NICT-Tib1 Tibetan ASR Corpus

This is not a new corpus. It is the NICT-Tib1 speech corpus (Soky, Gong & Li, 2022 — see citation below) repackaged into Hugging Face datasets format, with the speaker-disjoint train/test split used in the paper below. All audio and transcriptions originate from the original NICT-Tib1 release; no new speech data was recorded for this repo.

A speech-transcription corpus of modern Lhasa (Central) Tibetan news speech, hosted here as a ready-to-use Hugging Face dataset, reformatted and split for a benchmark of Tibetan ASR architectures and evaluation metrics:

J. Moore, S. Li and P. Lauren, "Evaluating Tibetan ASR With Segmented Word Error Rate: Beyond Character-Level Metrics," in IEEE Access, vol. 14, pp. 101790-101805, 2026, doi: 10.1109/ACCESS.2026.3709206.

Dataset Description

NICT-Tib1 contains transcribed speech from 20 unique speakers of Lhasa dialect Tibetan, recorded reading scripted news-domain sentences aloud. Each example pairs an audio recording with its Tibetan-script transcription and a speaker identifier.

Dataset Structure

Column	Type	Description
audio	audio	Speech recording
transcription	string	Reference Tibetan-script transcription
speaker_id	string	Speaker identifier

Splits:

Split	Utterances	Speakers
train	15,099	17
test	1,547	3

The split is speaker-disjoint: speaker IDs were partitioned 85/15 (fixed random seed) and every utterance from a given speaker was assigned entirely to one split, so no speaker appears in both training and test. This prevents models from being evaluated on voices seen during training, at the cost of a small absolute number of test speakers given the corpus's limited speaker pool.

Intended Uses

• Training and fine-tuning Tibetan ASR models
• Benchmarking ASR architectures and adaptation strategies (full fine-tuning, quantization, LoRA/QLoRA) on a low-resource language
• Evaluating word-, syllable-, and character-level ASR metrics for a language without explicit word boundaries in its orthography

Limitations

• Covers only modern Lhasa (Central) Tibetan; performance/behavior on Amdo, Kham, or Classical Tibetan is not represented.
• Read-speech news domain only; does not represent spontaneous or conversational speech.
• The speaker-disjoint test split contains only 3 speakers, limiting how confidently results generalize across speaker, age, and prosodic variation within Lhasa Tibetan itself.

Related Resources

• Model checkpoints fine-tuned on this dataset: billingsmoore/tibetan-asr-* (14 models across 5 architectures and multiple adaptation strategies)
• tibetan-wer (PyPI): open-source package implementing CER, SER, and Segmented WER for Tibetan ASR evaluation — https://pypi.org/project/tibetan-wer/
• Word segmentation annotations: billingsmoore/tibetan-word-segmentation-ds

Citation

If you use this dataset, please cite both the paper and the source NICT-Tib1 corpus:

@ARTICLE{11592371,
  author={Moore, Jacob and Li, Sheng and Lauren, Paula},
  journal={IEEE Access},
  title={Evaluating Tibetan ASR With Segmented Word Error Rate: Beyond Character-Level Metrics},
  year={2026},
  volume={14},
  number={},
  pages={101790-101805},
  keywords={Modeling;Automatic speech recognition;Error analysis;LoRa;Measurement;Ranking (statistics);Quantization (signal);Bit error rate;Standards;Training;Tibetan;automatic speech recognition;word error rate;low-resource language},
  doi={10.1109/ACCESS.2026.3709206}
}

@inproceedings{soky2022nict,
  title={Nict-tib1: A public speech corpus of lhasa dialect for benchmarking tibetan language speech recognition systems},
  author={Soky, Kak and Gong, Zhuo and Li, Sheng},
  booktitle={2022 25th Conference of the Oriental COCOSDA International Committee for the Co-ordination and Standardisation of Speech Databases and Assessment Techniques (O-COCOSDA)},
  pages={1--5},
  year={2022},
  organization={IEEE}
}

License

The original NICT-Tib1 corpus is released by Japan's National Institute of Information and Communications Technology (NICT) under the Creative Commons Attribution 4.0 International License (CC BY 4.0). This repo redistributes the same audio and transcriptions, reformatted and re-split, under the same license.