EN-MALAY-CS-FILTERED

English–Malay code-switching conversational speech from the IMDA National Speech Corpus (2021), segmented to utterance level and cleaned with a 3-model agreement filter: each utterance was transcribed by three ASR models — openai/whisper-large-v3, MERaLiON/MERaLiON-2-10B-ASR, and Qwen/Qwen3-ASR-1.7B — and an utterance is removed when all three models score WER > 60% against the reference transcript (all models agreeing the reference is unreliable). Utterances where at least one model matches the reference reasonably are kept.

Split	Kept / Original	Removed	Speakers	Duration
train	6843 / 10743	3900 (36.3%)	45	11 h 20 m
validation	1722 / 2093	371 (17.7%)	58	2 h 03 m
test	1383 / 1985	602 (30.3%)	66	1 h 25 m

Audio: 16 kHz, mono, 16-bit PCM WAV, one file per utterance.

Usage

from datasets import load_dataset

ds = load_dataset("yyhenggg/EN-MALAY-CS-FILTERED")
# ds["train"], ds["validation"], ds["test"]

Columns

• audio — the utterance waveform
• utt_id — imda-2021-<session>-<channel>-<start>-<end> (times in centiseconds)
• speaker_id — recording session/channel identifier
• duration — seconds
• transcription — reference transcript (verbatim, lowercased, unpunctuated)
• whisper_hyp, whisper_wer — whisper-large-v3 zero-shot hypothesis and its WER vs. the reference
• meralion_hyp, meralion_wer — MERaLiON-2-10B-ASR hypothesis and WER
• qwen3_hyp, qwen3_wer — Qwen3-ASR-1.7B hypothesis and WER

Kaldi recipe

kaldi/train, kaldi/valid, kaldi/test contain standard Kaldi data-dir files (text, wav.scp, utt2spk, spk2utt, utt2dur). wav.scp paths are relative to the repository root — after downloading, rewrite them to absolute paths, e.g.:

awk -v root="$PWD" '{print $1, root"/"$2}' kaldi/train/wav.scp > kaldi/train/wav_abs.scp

Provenance and license

Source audio and references derive from the IMDA National Speech Corpus (English–Malay code-switching, 2021). Use is subject to the IMDA NSC terms / Singapore Open Data Licence.