Implement long audio inference for k2

[JER-ry]

ReazonSpeech/pkg/k2-asr/src/transcribe.py

Lines 28 to 34 in b6916fd

	if duration > TOO_LONG_SECONDS:
	warnings.warn(
	f"Passing a long audio input ({duration:.1f}s) is not recommended, "
	"because K2 will require a large amount of memory. "
	"Read the upstream discussion for more details: "
	"https://github.com/k2-fsa/icefall/issues/1680"
	)

Since this is implemented for espnet, it's probably worth also copying that here.

ReazonSpeech/pkg/espnet-asr/src/transcribe.py

Lines 59 to 77 in b6916fd

	while pos < len(audio.waveform):
	samples = audio.waveform[pos:]
	# If the audio data is very long, find out the longest
	# non-speech region and perform decoding up to that point.
	if len(samples) > window:
	blank = find_blank(model, samples[:window])
	mid = int((blank.start + blank.end) / 2)
	samples = samples[:mid]
	asr = model(np.pad(samples, PADDING, mode="constant"))[0][0]
	fulltext += asr
	for start, end, text in split_text(model, samples, asr):
	segments.append(Segment(
	start_seconds=((pos + start) / audio.samplerate),
	end_seconds=((pos + end) / audio.samplerate),
	text=text,
	))

[fujimotos]

Unfortunately, it's not that easy.

This particular function find_blank() deeply depends on ESPnet's API interface.
For this reason, we cannot just copy it to reazonspeech.k2.asr.

Since this is implemented for espnet, it's probably worth also copying that here

For now, we recommend to use an external VAD model (such as Silero VAD)
to process a long audio sequence with our K2 model.