stukenov/kzcalm-mimi-codes-kk-v1

KZ-CALM Mimi Codec Codes — Kazakh v1

Discrete codec tokens extracted from 232,350 Kazakh speech utterances (~439 hours) using the Mimi neural audio codec (Kyutai).

This dataset is the B2 output of the KZ-CALM TTS pipeline — it bridges raw audio and the latent-space generative model.

Source

All audio comes from stukenov/kzcalm-tts-kk-v1, which merges:

Source	Samples	Hours	Speakers
KazakhTTS (ISSAI)	91,424	177.7	5 professional
KazEmoTTS (ISSAI)	140,926	261.1	3, 6 emotions
Total	232,350	438.8	8 unique

Codec Details

Parameter	Value
Codec	Mimi (Kyutai) via transformers.MimiModel
Weights	kyutai/mimi
Sample rate	24,000 Hz
Frame rate	12.5 Hz (1 frame = 80 ms)
Codebooks	8 (RVQ)
Codebook size	2,048 entries each

Schema

Column	Type	Description
text	string	Kazakh utterance text
speaker_id	string	Speaker identifier (e.g. ISSAI_KazakhTTS_M01)
source	string	KazakhTTS or KazEmoTTS
emotion	string	Emotion label (neutral, angry, happy, sad, surprised, scared, or empty)
duration	float	Audio duration in seconds
codes	list[list[int]]	Codec tokens — shape (8, T) where T = num_frames
num_frames	int	Number of codec frames (ceil(duration * 12.5))

Usage

from datasets import load_dataset

ds = load_dataset("stukenov/kzcalm-mimi-codes-kk-v1", split="train")
sample = ds[0]

codes = sample["codes"]   # list of 8 lists (one per codebook)
print(f"Text: {sample['text']}")
print(f"Codebooks: {len(codes)}, Frames: {sample['num_frames']}")
print(f"Duration: {sample['duration']:.1f}s")

Reconstruct audio

import torch
from transformers import MimiModel

model = MimiModel.from_pretrained("kyutai/mimi").cuda()
codes_tensor = torch.tensor([sample["codes"]], device="cuda")  # (1, 8, T)
waveform = model.decode(codes_tensor).audio_values  # (1, 1, T_samples)

Intended Use

• Training latent-space TTS models (flow matching, consistency models)
• Kazakh speech synthesis research
• Codec token language modeling

Encoding Process

1. Loaded audio via HuggingFace datasets streaming (24 kHz mono)
2. Batched (batch_size=16), padded to max length in batch
3. Encoded through MimiModel.encode() on GPU (RTX 3060)
4. Trimmed output codes to actual frame count per utterance
5. Total encoding time: ~5 hours on a single RTX 3060

License

CC-BY-4.0 (following the source datasets from ISSAI).

Citation

If you use this dataset, please cite the original corpora:

@inproceedings{mussakhojayeva2022kazakhtts,
  title={KazakhTTS: An Open-Source Kazakh Text-to-Speech Synthesis Dataset},
  author={Mussakhojayeva, Saida and Khassanov, Yerbolat and Varol, Huseyin Atakan},
  booktitle={Proc. LREC},
  year={2022}
}

@inproceedings{razakhan2024kazemotts,
  title={KazEmoTTS: A Dataset for Kazakh Emotional Text-to-Speech Synthesis},
  author={Razakhan, Adal and Mussakhojayeva, Saida and Khassanov, Yerbolat},
  booktitle={Proc. LREC-COLING},
  year={2024}
}