freococo/zomi_asr

This is the first public Zomi language ASR dataset in AI history.

Zomi ASR

This dataset contains audio recordings and aligned metadata in the Zomi language — a collective ethnolinguistic identity adopted by some Kuki-Chin language-speaking communities in Myanmar and India. The term Zomi means "Zo people", derived from the root word Zo (ancestral identity) and mi meaning "people." While originally coined to encompass all Zo-related communities, usage of the term varies regionally and politically.

All audio segments in this dataset were sourced from publicly available news broadcasts by Zoland Voice TV, an ethnic-language news channel affiliated with the National Unity Government (NUG) of Myanmar. These broadcasts promote information access in minority languages, including Zomi.

The dataset includes over 18.99 hours of segmented and labeled audio, prepared in WebDataset format, with paired .audio and .json files suitable for training automatic speech recognition (ASR) systems.

Acknowledgments

Special thanks to:

• Zoland Voice TV and PVTV for producing and releasing multilingual content freely
• National Unity Government (NUG) for supporting inclusive language outreach
• Volunteers and researchers advancing low-resource ASR for ethnic languages

Dataset Structure & Format

This dataset follows the WebDataset format. Each training sample consists of two paired files inside a tar archive:

• XXXX.audio — the audio chunk (in MP3 format)
• XXXX.json — the corresponding metadata (UTF-8 JSON)

🟢 Minimum chunk duration: 2.04 sec
🔴 Maximum chunk duration: 15.05 sec

Each .json file contains the following fields:

{
  "file_name": "XXXX.audio",
  "video_id": "YouTubeVideoID",
  "title": "Original broadcast title from Zoland Voice TV",
  "url": "https://www.youtube.com/watch?v=YouTubeVideoID",
  "duration": 13.24
}

Usage Example

You can load and stream this dataset using the Hugging Face datasets library:

    from datasets import load_dataset

    dataset = load_dataset(
        "freococo/zomi_asr",
        split="train",
        streaming=True
    )

    for sample in dataset:
        print(sample["audio"])        # Audio object
        print(sample["file_name"])    # Chunk filename
        print(sample["duration"])     # Duration in seconds
        print(sample["title"])        # Broadcast title
        print(sample["url"])          # YouTube source URL

Each sample includes:

• audio: the audio chunk (stored as .audio, typically MP3 format)
• file_name: filename of the chunk
• title: broadcast title in Zomi or Burmese
• url: original YouTube video link
• video_id: YouTube video ID
• duration: duration of the audio in seconds

Known Limitations

This dataset was segmented automatically from broadcast videos using pause-based or fixed-length chunking. As such:

• No transcriptions are included.
• Some chunks may contain background music, news jingles, or non-speech segments.
• No speaker labels, noise filtering, or speech-vs-music tagging is applied.
• Audio quality varies depending on the original broadcast conditions.

Despite these limitations, this dataset is the most comprehensive public resource available for developing ASR and pretraining models in the Zomi language.

Licensing & Use

All content is released under the Creative Commons Zero (CC0 1.0 Universal) public domain dedication.

You are free to:

- Use, adapt, and remix the data
- Train both open and commercial models
- Publish derivative works, applications, and papers

We ask users to respect the dignity and intent of the original community broadcasts.

📚 Citation

Freococo (2025).
Zomi ASR
https://huggingface.co/datasets/freococo/zomi_asr
Dataset compiled from Zoland Voice TV ethnic news broadcasts in the Zomi language.
Released under CC0 1.0 (Public Domain).