Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This corpus contains narrations of third-party Kazakh works and is provided for research only, with no blanket redistribution license. By requesting access you agree to use it for non-commercial Kazakh speech-technology research and not to redistribute the audio.
Log in or Sign Up to review the conditions and access this dataset content.
Til Audio Corpus — Kazakh 🇰🇿
Тіл аудиокорпусы (қазақша)
A 1043-hour Kazakh speech corpus of narrated audio paired with
transcripts: literature audiobooks, the Shalqar collection, and minbar
sermons. Each row = one chapter/segment (mp3 + full Kazakh transcript).
Published by TilQazyna as part of the Til foundation-model program for Kazakh.
⚠️ Gated, manual approval. Source-available, not openly licensed — see Provenance & licensing below.
📊 Overview / Шолу
| Total audio | 1042.8 hours (1042h 47m) |
| Clips (audio+text pairs) | 3,303 |
| Distinct works | 435 |
| Avg clip length | 1137 s (~18.9 min) |
| Transcript size | 6,800,866 words · 48,461,032 chars |
| Language | Kazakh (kk, Cyrillic) |
| Audio | MP3 |
| On-disk size | ~64 GB |
By source / Дереккөз бойынша
| source | clips | hours | description |
|---|---|---|---|
kitap |
1,896 | 439.2 | Көркем әдебиет аудиокітаптары — Kazakh-literature audiobooks (Abai Joly / M. Auezov, dastans, folk tales, poetry, …) |
minber |
295 | 97.9 | «Мінбер» уағыздары — minbar sermons / lectures |
shalqar |
1,112 | 505.7 | «Шалқар» жинағы — Shalqar collection (kü/recitation, named reciters) |
Largest works / Ең үлкен туындылар
| work | source | clips | hours |
|---|---|---|---|
| ШАЛҚАР ШАҢЫРАҒЫНДА | shalqar |
413 | 199.4 |
| Бүгінгі сөз | shalqar |
229 | 109.1 |
| (root) | minber |
295 | 97.9 |
| Күй шашу | shalqar |
61 | 49.9 |
| Нысана | shalqar |
78 | 46.8 |
| Шаһнама | kitap |
48 | 45.3 |
| «Жақсының жары» | shalqar |
130 | 43.8 |
| Сөз киесі | shalqar |
135 | 41.9 |
| Ұмытылған ұлылық Орталық Азияның араб шапқыншылығынан Әмір Темірге дейінгі алтын ғасыры | kitap |
42 | 25.7 |
| Өмір мектебі ІІІ | kitap |
21 | 19.5 |
| Абай жолы. II том | kitap |
39 | 18.8 |
| Мөлдір махаббат | kitap |
22 | 18.7 |
| Ботагөз | kitap |
12 | 17.4 |
| Абай жолы. III том | kitap |
34 | 16.8 |
| Жолдастар | kitap |
134 | 15.1 |
🗂 Format / Құрылым
Hugging Face audiofolder layout:
metadata.jsonl # one JSON row per clip
data/<source>/<work>/<chapter>.mp3 # audio
metadata.jsonl fields:
| field | type | description |
|---|---|---|
file_name |
string | relative path to the mp3 (data/...) |
transcription |
string | full Kazakh transcript of the clip |
source |
string | kitap · shalqar · minber |
book |
string | work / book / story title |
chapter |
string | chapter / segment title |
duration_s |
float | audio duration in seconds (ffprobe) |
🚀 Usage / Қолдану
from datasets import load_dataset
# request access first (gated); then authenticate with `hf auth login`
ds = load_dataset("TilQazyna/Til-Audio-Corpus-KK-v1", split="train")
print(ds[0]["transcription"])
print(ds[0]["duration_s"], "s", "/", ds[0]["source"])
ds[0]["audio"] # decoded waveform (needs `torchcodec` with datasets>=4)
Transcript-only (no audio backend needed):
ds = load_dataset("TilQazyna/Til-Audio-Corpus-KK-v1", split="train").remove_columns("audio")
🎯 Intended use
Research on Kazakh ASR / TTS / speech representation: long-form read-speech acoustic modelling, audiobook alignment, TTS voice data, pre-training for speech encoders. Pairs with the rest of the Til program.
⚠️ Limitations
- Long-form clips (chapter-level, avg ~19 min) — not utterance-segmented; segment/force-align before frame-level ASR training.
- Transcripts are chapter-level text, no word/phoneme timestamps.
minberfilenames are machine-scraped (noisy stems); content/pairing is intact.- Single narrator per work in most cases → limited speaker diversity within a work.
- Light normalization only; expect occasional ASR-style transcript noise.
📜 Provenance & licensing
Source-available, gated, manual-approval. Recordings are narrations of third-party Kazakh works (e.g. M. Auezov's Abai Joly, folk tales/dastans, the Shalqar collection, minbar sermons), aggregated for Kazakh speech-technology research. No blanket redistribution license is granted. Access is gated and approved manually; intended for non-commercial research. Rights holders may request removal of any clip via a repository discussion.
📚 Citation
@misc{tilqazyna_audio_corpus_kk_2026,
title = {Til Audio Corpus (Kazakh)},
author = {TilQazyna},
year = {2026},
url = {https://huggingface.co/datasets/TilQazyna/Til-Audio-Corpus-KK-v1}
}
- Downloads last month
- 15