README.md

---
datasets:
- freococo/rohingya_asr_audio
language:
- rhg
tags:
- speech
- audio
- voa
- rohingya
- self-supervised
- webdataset
- public-domain
pretty_name: VOA Rohingya ASR
license: pddl
task_categories:
- automatic-speech-recognition
- audio-to-audio
- audio-classification
language_creators:
  - found
source_datasets:
  - original
---

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

## Overview

This dataset contains broadcast audio recordings from the **Voice of America (VOA) Rohingya Service**. Each file represents a daily news segment, typically 30 minutes in length, automatically segmented into chunks of 5–15 seconds for use in **self-supervised ASR**, **pretraining**, **language identification**, and more.

The content was aired publicly as part of VOA’s Rohingya-language radio program and is therefore released under a **public domain dedication** (U.S. Government speech, [17 U.S.C. § 105](https://www.govinfo.gov/content/pkg/USCODE-2011-title17/html/USCODE-2011-title17-chap1-sec105.htm)).

The dataset is stored in **WebDataset format**, with each `.tar` archive containing paired `.audio` (MP3) and `.json` metadata files for each segment.

## Acknowledgments

This dataset would not exist without the dedication and professionalism of the **Voice of America Rohingya Service** — especially the **journalists, editors, producers, and engineers** who continue broadcasting trusted news and public service content to marginalized communities.

Special gratitude goes to:

- VOA multilingual teams who **created, edited, and voiced** this content
- The **American people**, whose hard-earned taxpayer contributions make public media like VOA possible
- The open-source, low-resource, and humanitarian tech community — for tools, models, and continued support

This dataset is released in the hope that it will:
- Advance multilingual speech technology
- Empower access to information
- Amplify underrepresented voices across the world

## Metrics

| Metric            | Value        |
|-------------------|--------------|
| Total audio hours | **357.55 h** |
| Audio chunks      | **131,860**  |
| Shard count       | **14**       |
| Average chunk size| 6–15 sec     |
| Format            | WebDataset   |
| License           | Public Domain (VOA / U.S. Gov) |

## Quick-start

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

    from datasets import load_dataset

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

    for sample in dataset:
        print(sample["audio"])           # Audio object
        print(sample["file_name"])       # Chunk file name
        print(sample["download_url"])    # Original source URL
        print(sample["duration"])        # Duration in seconds

## Known Limitations

This dataset was created through automatic chunking of full-length VOA Rohingya news broadcasts. As a result, developers should be aware of the following limitations:

- **No transcriptions** are included. This dataset is not aligned for supervised training unless transcribed independently.
- Some chunks may contain **non-speech segments** such as:
    - Music intros and outros
    - Jingles or filler transitions
    - Background crowd noise or environmental sounds
    - Silent or low-audio intervals
- **No speaker labeling** is provided. Voice diversity, accents, and gender variation exist, but are unlabeled.
- **Broadcast mixing artifacts** may affect ASR performance in noisy conditions (e.g., overlayed music, crossfades, background hum).

Despite these challenges, the dataset is suitable for:
- Pretraining ASR models (wav2vec2-style)
- Unsupervised learning
- Language ID and diarization
- Synthetic data generation

We recommend applying **speech detection filters**, **VAD**, or **manual quality control** for downstream supervised tasks.

## Dataset Details

Each training sample is stored as:

- `.audio` — MP3 audio content (~5–15 seconds)
- `.json` — metadata with:

    - `file_name`: full chunk filename (e.g., `20250310_0001.audio`)
    - `original_file`: e.g., `20250310`
    - `publish_date`: ISO 8601 format (e.g., `2025-03-10`)
    - `download_url`: original VOA source URL
    - `duration`: chunk duration in seconds

These files are stored in `.tar` archives, split into ~10,000-sample shards named like:

    rohingya-00000.tar
    rohingya-00001.tar
    ...

Each archive follows [WebDataset format](https://github.com/webdataset/webdataset), making it easy to use with PyTorch and Hugging Face streaming.

## License & Reuse

All content is in the **public domain** under U.S. law:

> U.S. Government speech recordings (VOA staff broadcasts) are public domain under [17 U.S.C. § 105](https://www.govinfo.gov/content/pkg/USCODE-2011-title17/html/USCODE-2011-title17-chap1-sec105.htm).

Some broadcasts may contain music or third-party clips. Please verify manually if using for commercial purposes.

## Citation

If you use this dataset in research, please cite:

> **Freococo (2025).**  
> *VOA Rohingya ASR*  
> Hugging Face: [https://huggingface.co/datasets/freococo/rohingya_asr_audio](https://huggingface.co/datasets/freococo/rohingya_asr_audio)  
> Public-domain speech segments from VOA Rohingya news programming.  
> Released under `pddl`.