Fleurs-mn: Mongolian Speech Recognition Dataset with 15.5 Hours of Audio

4,428 audio samples totaling 15 hours and 33 minutes of Mongolian speech, recombined and split into train and test sets. The dataset, created by Batuka0901, contains 3,985 training samples and 443 test samples. It was last updated on the platform in April 2026.

Use Cases

• Train Mongolian speech recognition models based on the described audio samples.
• Benchmark ASR system performance using the provided 10% test split.
• Develop language-specific acoustic models for Mongolian based on the 15.5-hour audio corpus.
• Fine-tune pre-trained multilingual speech models on Mongolian data.

Strengths

• Contains 4,428 total audio samples, providing a substantial corpus for model training.
• Offers a clear 90/10 train-test split with 3,985 training and 443 test samples.
• Total audio duration is precisely quantified at 15 hours, 33 minutes, and 10 seconds.
• Average audio clip duration is provided for each split, around 12-13 seconds.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Description metadata is limited; actual data quality requires manual inspection after download.
• The dataset's original source, collection methodology, and speaker demographics are not described.

Provenance

Source

Batuka0901 on Hugging Face

Collection Method

Described as recombined and split from an unspecified source.

Freshness

Last updated 2026-04-24 13:46:19; freshness should be verified.

Geography

Likely contains Mongolian language audio.