stukenov/sozkz-asr-mels-kk-v1

SozKZ ASR Mel Spectrograms (Kazakh) v1

Pre-computed log-mel spectrograms for Kazakh automatic speech recognition, ready for training encoder-decoder ASR models.

Dataset Details

	Train	Validation	Test	Total
Samples	967,731	51,475	10,295	1,029,501

Sources

Dataset	Samples	Description
KSC2	640,816	Kazakh Speech Corpus 2 (~1,200h, crowdsourced + podcasts)
OpenSLR-140	203,488	Read speech corpus (~554h)
KazEmoTTS	51,030	Emotional Kazakh TTS corpus (~75h)
KazakhTTS2	134,167	Multi-speaker TTS corpus (~271h, 5 speakers)

Total audio: ~2,100 hours of Kazakh speech

Mel Configuration

Parameter	Value
Sample rate	16,000 Hz
Mel bins	80
FFT size	400
Hop length	160
Max audio	30 seconds
Storage dtype	float16

Features

• mel — Log-mel spectrogram as a list of floats (shape: [n_mels, time_frames], stored flat, dtype float16)
• text — Transcription text (Kazakh)
• duration — Audio duration in seconds
• source — Source dataset name

Split Strategy

• Train: 94%
• Validation: 5%
• Test: 1%
• All splits are shuffled across all source datasets (seed=42)

Usage

from datasets import load_dataset

ds = load_dataset("stukenov/sozkz-asr-mels-kk-v1")

# Access a sample
sample = ds["train"][0]
mel = sample["mel"]       # flat list -> reshape to (80, T)
text = sample["text"]     # kazakh transcription
duration = sample["duration"]

License

MIT

Citation

Part of the SozKZ project — building open Kazakh language technology.