MERaLiON/sea_audiobench_datasets_TLoc

SEA-SpeechBench — TLoc (Temporal Localization)

This dataset is the temporal-localization (TLoc) task of SEA-SpeechBench, a large-scale multitask benchmark for speech understanding across Southeast Asia. It contains 6,736 evaluation examples across five languages, each pairing an audio recording of 30–180 seconds with an instruction and a reference answer.

Given the recording and the instruction, a model must identify when a described utterance occurs.

Quick start

Requires datasets>=4.0, which decodes audio through torchcodec and needs a system FFmpeg (versions 4–7):

pip install "datasets[audio]"

from datasets import load_dataset

ds = load_dataset(
    "MERaLiON/sea_audiobench_datasets_TLoc",
    "tloc_yodas2_en_30",
    split="train",
)

row = ds[0]
print(row["instruction"])
print(row["answer"])
print(row["language"], row["audio_length"])

# `context` is a torchcodec AudioDecoder, not a dict
samples = row["context"].get_all_samples()
waveform = samples.data          # torch.Tensor, shape (num_channels, num_samples)
sr = samples.sample_rate         # 16000
print(waveform.shape, sr)

If you need a numpy array — most feature extractors take one — squeeze the channel dimension (all TLoc audio is mono):

audio = samples.data.squeeze(0).numpy()

Subsets

All data lives in a single train split and is intended for evaluation only. Config names follow tloc_{source}_{language}_{context_length_seconds}.

Config	Source	Lang	Ctx (s)	Examples	Nominal hrs
tloc_sg_streets_30	SG Streets	en	30	95	0.8
tloc_sgpccsc_long_30	ASR-SgpCCSC	zh	30	512	4.3
tloc_yodas2_en_30	YODAS2	en	30	1,000	8.3
tloc_yodas2_id_30	YODAS2	id	30	1,000	8.3
tloc_yodas2_th_30	YODAS2	th	30	1,000	8.3
tloc_yodas2_vi_30	YODAS2	vi	30	489	4.1
tloc_yodas2_vi_60	YODAS2	vi	60	341	5.7
tloc_yodas2_vi_120	YODAS2	vi	120	163	5.4
tloc_yodas2_vi_180	YODAS2	vi	180	62	3.1
tloc_yodas2_zh_30	YODAS2	zh	30	1,000	8.3
tloc_yodas2_zh_60	YODAS2	zh	60	754	12.6
tloc_yodas2_zh_120	YODAS2	zh	120	271	9.0
tloc_yodas2_zh_180	YODAS2	zh	180	49	2.5
Total				6,736	~80.7

Data fields

Field	Type	Description
context	Audio(sampling_rate=16000)	The audio recording.
instruction	string	The instruction given to the model.
answer	string	The reference answer.
audio_length	float64	Duration of context, in seconds.
language	string	Language code of the subset.

See the paper for the instruction and answer formats, how segments and annotations were constructed, and the evaluation protocol.

Source data

TLoc introduces no new recordings. All audio comes from existing corpora; the contribution is the segmentation, instructions, and reference answers.

Source	Lang	Upstream license
YODAS2	en, id, zh, th, vi	CC BY 3.0
SG Streets	en	Not specified
ASR-SgpCCSC	zh	CC BY-NC-ND 4.0

Citation

If you use TLoc, please cite our benchmark and the source corpora.

This benchmark

@inproceedings{liao2026seaspeechbench,
  title     = {{SEA-SpeechBench}: A Large-Scale Multitask Benchmark for
               Speech Understanding Across Southeast Asia},
  author    = {Liao, Jingyi and Zhang, Wenyu and Liu, Zhuohan and
               He, Yingxu and Lin, Geyu and Zou, Xunlong and Sun, Shuo and
               Alsagoff, Syed Ali Redha and Aw, Ai Ti},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in
               Natural Language Processing (EMNLP)},
  year      = {2026}
}

Source corpora

@inproceedings{li2023yodas,
  title        = {YODAS: Youtube-Oriented Dataset for Audio and Speech},
  author       = {Li, Xinjian and Takamichi, Shinnosuke and Saeki, Takaaki
                  and Chen, William and Shiota, Sayaka and Watanabe, Shinji},
  booktitle    = {2023 IEEE Automatic Speech Recognition and Understanding
                  Workshop (ASRU)},
  pages        = {1--8},
  year         = {2023},
  organization = {IEEE}
}

@misc{sgpccsc,
  title  = {{ASR-SgpCCSC: A Singapore Chinese Conversational Speech Corpus}},
  author = {Magic Data Tech},
  year   = {2024},
  url    = {https://magichub.com/datasets/[SLUG]/}
}

@inproceedings{khassanov2019enriching,
  title     = {Enriching Rare Word Representations in Neural Language Models
               by Embedding Matrix Augmentation},
  author    = {Khassanov, Yerbolat and Zeng, Zhiping and Pham, Van Tung and
               Xu, Haihua and Chng, Eng Siong},
  booktitle = {Interspeech 2019},
  pages     = {3505--3509},
  year      = {2019},
  month     = sep,
  publisher = {ISCA},
  doi       = {10.21437/Interspeech.2019-1858},
  url       = {http://dx.doi.org/10.21437/Interspeech.2019-1858}
}