btsee/common-voices-24-mn

Common Voice 24.0 Mongolian - Cleaned Dataset

A cleaned and preprocessed version of the Mozilla Common Voice 24.0 Mongolian (Khalkha) dataset, optimized for Text-to-Speech (TTS) training.

Dataset Description

This dataset contains high-quality, cleaned audio recordings of Mongolian speech derived from Mozilla Common Voice version 24.0.

Processing Pipeline

All audio samples have been processed through the following pipeline:

1. Quality Filtering: Only samples with >0 upvotes and 0 downvotes were selected
2. Noise Suppression: DeepFilterNet3 applied for background noise removal
3. Resampling: Audio resampled to 22,050 Hz (optimal for TTS)
4. Normalization: Peak amplitude normalized to 0.95
5. Silence Trimming: Leading/trailing silence removed (20dB threshold)
6. Duration Filtering: Only clips between 0.5-15 seconds retained

Dataset Statistics

Metric	Value
Total Clips	~28,900
Total Duration	~40 hours
Sample Rate	22,050 Hz
Audio Format	WAV (16-bit PCM)
Language	Mongolian (Khalkha Cyrillic)
Unique Speakers	~1,500+

Dataset Structure

DatasetDict({
    train: Dataset({
        features: ['audio', 'text', 'speaker_id', 'duration'],
        num_rows: ...
    })
})

Features

• audio: Audio waveform (automatically decoded)
• text: Mongolian Cyrillic transcription
• speaker_id: Anonymized speaker identifier (first 8 chars of hash)
• duration: Audio duration in seconds

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("btsee/common-voices-24-mn")

# Access a sample
sample = dataset["train"][0]
print(sample["text"])  # Mongolian transcription
print(sample["audio"]["array"])  # Audio waveform
print(sample["audio"]["sampling_rate"])  # 22050

For TTS Training

from datasets import load_dataset

dataset = load_dataset("btsee/common-voices-24-mn", split="train")

# Filter by duration for VITS2 training
dataset = dataset.filter(lambda x: 1.0 <= x["duration"] <= 10.0)

# Group by speaker for multi-speaker TTS
speakers = dataset.unique("speaker_id")

Intended Uses

• Text-to-Speech (TTS): Training Mongolian speech synthesis models (VITS2, Tacotron2, etc.)
• Automatic Speech Recognition (ASR): Fine-tuning Whisper or Wav2Vec2 for Mongolian
• Voice Cloning: Multi-speaker TTS development
• Linguistic Research: Mongolian phonetics and prosody studies

Source

• Original Dataset: Mozilla Common Voice 24.0
• Processing Code: OronTTS

License

This dataset inherits the CC0 1.0 Universal license from Mozilla Common Voice.

Citation

If you use this dataset, please cite:

@dataset{common_voices_24_mn_cleaned,
  title = {Common Voice 24.0 Mongolian - Cleaned Dataset},
  author = {OronTTS Contributors},
  year = {2026},
  url = {https://huggingface.co/datasets/btsee/common-voices-24-mn},
  note = {Cleaned and preprocessed for TTS training}
}

@inproceedings{commonvoice:2020,
  author = {Ardila, Rosana and others},
  title = {Common Voice: A Massively-Multilingual Speech Corpus},
  booktitle = {LREC},
  year = {2020}
}

Acknowledgements

• Mozilla Foundation for the Common Voice project
• DeepFilterNet team for the noise suppression model
• All Mongolian speakers who contributed to Common Voice