Incorrect validation in validate_for_asr()

[csukuangfj]

See

lhotse/lhotse/dataset/speech_recognition.py

Lines 206 to 215 in aa38c0f

	tol = 2e-3 # 1ms
	for cut in cuts:
	for supervision in cut.supervisions:
	assert supervision.start >= -tol, (
	f"Supervisions starting before the cut are not supported for ASR"
	f" (sup id: {supervision.id}, cut id: {cut.id})"
	)
	# Supervision start time is relative to Cut ...
	# https://lhotse.readthedocs.io/en/v0.10_e/cuts.html

The code assumes supervision.start is positive.

However,

lhotse/lhotse/dataset/speech_recognition.py

Line 214 in aa38c0f

# Supervision start time is relative to Cut ...

says supervision.start is relative to cut, which means it can be negative.

See also

lhotse/lhotse/cut/base.py

Line 77 in aa38c0f

In some cases, the supervision might have a negative start, or a duration exceeding the duration of the cut;

---

I find this since I have the following cut that fails to pass the validation.

# Edited to remove some info

MonoCut(id='xxxx', start=555.4219954648526, duration=24.081995464852607, channel=0, 
supervisions=[SupervisionSegment(id='xxxx', recording_id='xxx', start=-8.854, 
duration=8.954000000000065, channel=0, text='xxxx', language='zh', 
speaker=None, gender=None, alignment=None)], 
features=Features(type='kaldi-fbank', num_frames=2408, 
num_features=80, frame_shift=0.01, sampling_rate=16000, 
start=555.4219954648526, duration=24.082, 
storage_type='lilcom_chunky',
 storage_path='/xxxx/feats-0.lca', storage_key='501861,44277,43037,43447,43043,3
5284', recording_id='None', channels=0), 
recording=Recording(id='xxxx, 
sources=[AudioSource(type='file', channels=[0, 1], source='xxxx.m4a')], 
sampling_rate=16000, num_samples=19222768, 
duration=1201.423, channel_ids=[0, 1], 
transforms=[Resample(source_sampling_rate=44100, 
target_sampling_rate=16000)]), custom=None)

You can see that supervisions[0].start is -8.854, which is negative.

[pzelasko]

It depends on whether this kind of supervision is acceptable for your experimental setup. In most cases if you have a supervision that began 9s before the cut, a huge part of the transcript in the label would be missing in the audio signal. Perhaps this validation could be a warning rather than error, if that's OK with you can you make a PR?