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 __init__(
self,
data_paths,
text_pipelines: typing.List[
language_processing.ProcessingPipeline],
sample_rate,
frontend=None,
speaker_names=None,
waveform_transform_debug_dir=None,
min_duration=None,
max_duration=None,
duration_filter=True,
min_ref_len=None,
max_ref_len=None,
max_num_channels=2,
ref_len_filter=True,
mono=True,
audio_dtype='float32',
segmented=False,
time_padding_multiple=1,
audio_backend=None,
exclude=set(),
join_transcript=False,
bucket=None,
pop_meta=False,
string_array_encoding='utf_16_le',
_print=print,
debug_short_long_records_features_from_whole_normalized_signal=False
):
self.debug_short_long_records_features_from_whole_normalized_signal = debug_short_long_records_features_from_whole_normalized_signal
self.join_transcript = join_transcript
self.max_duration = max_duration
self.text_pipelines = text_pipelines
self.frontend = frontend
self.sample_rate = sample_rate
self.waveform_transform_debug_dir = waveform_transform_debug_dir
self.segmented = segmented
self.time_padding_multiple = time_padding_multiple
self.mono = mono
self.audio_backend = audio_backend
self.audio_dtype = audio_dtype
data_paths = data_paths if isinstance(data_paths,
list) else [data_paths]
exclude = set(exclude)
tic = time.time()
transcripts_read = list(map(transcripts.load, data_paths))
_print('Dataset reading time: ',
time.time() - tic)
tic = time.time()
#TODO group only segmented = True
segments_by_audio_path = []
for transcript in transcripts_read:
transcript = sorted(transcript, key=transcripts.sort_key)
transcript = itertools.groupby(transcript,
key=transcripts.group_key)
for _, example in transcript:
segments_by_audio_path.append(list(example))
speaker_names_filtered = set()
examples_filtered = []
examples_lens = []
transcript = []
duration = lambda example: sum(
map(transcripts.compute_duration, example))
segments_by_audio_path.sort(key=duration)
# TODO: not segmented mode may fail if several examples have same audio_path
for example in segments_by_audio_path:
exclude_ok = ((not exclude) or
(transcripts.audio_name(example[0]) not in exclude))
duration_ok = (
(not duration_filter) or
(min_duration is None or min_duration <= duration(example)) and
(max_duration is None or duration(example) <= max_duration))
if duration_ok and exclude_ok:
b = bucket(example) if bucket is not None else 0
for t in example:
t['bucket'] = b
t['ref'] = t.get('ref', transcripts.ref_missing)
t['begin'] = t.get('begin', transcripts.time_missing)
t['end'] = t.get('end', transcripts.time_missing)
t['channel'] = t.get('channel',
transcripts.channel_missing)
examples_filtered.append(example)
transcript.extend(example)
examples_lens.append(len(example))
self.speaker_names = transcripts.collect_speaker_names(
transcript,
speaker_names=speaker_names or [],
num_speakers=max_num_channels,
set_speaker=True)
_print('Dataset construction time: ',
time.time() - tic)
tic = time.time()
self.bucket = torch.ShortTensor(
[e[0]['bucket'] for e in examples_filtered])
self.audio_path = utils.TensorBackedStringArray(
[e[0]['audio_path'] for e in examples_filtered],
encoding=string_array_encoding)
self.ref = utils.TensorBackedStringArray(
[t['ref'] for t in transcript], encoding=string_array_encoding)
self.begin = torch.DoubleTensor([t['begin'] for t in transcript])
self.end = torch.DoubleTensor([t['end'] for t in transcript])
self.channel = torch.CharTensor([t['channel'] for t in transcript])
self.speaker = torch.LongTensor([t['speaker'] for t in transcript])
self.cumlen = torch.ShortTensor(examples_lens).cumsum(
dim=0, dtype=torch.int64)
if pop_meta:
self.meta = {}
else:
self.meta = {self.example_id(t): t for t in transcript}
if self.join_transcript:
#TODO: harmonize dummy transcript of replace_transcript case (and fix channel)
self.meta.update({
self.example_id(t_src): t_tgt
for e in examples_filtered for t_src, t_tgt in [(
dict(audio_path=e[0]['audio_path'],
begin=transcripts.time_missing,
end=transcripts.time_missing,
channel=transcripts.channel_missing,
speaker=transcripts.speaker_missing),
dict(audio_path=e[0]['audio_path'],
begin=0.0,
end=audio.compute_duration(e[0]['audio_path'],
backend=None),
channel=transcripts.channel_missing,
speaker=transcripts.speaker_missing,
ref=' '.join(
filter(bool, [t.get('ref', '')
for t in e]))))]
})
_print('Dataset tensors creation time: ', time.time() - tic)
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 __init__(
self,
data_paths: typing.List[str],
text_pipelines: typing.List[text_processing.ProcessingPipeline],
sample_rate: int,
mode: str = DEFAULT_MODE,
frontend: typing.Optional[torch.nn.Module] = None,
speaker_names: typing.Optional[typing.List[str]] = None,
max_audio_file_size: typing.Optional[float] = None, #bytes
min_duration: typing.Optional[float] = None,
max_duration: typing.Optional[float] = None,
max_num_channels: int = 2,
mono: bool = True,
audio_dtype: str = 'float32',
time_padding_multiple: int = 1,
audio_backend: typing.Optional[str] = None,
exclude: typing.Optional[typing.Set] = None,
bucket_fn: typing.Callable[[typing.List[typing.Dict]],
int] = lambda transcript: 0,
pop_meta: bool = False,
string_array_encoding: str = 'utf_16_le',
_print: typing.Callable = print,
debug_short_long_records_features_from_whole_normalized_signal:
bool = False,
duration_from_transcripts: bool = False,
):
self.debug_short_long_records_features_from_whole_normalized_signal = debug_short_long_records_features_from_whole_normalized_signal
self.mode = mode
self.min_duration = min_duration
self.max_duration = max_duration
self.max_audio_file_size = max_audio_file_size
self.text_pipelines = text_pipelines
self.frontend = frontend
self.sample_rate = sample_rate
self.time_padding_multiple = time_padding_multiple
self.mono = mono
self.audio_backend = audio_backend
self.audio_dtype = audio_dtype
data_paths = data_paths if isinstance(data_paths,
list) else [data_paths]
data_paths_ = []
for data_path in data_paths:
if os.path.isdir(data_path):
data_paths_.extend(
os.path.join(data_path, filename) for filename in filter(
audio.is_audio, os.listdir(data_path)))
else:
data_paths_.append(data_path)
data_paths = data_paths_
tic = time.time()
segments = []
for path in data_paths:
if audio.is_audio(path):
assert self.mono or self.mode != AudioTextDataset.DEFAULT_MODE, 'Only mono audio files allowed as dataset input in default mode'
if self.mono:
transcript = [
dict(audio_path=path,
channel=transcripts.channel_missing)
]
else:
transcript = [
dict(audio_path=path, channel=c)
for c in range(max_num_channels)
]
else:
transcript = transcripts.load(path)
segments.extend(transcript)
_print('Dataset reading time: ', time.time() - tic)
tic = time.time()
# get_or_else required because dictionary could contain None values which we want to replace.
# dict.get doesnt work in this case
get_or_else = lambda dictionary, key, default: dictionary[
key] if dictionary.get(key) is not None else default
for t in segments:
t['ref'] = get_or_else(t, 'ref', transcripts.ref_missing)
t['begin'] = get_or_else(t, 'begin', transcripts.time_missing)
t['end'] = get_or_else(t, 'end', transcripts.time_missing)
t['channel'] = get_or_else(
t, 'channel', transcripts.channel_missing
) if not self.mono else transcripts.channel_missing
transcripts.collect_speaker_names(segments,
speaker_names=speaker_names or [],
num_speakers=max_num_channels,
set_speaker_data=True)
buckets = []
grouped_segments = []
transcripts_len = []
speakers_len = []
if self.mode == AudioTextDataset.DEFAULT_MODE:
groupped_transcripts = ((i, [t]) for i, t in enumerate(segments))
else:
groupped_transcripts = itertools.groupby(
sorted(segments, key=transcripts.group_key),
transcripts.group_key)
for group_key, transcript in groupped_transcripts:
transcript = sorted(transcript, key=transcripts.sort_key)
if self.mode == AudioTextDataset.BATCHED_CHANNELS_MODE:
transcript = transcripts.join_transcript(
transcript,
self.mono,
duration_from_transcripts=duration_from_transcripts)
if exclude is not None:
allowed_audio_names = set(
transcripts.audio_name(t) for t in transcript
if transcripts.audio_name(t) not in exclude)
else:
allowed_audio_names = None
transcript = transcripts.prune(
transcript,
allowed_audio_names=allowed_audio_names,
duration=(
min_duration if min_duration is not None else 0.0,
max_duration if max_duration is not None else 24.0 * 3600,
), #24h
max_audio_file_size=max_audio_file_size)
transcript = list(transcript)
for t in transcript:
t['example_id'] = AudioTextDataset.get_example_id(t)
if len(transcript) == 0:
continue
bucket = bucket_fn(transcript)
for t in transcript:
t['bucket'] = bucket
speakers_len.append(
len(t['speaker']) if (
isinstance(t['speaker'], list)) else 1)
buckets.append(bucket)
grouped_segments.extend(transcript)
transcripts_len.append(len(transcript))
_print('Dataset construction time: ', time.time() - tic)
tic = time.time()
self.bucket = torch.tensor(buckets, dtype=torch.short)
self.audio_path = utils.TensorBackedStringArray(
[t['audio_path'] for t in grouped_segments],
encoding=string_array_encoding)
self.ref = utils.TensorBackedStringArray(
[t['ref'] for t in grouped_segments],
encoding=string_array_encoding)
self.begin = torch.tensor([t['begin'] for t in grouped_segments],
dtype=torch.float64)
self.end = torch.tensor([t['end'] for t in grouped_segments],
dtype=torch.float64)
self.channel = torch.tensor([t['channel'] for t in grouped_segments],
dtype=torch.int8)
self.example_id = utils.TensorBackedStringArray(
[t['example_id'] for t in grouped_segments],
encoding=string_array_encoding)
if self.mode == AudioTextDataset.BATCHED_CHANNELS_MODE:
self.speaker = torch.tensor([
speaker for t in grouped_segments for speaker in t['speaker']
],
dtype=torch.int64)
else:
self.speaker = torch.tensor(
[t['speaker'] for t in grouped_segments], dtype=torch.int64)
self.speaker_len = torch.tensor(speakers_len, dtype=torch.int16)
self.transcript_cumlen = torch.tensor(transcripts_len,
dtype=torch.int16).cumsum(
dim=0, dtype=torch.int64)
if pop_meta:
self.meta = {}
else:
self.meta = {t['example_id']: t for t in grouped_segments}
_print('Dataset tensors creation time: ', time.time() - tic)