Use of this dataset for commercial purposes is strictly prohibited.
Please provide your real details when requesting access.
The owner of the dataset reserves the right to reject any request without explanation.

Kazakh Songs ASR Dataset

Dataset Summary

This dataset consists of manually aligned audio–text pairs extracted from Kazakh songs and designed for research in automatic speech recognition (ASR) for low-resource languages. The primary goal of the dataset is to investigate whether sung speech can serve as a complementary training resource for Kazakh ASR systems.

The corpus contains line-level vocal segments obtained from commercially released songs, with manually verified transcriptions and timestamps. It is intended primarily for research on low-resource ASR, domain adaptation, and robustness to non-speech vocal styles.

This dataset was introduced in the paper:

Using Songs to Improve Kazakh Automatic Speech Recognition (LREC 2026)

Supported Tasks

Automatic Speech Recognition (ASR)
Low-resource speech modelling
Domain adaptation (speech vs sung vocals)
Robust ASR evaluation under non-standard acoustic conditions

Languages

Kazakh (kk)

Dataset Structure

Each example in the dataset contains the following fields:

Field	Description
`audio`	Vocal audio segment (speech extracted from songs)
`transcription`	Manually verified reference transcription in Kazakh
`transcription_normalized`	Lowercased and punctuation-normalised transcription (for normalised WER evaluation)
`start_time`	Start timestamp (in seconds) within the original song
`end_time`	End timestamp (in seconds) within the original song
`duration_sec`	Duration of the segment in seconds
`artist_name`	Name of the performing artist
`song_title`	Title of the song
`artist_gender`	Gender of the performing artist (m/f)

Data Collection Process

Source Material

The dataset was constructed from Kazakh songs performed by 36 artists across multiple mainstream genres (e.g., pop, pop-estrada, folk-pop, and R&B).

Selection Criteria

Songs were selected based on:

Prominent solo vocals
Clear lyrical structure
Diversity of artists and styles
Availability of reliable lyrics

Instrumental-dominated tracks and choral arrangements were excluded.

Audio Processing

Songs were obtained from publicly available online sources.
Vocal tracks were separated from instrumental accompaniment using Spleeter.
Residual background music may remain due to imperfect source separation.

Alignment and Segmentation

All audio–text pairs were manually aligned at the line level using Audacity.
Each segment was created by listening to the vocal track and synchronising it with the corrected lyrics. The exported timestamps (start_time, end_time) ensure full reproducibility of the segmentation process.

Transcription Quality

Lyrics were collected from online repositories and official sources, then manually reviewed and corrected to match the actual sung content. This includes:

Colloquial pronunciations
Repetitions
Minor lyrical variations in performance

Transcriptions preserve original casing and punctuation, while a normalised version is provided for evaluation.

Dataset Statistics

Number of artists: 36
Number of songs: 195
Number of segments: 3,013
Total duration: ~4.5 hours
Average segment length: ~5.4 seconds

The dataset includes both female and male artists and spans multiple musical genres to ensure stylistic diversity.

Intended Use

Primary Use

Fine-tuning ASR models for Kazakh
Research on low-resource ASR
Domain adaptation using sung speech
Analysis of ASR robustness to non-standard vocal styles

Secondary Use

Linguistic analysis of sung Kazakh speech
Prosody-aware ASR research
Cross-domain speech modelling

Limitations

The total duration (4.5 hours) is relatively small compared to large-scale ASR corpora.
Residual background music may remain after vocal separation.
Manual alignment may introduce minor subjective inconsistencies.
The dataset focuses on Kazakh and may not generalise to other languages or musical traditions.
Genre coverage, while diverse, is not exhaustive.

This dataset should be considered a proof-of-concept resource rather than a comprehensive ASR corpus.

Ethical and Legal Considerations

The dataset is derived from copyrighted musical works.
To respect intellectual property rights:

Only segmented vocal excerpts are provided for research purposes.
The dataset is intended strictly for non-commercial academic use.
Metadata (artist and song title) are retained for transparency and traceability.

Users are responsible for ensuring compliance with applicable copyright regulations in their jurisdiction.

Citation

If you use this dataset, please cite:

@misc{yeshpanov2026usingsongsimprovekazakh,
      title={Using Songs to Improve Kazakh Automatic Speech Recognition}, 
      author={Rustem Yeshpanov},
      year={2026},
      eprint={2603.00961},
      archivePrefix={arXiv},
      primaryClass={eess.AS},
      url={https://arxiv.org/abs/2603.00961}, 
}

Downloads last month: 2

Paper for yeshpanovrustem/kazakh_songs_asr

Using Songs to Improve Kazakh Automatic Speech Recognition

Paper • 2603.00961 • Published 3 days ago