Kazakh Songs ASR: 1,000-10,000 Manually Aligned Audio-Text Pairs

Aggregating between 1,000 and 10,000 manually aligned audio-text pairs from Kazakh commercial songs, released by yeshpanovrustem in 2026. It provides line-level vocal segments designed to investigate the utility of sung speech for low-resource automatic speech recognition (ASR) systems.

Use Cases

• Training ASR models using line-level vocal segments to map melodic audio to text transcriptions
• Benchmarking the robustness of Kazakh speech-to-text systems against rhythmic and melodic vocal inputs
• Researching transfer learning capabilities from sung speech to spoken language models in low-resource contexts

Strengths

• Manually verified line-level alignments between audio and text
• Contains between 1,000 and 10,000 records for a low-resource language
• Sourced from high-quality commercial song recordings

Limitations

• Small sample size of under 10,000 records compared to standard ASR corpora
• Acoustic bias toward sung speech and musical accompaniment which may not generalize to spoken Kazakh
• Lack of demographic diversity data for the vocalists

Provenance

Source

yeshpanovrustem (Arxiv: 2603.00961)

Collection Method

Manual alignment of audio-text pairs from commercially released Kazakh songs

Freshness

Last updated March 2026.

Geography

Kazakhstan