The Revolab ASR Benchmark is a human-annotated evaluation dataset for Malaysian Malay speech recognition. This is the public split - a downloadable subset that anyone can use to run their own evaluation. A larger private split is used to maintain the official leaderboard.

Dataset Details

Audio samples: 820
Duration: ~2.2 hours
Languages: Bahasa Malaysia and English (including code-switching)
Sample rate: 16kHz
License: CC-BY-4.0 (for annotations; underlying audio recordings are subject to their original terms)

Dataset Fields

Each row contains:

audio: Audio waveform (16kHz)
text: Human-verified transcript (verbatim)
normalized_text: Normalized version (numbers/currency spelled out)
category: Domain/category label
canonical: Map of word variants that should score as correct
id: Unique identifier based on audio fingerprint

Data Sources

Most categories were collected, sampled, and manually labelled by Revolab specifically for this benchmark. The dataset incorporates the following sources:

Read speech: Incorporates audio from Raz/tufs_malay, a corpus of Malay speech recordings, extended with additional samples and our own transcription and normalization pipeline
FLEURS: From Google Research (CC-BY-4.0)
Common Voice: From Mozilla Foundation (CC0)

Intended Use

This dataset is intended for:

Evaluation of ASR systems on Malaysian Malay speech
Research on code-switching and accented speech recognition
Benchmarking model performance across diverse speech domains

Limitations

Covers Malaysian Malay and English code-switching only
~2.2 hours is sufficient for evaluation, not training from scratch
Public telephony samples are role-played (not real customer calls)

Citation

If you use this dataset, please cite:

@dataset{revolab_asr_benchmark,
  title = {Revolab ASR Benchmark},
  author = {Revolab Research},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {https://huggingface.co/datasets/Revolab/ASR-Benchmark-Public}
}

License

We do not claim ownership of any original content included in this dataset. All copyrights and other intellectual property rights belong to their respective owners.

This dataset is shared solely to facilitate research and educational purposes. If you are the copyright owner of any content and believe it should not be included, please contact us and we will promptly remove the relevant material.

Contact

For questions or any feedback about this dataset, please contact the Revolab team at support@revolab.ai

Downloads last month: 100

Total file size:

242 MB

Category	Description	Public Samples
Telephony	Compressed, 8kHz customer support and IVR audio	66
Parliament	Formal register with technical vocabulary	52
Podcast	Long-form, unscripted conversational speech	52
Read speech	Clean, controlled recordings	51
Street interview	Outdoor ambient noise with spontaneous speech	48
Short inputs	Single words and short phrases (TTS-generated)	52
Animation	Highly expressive, exaggerated voice acting	50
Singing	Extreme pronunciation and rhythm variation	51
Drama	Emotionally inflected, scripted dialogue	52
News	Professionally delivered speech	52
FLEURS	Google's multilingual benchmark subset	153
Common Voice	Mozilla's crowdsourced corpus subset	141

You need to agree to share your contact information to access this dataset

Revolab ASR Benchmark (Public Split)