Add GigaSpeech 2 recipe

[yfyeung]

This PR adds a recipe for GigaSpeech 2.
GigaSpeech 2 raw comprises about 30,000 hours of automatically transcribed speech across Thai, Indonesian, and Vietnamese. GigaSpeech 2 refined consists of 10,000 hours of Thai, 6,000 hours each for Indonesian and Vietnamese. GigaSpeech 2 test sets more realistically reflect speech recognition scenarios and mirror the real performance of an ASR system for low-resource languages.

For more details, please visit:
Dataset: https://huggingface.co/datasets/speechcolab/gigaspeech2
Preprint paper: https://arxiv.org/pdf/2406.11546

[pzelasko]

Thanks!! The recipe looks good to me, although I have one suggestion. If you could re-use the streaming manifest writing mechanism from GigaSpeech 1 recipe, it would allow users to prepare this dataset with minimal memory usage. As-is, it will take a lot of CPU memory to hold the entire manifest in memory before writing it to disk. See:

lhotse/lhotse/recipes/gigaspeech.py

Lines 96 to 129 in da4d70d

	with RecordingSet.open_writer(
	output_dir / f"gigaspeech_recordings_{part}.jsonl.gz"
	) as rec_writer, SupervisionSet.open_writer(
	output_dir / f"gigaspeech_supervisions_{part}.jsonl.gz"
	) as sup_writer, CutSet.open_writer(
	output_dir / f"gigaspeech_cuts_{part}.jsonl.gz"
	) as cut_writer:
	for recording, segments in tqdm(
	parallel_map(
	parse_utterance,
	gigaspeech.audios("{" + part + "}"),
	repeat(gigaspeech.gigaspeech_dataset_dir),
	num_jobs=num_jobs,
	),
	desc="Processing GigaSpeech JSON entries",
	):
	# Fix and validate the recording + supervisions
	recordings, segments = fix_manifests(
	recordings=RecordingSet.from_recordings([recording]),
	supervisions=SupervisionSet.from_segments(segments),
	)
	validate_recordings_and_supervisions(
	recordings=recordings, supervisions=segments
	)
	# Create the cut since most users will need it anyway.
	# There will be exactly one cut since there's exactly one recording.
	cuts = CutSet.from_manifests(
	recordings=recordings, supervisions=segments
	)
	# Write the manifests
	rec_writer.write(recordings[0])
	for s in segments:
	sup_writer.write(s)
	cut_writer.write(cuts[0])

[yfyeung]

Thanks!! The recipe looks good to me, although I have one suggestion. If you could re-use the streaming manifest writing mechanism from GigaSpeech 1 recipe, it would allow users to prepare this dataset with minimal memory usage. As-is, it will take a lot of CPU memory to hold the entire manifest in memory before writing it to disk. See:

lhotse/lhotse/recipes/gigaspeech.py

Lines 96 to 129 in da4d70d

	with RecordingSet.open_writer(
	output_dir / f"gigaspeech_recordings_{part}.jsonl.gz"
	) as rec_writer, SupervisionSet.open_writer(
	output_dir / f"gigaspeech_supervisions_{part}.jsonl.gz"
	) as sup_writer, CutSet.open_writer(
	output_dir / f"gigaspeech_cuts_{part}.jsonl.gz"
	) as cut_writer:
	for recording, segments in tqdm(
	parallel_map(
	parse_utterance,
	gigaspeech.audios("{" + part + "}"),
	repeat(gigaspeech.gigaspeech_dataset_dir),
	num_jobs=num_jobs,
	),
	desc="Processing GigaSpeech JSON entries",
	):
	# Fix and validate the recording + supervisions
	recordings, segments = fix_manifests(
	recordings=RecordingSet.from_recordings([recording]),
	supervisions=SupervisionSet.from_segments(segments),
	)
	validate_recordings_and_supervisions(
	recordings=recordings, supervisions=segments
	)
	# Create the cut since most users will need it anyway.
	# There will be exactly one cut since there's exactly one recording.
	cuts = CutSet.from_manifests(
	recordings=recordings, supervisions=segments
	)
	# Write the manifests
	rec_writer.write(recordings[0])
	for s in segments:
	sup_writer.write(s)
	cut_writer.write(cuts[0])

Sure, I will implement this later.