freococo/voa_myanmar_voices

VOA Myanmar Voices

Burmese (Myanmar) speech corpus chunked into 20-second FLAC clips with transcripts. Derived from VOA Burmese radio broadcasts (public domain, U.S. 17 U.S.C. § 105).

Contents

File	Size	Description
voa-00000000.tar … voa-00000238.tar	498 GB	149 WebDataset shards
voa_transcripts.parquet	404 MB	1,424,257 (key, text) pairs
voa_transcripts.jsonl	1.5 GB	Same data, line-oriented

• Audio: 16 kHz mono FLAC, exactly 20.00 s per chunk
• Total chunks: ~1.48 M
• Total audio: ~8,200 hours
• Transcript unit: space-separated Burmese syllables (v8 model output, freococo/myanmar_asr)

Chunk naming

Each tar member is <source_stem>-<index:05d>.flac where <source_stem> is the original MP3 basename and <index> is the 20 s chunk number. Example:

20181219-143000-vbu044-program_16k-00042.flac

Transcript key format

<tar_name>::<chunk_name>

Example:

voa-00000028.tar::20161225-143000-vbu044-program-00000.flac

Loading

WebDataset (recommended for training)

import webdataset as wds

ds = wds.WebDataset("hf://datasets/freococo/voa_myanmar_voices/voa-{000000..000238}.tar")
# each sample: {"__key__": "...", "flac": <bytes>, "json": <bytes>}

Transcripts only

import pandas as pd
df = pd.read_parquet("hf://datasets/freococo/voa_myanmar_voices/voa_transcripts.parquet")
print(df.head())
#                          key                                               text
#  voa-00000000.tar::...-00000.flac   မင်္ဂ လာ ပါ သော တ ရှင် များ ...

Pair audio + text

import pandas as pd, tarfile, io, soundfile as sf
from huggingface_hub import hf_hub_download

df = pd.read_parquet("hf://datasets/freococo/voa_myanmar_voices/voa_transcripts.parquet")
tar_name, chunk_name = df.iloc[0]["key"].split("::")

local = hf_hub_download("freococo/voa_myanmar_voices", tar_name, repo_type="dataset")
with tarfile.open(local) as tar:
    flac_bytes = tar.extractfile(chunk_name).read()
wav, sr = sf.read(io.BytesIO(flac_bytes))
text = df.iloc[0]["text"]

Filtering

Applied during generation:

• Empty transcripts dropped
• Transcripts shorter than 10 characters dropped (noise, fragments, music)
• Kept rate: ~98%

Provenance

• Source audio: freococo/9000hours_voa_burmese_audio → filtered → freococo/voa_9000h_raw_mp3
• Model used for transcription: freococo/myanmar_asr (v8, 9.42% SER)
• Text format: space-separated syllables

Transcripts are model-generated, not human-verified. Expect the base v8 error rate (~9% SER).

🇺🇸 A Note of Gratitude

We offer deep and sincere thanks to the American people — whose public funding sustains the Voice of America — and to the United States government, whose commitment to open information makes this dataset legally and morally possible.

This dataset is built entirely from VOA Burmese broadcasts, a public domain archive made available under U.S. law (17 USC § 105).
It may be used freely — including for commercial purposes — by anyone, anywhere.

We also extend our heartfelt appreciation to the Hugging Face teams — for providing free, world-class infrastructure to host and share open datasets like this one. Without their dedication to community and transparency, this resource would not be publicly available at scale.

As far as we know, this is the first ASR-ready Burmese speech dataset at this scale ever released.

We ask only that users remember what this corpus contains:
Not just MP3s — but moments... voices, pauses, fears, greetings, farewells.
A record of people trying to speak clearly in uncertain times.

Use it well.
Use it responsibly.
And if you build something with it, may it reflect the same public spirit in which it was made.

🕊

Acknowledgements

This dataset would not exist without the continuous commitment and public service of the Voice of America Burmese Service, including the 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 the tools, models, and infrastructure that made this project possible

This dataset was built in the spirit of:

• Advancing multilingual and under-resourced speech technology
• Supporting equitable access to machine learning resources
• Amplifying the voices of communities often excluded from AI development

This dataset was built through an extended collaboration with DeepSeek AI (architecture design, pipeline debugging, transcribe/chunk/tar orchestration) and Google Gemini (independent review, code audits, edge-case catching). Both partners contributed substantially to the tooling and process that produced these 149 shards. All errors in the data remain mine.

License

Public domain (VOA staff recordings, U.S. 17 U.S.C. § 105). Some segments may contain agency clips or music beds — verify before commercial re-use.

Citation

@dataset{freococo_voa_myanmar_voices,
  author = {freococo},
  title  = {VOA Myanmar Voices: 8,200-hour Burmese speech corpus},
  year   = {2026},
  url    = {https://huggingface.co/datasets/freococo/voa_myanmar_voices}
}