metadata

license: cc0-1.0
tags:
  - automatic-speech-recognition
  - audio
  - burmese
  - myanmar
  - public-domain
  - webdataset
  - civic-media
  - news-broadcast
  - low-resource
  - spoken-burmese
language:
  - my
pretty_name: 115 Hours PVTV Myanmar ASR
task_categories:
  - automatic-speech-recognition
  - audio-classification

115 Hours PVTV Myanmar ASR

This dataset contains 156,262 audio-transcript pairs of spoken Burmese, totaling approximately 115.31 hours. The audio segments were extracted from publicly available YouTube videos published by PVTV and aligned using subtitle timestamps.

Dedication

This dataset would not exist without the persistent voices of PVTV editors, journalists, narrators, and production teams, who continue to speak to the people under difficult conditions. PVTV is the official media channel of the National Unity Government (NUG) of Myanmar — the democratic government formed by the people's elected representatives after the 2021 coup.

Though their names may not appear in the metadata, their voices carry the dignity and defiance of a nation still fighting for its future.

To those who speak truth in the dark,
to those who read the news when others remain silent,
to those who educate under censorship —
your voices live here, in waveform and in memory.

This dataset is for the people of Myanmar — free, public, and open to all.

License

This dataset is released under the CC0 1.0 Universal license, placing it fully in the public domain.

All audio was extracted from publicly available content published by PVTV, the official media organ of Myanmar’s National Unity Government (NUG). As the NUG represents the will of the people, this dataset is offered as public property — for the people, by the people.

The dataset was created voluntarily, without sponsorship or payment, in solidarity with the open education and civic information efforts of the Myanmar people.

Acknowledgements

This dataset was prepared independently by @freococo, using only public media aligned with the values of the democratic movement in Myanmar.

It is dedicated to those who continue to speak, teach, and inform under conditions of risk — and to the researchers and students who will carry this language forward in open-source speech technology.

To those who speak the truth in the dark,
to those who read the news when others remain silent,
to those who narrate education under censorship —
your voice is not forgotten. It lives here, in waveform and in memory.

This dataset is shared in that same spirit: free, public, and open to all.

Dataset Structure

The dataset is organized in WebDataset format using .tar archives.

Each archive contains:

XXXXXX.mp3 — short audio segment (1–15 seconds)
XXXXXX.json — metadata with:
- "transcript": spoken Burmese sentence
- "duration": audio length in seconds

All .tar files are stored in the train/ folder.

Usage Example

from datasets import load_dataset

ds = load_dataset(
    "freococo/115hours_pvtv_myanmar_asr",
    data_dir="train",
    split="train",
    streaming=True
)

sample = next(iter(ds))
print(sample["json"])          # {'transcript': '...', 'duration': 1.92}
print(sample["mp3"]["array"])  # NumPy waveform

Known Issues and Suggestions

Some transcripts may contain small alignment offsets due to the timing accuracy of the original .srt files.
Speaker identity is not annotated.
Regional accents and tone variation are not labeled.
Some very short clips (under 1.5 seconds) are included for completeness, but may not be ideal for certain models.

Suggestions

If you are a university undergraduate or postgraduate student working in speech recognition, linguistics, or Burmese NLP — this dataset is a valuable opportunity. You can:

Train baseline ASR models (e.g. Whisper, Wav2Vec2)
Analyze spoken Burmese syntax and phonology
Explore dialect and speaker variation
Build tools for civic education or accessibility

The dataset is fully open and public domain — you are free to use it without restrictions. Please share your work so others can build on it.

Citation

If you use this dataset in your research or applications, please cite:

@misc{freococo2025pvtv,
  title = {115 Hours PVTV Myanmar ASR},
  author = {freococo},
  year = {2025},
  howpublished = {\url{https://huggingface.co/datasets/freococo/115hours_pvtv_myanmar_asr}},
  note = {Public-domain Burmese ASR dataset from PVTV broadcasts.}
}