in ref_segments).strip(), '\n'.join(
transcripts.join(hyp=h) for h in hyp_segments).strip()
if args.verbose:
print('HYP:', hyp)
print('CER: {cer:.02%}'.format(cer=metrics.cer(hyp=hyp, ref=ref)))
tic_alignment = time.time()
if args.align and y.numel() > 0:
#if ref_full:# and not ref:
# #assert len(set(t['channel'] for t in meta)) == 1 or all(t['type'] != 'channel' for t in meta)
# #TODO: add space at the end
# channel = torch.ByteTensor(channel).repeat_interleave(log_probs.shape[-1]).reshape(1, -1)
# ts = ts.reshape(1, -1)
# log_probs = log_probs.transpose(0, 1).unsqueeze(0).flatten(start_dim = -2)
# olen = torch.tensor([log_probs.shape[-1]], device = log_probs.device, dtype = torch.long)
# y = y_full[None, None, :].to(y.device)
# ylen = torch.tensor([[y.shape[-1]]], device = log_probs.device, dtype = torch.long)
# segments = [([], sum([h for r, h in segments], []))]
alignment = ctc.alignment(
log_probs.permute(2, 0, 1),
y.squeeze(1),
olen,
ylen.squeeze(1),
blank=labels.blank_idx,
pack_backpointers=args.pack_backpointers)
ref_segments = [
labels.decode(y[i, 0, :ylen[i]].tolist(),
ts[i],
alignment[i],
channel=channel[i],
speaker=speaker[i],
key='ref',
speakers=val_dataset.speakers)
for i in range(len(decoded))
]
oom_handler.reset()
except:
if oom_handler.try_recover(model.parameters()):
print(f'Skipping {i} / {num_examples}')
continue
else:
raise
print('Alignment time: {:.02f} sec'.format(time.time() -
tic_alignment))
if args.max_segment_duration:
ref_transcript, hyp_transcript = [
list(sorted(sum(segments, []), key=transcripts.sort_key))
for segments in [ref_segments, hyp_segments]
]
if ref:
ref_segments = list(
transcripts.segment(ref_transcript,
args.max_segment_duration))
hyp_segments = list(
transcripts.segment(hyp_transcript, ref_segments))
else:
hyp_segments = list(
transcripts.segment(hyp_transcript,
args.max_segment_duration))
ref_segments = [[] for _ in hyp_segments]
transcript = [
dict(audio_path=audio_path,
ref=ref,
hyp=hyp,
speaker=transcripts.speaker(ref=ref_transcript,
hyp=hyp_transcript),
cer=metrics.cer(hyp=hyp, ref=ref),
words=metrics.align_words(hyp=hyp, ref=ref)[-1]
if args.align_words else [],
alignment=dict(ref=ref_transcript, hyp=hyp_transcript),
**transcripts.summary(hyp_transcript)) for ref_transcript,
hyp_transcript in zip(ref_segments, hyp_segments)
for ref, hyp in [(transcripts.join(ref=ref_transcript),
transcripts.join(hyp=hyp_transcript))]
]
filtered_transcript = list(
transcripts.prune(transcript,
align_boundary_words=args.align_boundary_words,
cer=args.cer,
duration=args.duration,
gap=args.gap,
unk=args.unk,
num_speakers=args.num_speakers))
print('Filtered segments:', len(filtered_transcript), 'out of',
len(transcript))
if args.output_json:
transcript_path = os.path.join(args.output_path,
audio_name + '.json')
print(transcript_path)
with open(transcript_path, 'w') as f:
json.dump(filtered_transcript,
f,
ensure_ascii=False,
sort_keys=True,
indent=2)
if args.output_html:
transcript_path = os.path.join(args.output_path,
audio_name + '.html')
print(transcript_path)
vis.transcript(transcript_path, args.sample_rate, args.mono,
transcript, filtered_transcript)
if args.output_txt:
transcript_path = os.path.join(args.output_path,
audio_name + '.txt')
print(transcript_path)
with open(transcript_path, 'w') as f:
f.write(hyp)
if args.output_csv:
output_lines.append(
csv_sep.join((audio_path, hyp, str(begin), str(end))) + '\n')
print('Done: {:.02f} sec\n'.format(time.time() - tic))
if args.output_csv:
with open(os.path.join(args.output_path, 'transcripts.csv'), 'w') as f:
f.writelines(output_lines)
def ref(input_path, output_path, sample_rate, window_size, device,
max_duration, debug_audio, html, ext):
os.makedirs(output_path, exist_ok=True)
audio_source = ([
(input_path, audio_name) for audio_name in os.listdir(input_path)
] if os.path.isdir(input_path) else [(os.path.dirname(input_path),
os.path.basename(input_path))])
for i, (input_path, audio_name) in enumerate(audio_source):
print(i, '/', len(audio_source), audio_name)
audio_path = os.path.join(input_path, audio_name)
noextname = audio_name[:-len(ext)]
transcript_path = os.path.join(output_path, noextname + '.json')
rttm_path = os.path.join(output_path, noextname + '.rttm')
signal, sample_rate = audio.read_audio(audio_path,
sample_rate=sample_rate,
mono=False,
dtype='float32',
duration=max_duration)
speaker_id_ref, speaker_id_ref_ = select_speaker(
signal.to(device),
silence_absolute_threshold=0.05,
silence_relative_threshold=0.2,
kernel_size_smooth_signal=128,
kernel_size_smooth_speaker=4096,
kernel_size_smooth_silence=4096)
transcript = [
dict(audio_path=audio_path,
begin=float(begin) / sample_rate,
end=(float(begin) + float(duration)) / sample_rate,
speaker=speaker,
speaker_name=transcripts.default_speaker_names[speaker])
for speaker in range(1, len(speaker_id_ref_))
for begin, duration, mask in zip(
*models.rle1d(speaker_id_ref_[speaker])) if mask == 1
]
#transcript = [dict(audio_path = audio_path, begin = float(begin) / sample_rate, end = (float(begin) + float(duration)) / sample_rate, speaker_name = str(int(speaker)), speaker = int(speaker)) for begin, duration, speaker in zip(*models.rle1d(speaker_id_ref.cpu()))]
transcript_without_speaker_missing = [
t for t in transcript
if t['speaker'] != transcripts.speaker_missing
]
transcripts.save(transcript_path, transcript_without_speaker_missing)
print(transcript_path)
transcripts.save(rttm_path, transcript_without_speaker_missing)
print(rttm_path)
if debug_audio:
audio.write_audio(
transcript_path + '.wav',
torch.cat([
signal[..., :speaker_id_ref.shape[-1]],
convert_speaker_id(speaker_id_ref[..., :signal.shape[-1]],
to_bipole=True).unsqueeze(0).cpu() *
0.5, speaker_id_ref_[..., :signal.shape[-1]].cpu() * 0.5
]),
sample_rate,
mono=False)
print(transcript_path + '.wav')
if html:
html_path = os.path.join(output_path, audio_name + '.html')
vis.transcript(html_path,
sample_rate=sample_rate,
mono=True,
transcript=transcript,
duration=max_duration)
audio_source = ([
(args.input_path, audio_name) for audio_name in os.listdir(args.input_path)
] if os.path.isdir(args.input_path) else [(os.path.dirname(args.input_path),
os.path.basename(args.input_path))])
for i, (input_path, audio_name) in enumerate(audio_source):
print(i, '/', len(audio_source), audio_name)
audio_path = os.path.join(input_path, audio_name)
transcript_path = os.path.join(args.output_path, audio_name + '.json')
html_path = os.path.join(args.output_path, audio_name + '.html')
diarization = pipeline(dict(audio=audio_path))
transcript = [
dict(audio_path=audio_path,
begin=turn.start,
end=turn.end,
speaker_name=speaker)
for turn, _, speaker in diarization.itertracks(yield_label=True)
]
speaker_names = [None] + list(set(t['speaker_name'] for t in transcript))
for t in transcript:
t['speaker'] = speaker_names.index(t['speaker_name'])
json.dump(transcript, open(transcript_path, 'w'), indent=2, sort_keys=True)
vis.transcript(html_path,
sample_rate=args.sample_rate,
mono=True,
transcript=transcript_path)