---
license: pddl
task_categories:
- automatic-speech-recognition
- audio-to-audio
- audio-classification
language:
- my
tags:
- speech
- audio
- burmese
- voa
- webdataset
- asr
- self-supervised
pretty_name: VOA Burmese – Self-Supervised ASR-Ready WebDataset
size_categories:
- 100K<n<1M
---

⸻

# Overview

This dataset was created by scraping and segmenting over 4,000 episodes of the VOA Burmese morning radio program. From that archive, 3,687 MP3 files were extracted and processed. This dataset contains sentence-level audio chunks suitable for ASR and speech-related model training.

The current release (`voa_batch_001.tar` and `voa_batch_003.tar`) contains a combined total of **~152,300 sentence-level audio chunks** derived from the first **420 MP3 files** in the archive, totaling approximately **420 hours** of segmented audio.

- 🗂️ Contains ~152,300 sentence-level audio chunks
- ⏱️ Total duration: ~420 hours (1.5 million seconds)
- 🔊 Average chunk length: ~9.9 seconds
- 🎵 Audio format: 16kHz mono MP3
- 📦 2 WebDataset .tar files 
- 📏 Min chunk: 0.04 sec | Max chunk: 15.00 sec
  
Each .mp3 is paired with a .json file containing structured metadata including file_name, broadcast_date, url, and duration.

⸻

## Usage

```python
from datasets import load_dataset

# Stream all TARs under /train/
ds = load_dataset("freococo/voa_myanmar_asr_audio_2", split="train", streaming=True)

sample = next(iter(ds))
meta = sample["json"]
print(meta["file_name"], meta["broadcast_date"], meta["duration"])
```
Each item will have:
```
{
  "file_name": "000123.mp3",
  "broadcast_date": "2020-11-02",
  "url": "https://voa-audio.voanews.eu/...",
  "duration": 9.82
}
```
⸻

## Dataset Construction
	•	Source audio was scraped from VOA Burmese morning radio
	•	Audio was chunked using ffmpeg into 5–15 second segments
	•	Metadata was saved alongside each chunk using Hugging Face-compatible .json files
	•	TAR files were created in WebDataset format with {file_name}.mp3 + {file_name}.json
	•	Each TAR contains up to ~75k examples to balance file size and upload speed

This structure is ideal for:
```
	•	Self-supervised learning (SSL)
	•	ASR fine-tuning
	•	TTS timbre modelling
	•	Speaker diarisation and segmentation tasks
```
⸻

## Acknowledgments

This dataset would not exist without the consistent commitment and public service of the Voice of America Burmese Service, including the dedicated journalists, producers, editors, and engineers who broadcast timely, trustworthy news to communities across Myanmar and the diaspora.

We offer special appreciation to:

	•	VOA multilingual teams who wrote, voiced, and produced the original programming
	•	The American people, whose public funding supports VOA’s mission and open access to information
	•	The global open-source and humanitarian tech community for providing tools, models, and infrastructure that made this project possible

We release this dataset in the spirit of:

	•	Advancing multilingual and under-resourced speech technology
	•	Supporting equitable access to information
	•	Amplifying the voices of communities often left out of the machine learning loop

## Motivation & Intent

This dataset was created out of deep love for language, sound, and the voices of ordinary people.  
No grants, scholarships, payments, or institutional support backed this project — only dedication, long nights, and a belief in the power of open knowledge.

I built this for students, researchers, developers, and wanderers — especially those with no funding but strong intention.  
To help Burmese voices be heard in the future of AI.  
To offer a foundation where none existed before.

This work is not for profit — it is for purpose.  
May it serve yours.  
And may it carry the quiet echo of care behind every `.mp3`.

## Known Limitations

This dataset was created through automatic chunking of full-length VOA Burmese radio broadcasts. As a result, users 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 include non-speech content, such as:
	•	Music intros and outros
	•	Jingles or filler segments
	•	Environmental noise or crowd sounds
	•	Silence or low-volume transitions
	•	🗣 Speaker identities are unlabeled. While there is speaker diversity (accents, gender, prosody), no segmentation or diarisation labels are provided.
	•	🎚 Broadcast production artifacts (e.g., fades, overlays, noise beds) may introduce noise that impacts ASR quality in some cases.

Despite these limitations, the dataset is highly suitable for:

	•	🔊 Pretraining speech models (e.g., wav2vec2-style)
	•	🎯 Self-supervised learning
	•	🧬 Voice activity detection (VAD) and segmentation
	•	🌍 Language ID and speaker clustering
	•	🧪 Synthetic sentence generation and augmentation

We recommend applying pre-processing filters such as VAD or manual review for tasks requiring high-quality or speech-only segments.

## Citation

If you use this dataset, please cite it as:
```
@dataset{freococo_voa_myanmar_asr_2025,
  author       = {freococo},
  title        = {VOA Burmese – ASR-Ready WebDataset},
  year         = {2025},
  url          = {https://huggingface.co/datasets/freococo/voa_myanmar_asr_audio_2},
  note         = {Segmented WebDataset format, public-domain VOA speech}
}
```
⸻