Datasetpublic

pengyizhou/IALP-2026-data

IALP-2026: Whisper Open-Set Data-Selection — Query / Dev / Test Sets Supporting data for the study "Whisper-Based Open-Set Data Selection for NSC Adaptation." This repository holds the fixed target-query, validation, and evaluation sets used across all experiments. Each part is a self-contained .tar.gz. All audio is 16 kHz mono. Each split ships with: audio/ — audio files (FLAC, except GigaSpeech which is WAV PCM_16) wav.scp — <utt_id> audio/<file> (Kaldi-style, relative paths)… See the full description on the dataset page: https://huggingface.co/datasets/pengyizhou/IALP-2026-data.

sourceHugging Faceotherupdated 3mo agoView on Hugging Face

0likes20downloads

Dataset Card

IALP-2026: Whisper Open-Set Data-Selection — Query / Dev / Test Sets

Supporting data for the study "Whisper-Based Open-Set Data Selection for NSC Adaptation." This repository holds the fixed target-query, validation, and evaluation sets used across all experiments. Each part is a self-contained .tar.gz.

All audio is 16 kHz mono. Each split ships with:

—audio/ — audio files (FLAC, except GigaSpeech which is WAV PCM_16)
—wav.scp — <utt_id> audio/<file> (Kaldi-style, relative paths)
—text — <utt_id> <transcript>
—utt2spk — <utt_id> <speaker_id>
—manifest.jsonl — one JSON per utterance: {id, speaker, duration, text, audio}

Parts

Archive	Split	Source	#utts	Hours	Role
`nsc-query.tar.gz`	`nsc_query_5h`	NSC IMDA Part 6 (`train`)	~2000	~5.0	Target-domain query set for data selection / EXP0
`nsc-dev.tar.gz`	`nsc_dev_3h`	NSC IMDA Part 6 (`train`)	~1200	~3.0	Validation / checkpoint selection
`nsc-test.tar.gz`	`nsc_test`	NSC IMDA Part 6 (`test`)	3684	~7	In-domain final evaluation
`gigaspeech-test.tar.gz`	`gigaspeech_test`	GigaSpeech `test`	19930	35.4	OOD evaluation (normalized refs)
`librispeech-test.tar.gz`	`librispeech_test_clean` + `librispeech_test_other`	LibriSpeech	2620 + 2939	5.4 + 5.3	OOD evaluation

Split construction (NSC)

—Speaker is the 4th hyphen field of the utterance id (imda-2021-part6-`00006`-channel001m-...).
—nsc_query_5h, nsc_dev_3h, and nsc_test are mutually speaker-disjoint. Query/dev speakers are sampled (seed=42) from shards spread across the corpus and exclude every speaker appearing in the official test partition.
—Query ≈ 111 speakers; dev ≈ 64 speakers.

Notes

—GigaSpeech text is the standard normalized reference (punctuation tags removed, garbage-tagged segments filtered); it is the icefall/Kaldi test set.
—Licensing follows the upstream corpora: NSC/IMDA (IMDA terms), GigaSpeech (speechcolab/gigaspeech agreement), LibriSpeech (CC BY 4.0). Use is restricted to the respective original licenses.

Usage

bash

tar xzf nsc-query.tar.gz   # -> nsc_query_5h/{audio,wav.scp,text,utt2spk,manifest.jsonl}