CS-LID-Dataset

Training mix for a 4-language code-switch LID head (languages: English, Mandarin Chinese, Bahasa Indonesia, Bahasa Malay; code-switch pairs: zh-en, id-en, ms-en). Mono corpora are speaker-capped ~20 h subsets sampled with seed 0 (per-speaker duration caps chosen by binary search to maximize speaker diversity; valid splits hold out whole speakers).

Config	Language	Train	Valid	Test/Dev	Source
swb	English	19.0 h / 17,527 utts / 4,853 spk-sides	1.0 h	—	Switchboard-1 (LDC)
wenet	Chinese	19.1 h / 37,648 utts / 35,293 videos	1.0 h	—	WenetSpeech subset-S
datatang	Indonesian	19.0 h / 14,241 utts / 128 spk	0.9 h	—	Datatang mobile-phone (NTU)
malay	Malay	11.75 h / 3,410 utts	0.93 h	—	FLEURS ms_my (CommonVoice ms no longer distributable via HF)
seame	zh-en CS	96.3 h / 89,339 utts	5.1 h	devman 7.5 h, devsge 3.9 h	SEAME (LDC/NTU)
en_indo	id-en CS	15.8 h / 19,080 utts	2.1 h	test 1.6 h	NTU
en_malay	ms-en CS	11.3 h / 6,836 utts	2.1 h	test 1.4 h	NTU (IMDA), 3-model-agreement cleaned

All audio 16 kHz mono (wav; SEAME is flac; FLEURS files are float32 wav). Utterances outside 0.3–30 s were excluded.

Columns

• audio — utterance waveform
• utt_id, speaker_id, duration
• language — en/zh/id/ms for mono corpora, zh-en/id-en/ms-en for CS corpora
• cs — 1 = code-switched: fixed 0 for mono corpora, fixed 1 for en_indo/en_malay (corpus-level; a minority of utterances are actually monolingual), script-derived (CJK vs ASCII) per-utterance for seame
• transcription — reference transcript

Usage

from datasets import load_dataset
seame = load_dataset("yyhenggg/CS-LID-Dataset", "seame")
swb   = load_dataset("yyhenggg/CS-LID-Dataset", "swb", split="train")

Licensing — DO NOT MAKE PUBLIC

This mix contains LDC-licensed corpora (Switchboard-1, SEAME), IMDA NSC derived material, Datatang commercial data, and FLEURS (CC-BY). It is for internal research use; the repository must remain private.

Build manifests

kaldi/<corpus>/<split>/ holds the Kaldi-style manifests used to build this dataset (text, wav.scp, utt2spk, utt2dur); wav.scp paths are local to the build machine and document the exact speaker-capped sample.