stukenov/kzcalm-tts-kk-v1

KZ-CALM TTS Dataset v1 (Kazakh)

A unified Kazakh speech dataset for training text-to-speech (TTS) models. Compiled from two open-source ISSAI datasets — KazakhTTS and KazEmoTTS — with consistent preprocessing, quality filtering, and a single standardized schema.

Dataset Summary

Property	Value
Samples	232,350
Total audio	438.8 hours
Speakers	8 (5 professional studio + 3 emotional)
Emotions	6 (neutral, happy, sad, angry, scared, surprise)
Sample rate	24 kHz, mono, float32
Language	Kazakh (kk)
Format	176 Parquet shards with HF Audio feature
License	CC BY-SA 4.0

Sources

This dataset merges two publicly available Kazakh speech corpora produced by the Institute of Smart Systems and Artificial Intelligence (ISSAI) at Nazarbayev University:

Source	Samples	Hours	Description
issai/KazakhTTS	177,731	~364 h	5 professional speakers (3F + 2M), neutral read speech recorded in studio conditions. Covers news, books, and Wikipedia domains.
issai/KazEmoTTS	54,619	~75 h	3 speakers performing 6 emotions (neutral, happy, sad, angry, scared, surprise). Designed for emotional speech synthesis research.

Schema

Each sample contains the following fields:

Field	Type	Description
audio	Audio(sampling_rate=24000)	Audio waveform resampled to 24 kHz mono
text	string	Transcription text in Kazakh
speaker_id	string	Speaker identifier (e.g., F1, M2, 399172782)
source	string	Origin dataset: KazakhTTS or KazEmoTTS
emotion	string	Emotion label: neutral, happy, sad, angry, scared, surprise (KazakhTTS samples are all neutral)
duration	float64	Audio duration in seconds

Speakers

Speaker ID	Source	Gender	Notes
F1	KazakhTTS	Female	Professional studio recording
F2	KazakhTTS	Female	Professional studio recording
F3	KazakhTTS	Female	Professional studio recording
M1	KazakhTTS	Male	Professional studio recording
M2	KazakhTTS	Male	Professional studio recording
399172782	KazEmoTTS	—	6 emotions
805570882	KazEmoTTS	—	6 emotions
1263201035	KazEmoTTS	—	6 emotions

Preprocessing Pipeline

All audio was processed through a uniform pipeline before inclusion:

1. Resampling: All audio converted to 24 kHz mono using librosa.resample (Kaiser best).
2. Amplitude normalization: Peak-normalized to 0.95 to ensure consistent volume levels across speakers and sources.
3. Quality control filtering — samples were rejected if any of the following conditions were met:
   • Estimated SNR < 15 dB (too noisy)
   • Clipping detected (>1% of samples at max amplitude)
   • Duration outside the 1–30 second range
4. Rejection rate: 688 out of 233,038 total samples were filtered out (0.3% rejection rate).

Original Format Handling

• KazakhTTS: Distributed as tar.gz archives on HuggingFace. Extracted speaker directories containing Audios/*.wav and Transcripts/*.txt pairs, matched by filename stem.
• KazakhTTS2 (included in KazakhTTS): Nested structure with category subdirectories (Book, News, Wiki) containing Audio/*.wav and Transcripts/*.txt.
• KazEmoTTS: Distributed as zip archive. Structure: EmoKaz/{speaker_id}/{train|eval}/{speaker}_{emotion}_{utterance_id}.{wav|txt}.

Usage

from datasets import load_dataset

# Load full dataset
ds = load_dataset("stukenov/kzcalm-tts-kk-v1", split="train")

# Access a sample
sample = ds[0]
print(sample["text"])          # Kazakh transcription
print(sample["speaker_id"])    # e.g., "F1"
print(sample["emotion"])       # e.g., "neutral"
print(sample["duration"])      # e.g., 4.32 (seconds)

# Audio array and sample rate
audio = sample["audio"]["array"]      # numpy float32 array
sr = sample["audio"]["sampling_rate"] # 24000

# Filter by speaker
f1_data = ds.filter(lambda x: x["speaker_id"] == "F1")

# Filter emotional speech only
emotional = ds.filter(lambda x: x["emotion"] != "neutral")

# Streaming mode (no disk download)
ds_stream = load_dataset("stukenov/kzcalm-tts-kk-v1", split="train", streaming=True)
for sample in ds_stream:
    print(sample["text"])
    break

Intended Use

This dataset is designed for:

• Training Kazakh TTS models (the primary purpose — part of the KZ-CALM project)
• Kazakh ASR (automatic speech recognition) research
• Speaker verification / identification experiments
• Emotional speech synthesis research (KazEmoTTS subset)
• Kazakh language processing benchmarks

Limitations

• All speakers are from a limited pool (8 total). The dataset does not represent the full diversity of Kazakh dialects or accents.
• KazEmoTTS emotion labels are acted (not spontaneous), which may limit naturalness in emotional TTS.
• KazakhTTS recordings are studio-quality; models trained on this data may not generalize well to noisy environments.
• Text normalization is minimal — numbers, abbreviations, and special characters may appear in raw form.

Citation

If you use this dataset, please cite the original ISSAI publications:

@inproceedings{mussakhojayeva2021kazakhtts,
  title={KazakhTTS: An Open-Source Kazakh Text-to-Speech Synthesis Dataset},
  author={Mussakhojayeva, Saida and Khassanov, Yerbolat and Varol, Huseyin Atakan},
  year={2021}
}

@inproceedings{mussakhojayeva2024kazemotts,
  title={KazEmoTTS: A Dataset for Kazakh Emotional Text-to-Speech Synthesis},
  author={Mussakhojayeva, Saida and others},
  year={2024}
}