freococo/voa_myanmar_asr_audio_1

📢 This is the first publicly released ASR-ready Burmese speech dataset with over 1 million audio chunks — a milestone in the history of Myanmar language technology.

Overview

This dataset was created by scraping and segmenting the full archive of the VOA Burmese morning radio program. Out of a total of 3,687 full-length MP3 broadcasts, this release processes 3,267 of them, resulting in approximately 1.8 million sentence-level audio chunks, totaling ~3,267 hours of segmented audio. The dataset spans nearly a decade of daily VOA Burmese morning broadcasts, from 2015 to 2025, capturing speech across changing events, voices, and years.

This release includes 121 WebDataset .tar files (voa_myanmar_000001.tar → voa_myanmar_000121.tar) and provides a high-quality unsupervised audio corpus for speech modeling and self-supervised learning in the Burmese (Myanmar) language.

• 🔢 ~1.8 million sentence-level audio chunks
• ⏱️ Total duration: ~3,267 hours
• 🎧 Source MP3s: 3,267 of 3,687 original broadcasts
• 🔊 Average chunk length: ~6.5–7.5 seconds
• 📦 121 WebDataset TAR files
• 🧾 Each .mp3 is paired with a .json file containing:
  • file_name
  • original_file (e.g., voa_2105.mp3)
  • broadcast_date
  • url
  • duration
  • __key__

Note: Hugging Face’s streaming estimate may show a lower count (~1.43M) due to internal preview sampling. The actual total is ~1.8 million complete audio chunks.

Usage

from datasets import load_dataset

# Stream all TARs under /train/
ds = load_dataset("freococo/voa_myanmar_asr_audio_1", split="train", streaming=True)

sample = next(iter(ds))
meta = sample["json"]
print(meta["file_name"], meta["original_file"], meta["broadcast_date"], meta["duration"])

Each item is structured like:

{
  "__key__": "f9d3723f0f354f21",
  "mp3": ...,
  "json": {
    "file_name": "1260462.mp3",
    "original_file": "voa_2105.mp3",
    "broadcast_date": "2016-01-01",
    "url": "https://voa-audio.voanews.eu/vbu/2016/01/01/...",
    "duration": 9.63
  }
}

Dataset Construction

• All audio originates from Voice of America (VOA) Burmese radio broadcasts.
• A total of 3,267 original MP3 files were downloaded and processed.
• Each original .mp3 file was segmented into 5–15 second audio chunks using ffmpeg, with silence-based boundary detection.
• Python was used to build the processing pipeline, including:
  • pydub and ffmpeg for audio slicing
  • uuid for generating unique sample keys (__key__)
  • pandas for metadata management
  • tqdm for monitoring progress
  • webdataset to generate the final .tar archives
• Each chunk is paired with a .json sidecar containing:
  • file_name (e.g., 1260462.mp3)
  • original_file (e.g., voa_2105.mp3)
  • url pointing to the full original VOA broadcast
  • broadcast_date and chunk duration
  • __key__ (WebDataset sample ID)
• TARs were capped at 10,000 examples each to ensure optimal upload speed and streaming compatibility.
• The final dataset includes 121 .tar files, ranging from voa_myanmar_000001.tar to voa_myanmar_000121.tar.

Ideal For

This dataset is well-suited for training and evaluation tasks in speech technology, particularly for under-resourced and low-resource language modeling. Ideal applications include:

• 🧠 Self-supervised speech model pretraining (e.g., wav2vec2, HuBERT, Whisper-style models)
• 🗣️ Fine-tuning ASR models in Burmese (with external transcriptions)
• 🎚 Voice activity detection (VAD) and silence segmentation
• 🧬 Synthetic sentence generation and augmentation
• 🎙️ Speaker segmentation, diarisation, and prosody modeling
• 🌍 Language ID or multilingual audio classification
• 📦 WebDataset-based large-scale audio streaming pipelines

Acknowledgments

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

Motivation & Intent

This dataset was created out of deep love for language, sound, and the voices of ordinary people.

No grants, scholarships, or institutional support backed this work — only long nights, quiet persistence, and a belief that Burmese voices should be part of the future of AI.

It was built for:

• Students with no GPU but strong intention
• Researchers working in silence, far from labs
• Developers with one machine and big dreams
• Linguists, builders, wanderers — and you

This dataset is not for profit. It is for purpose. May it serve yours.
May it empower unheard voices.
May it carry the quiet echo of care behind every .mp3.

Known Limitations

This dataset was created through automated segmentation of full-length VOA Burmese radio broadcasts. As a result, users should be aware of the following limitations:

• ❌ No transcriptions are included — this is an unsupervised speech corpus
• 🔉 Some chunks may contain:
  • Intro or outro music
  • Commercial jingles or program IDs
  • Crowd noise, ambient background, or silence
• 🎙 Speaker identities are unlabeled, although the dataset contains a variety of speakers
• 🎧 Production artifacts such as fades, overlays, or compression may affect audio quality in some cases
• 🧩 Segment boundaries were determined via ffmpeg-based silence detection, not linguistically aligned

Despite these limitations, this dataset remains highly valuable for:

• Self-supervised learning (SSL)
• Voice activity detection (VAD)
• Sentence segmentation
• Pretraining large-scale ASR models

🇺🇸 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.

🕊

Tools & Reproducibility

This project was built entirely using open tools and reproducible Python code.

The following WebDataset creation script was used in practice to build over 1.5 million sentence-level audio chunks, split into 121 TAR archives. It was tested on real-world Burmese speech data, verified for alignment, metadata integrity, and upload stability.

You can adapt this script to build your own datasets — for audio, images, video, or any other chunked data paired with metadata.

---

Supported Use Cases

This script supports:

• ✅ Audio chunks with metadata (.mp3 + .json)
• ✅ Images with captions or tags (.jpg + .json)
• ✅ Video snippets with subtitles (.mp4 + .json)
• ✅ Any binary file + .json sidecar
• ✅ Hugging Face-compatible .tar output for WebDataset or streaming

---

🔧 Requirements

To use the script, first install the following Python packages:

pip install pandas webdataset tqdm

How To Use

Just update the paths and column names in the config section. Then run:

python createtar.py

Full Example Code

Save this code as createtar.py in your working directory.

import os
import uuid
import json
import shutil
import pandas as pd
import webdataset as wds
from tqdm import tqdm

# === CONFIGURATION ===
BATCH_ID = "001"  # Customize this batch label
CHUNKS_DIR = f"./chunks_batch_{BATCH_ID}"          # Folder with your .mp3 or other media chunks
METADATA_PATH = f"./metadata_batch_{BATCH_ID}.csv" # CSV file with metadata per chunk
OUTPUT_DIR = "./wds_output"                        # Where .tar files will be saved
TAR_PREFIX = "data_"                               # Prefix for each .tar file
TAR_SIZE = 10000                                   # Max samples per TAR
START_SHARD = 0                                    # Starting index for tar naming

# === Ensure output folder exists
os.makedirs(OUTPUT_DIR, exist_ok=True)

# === Load and clean metadata
df = pd.read_csv(METADATA_PATH)
df = df[df["file_name"].notnull()].reset_index(drop=True)
print(f"🔢 Total chunks to process: {len(df)}")

sample_count = 0
shard_count = 0
shard_writer = None

# === Process each row and write to WebDataset
for idx, row in tqdm(df.iterrows(), total=len(df)):
    file_path = os.path.join(CHUNKS_DIR, row["file_name"])
    if not os.path.exists(file_path):
        continue

    # Start new shard every TAR_SIZE
    if sample_count % TAR_SIZE == 0:
        if shard_writer:
            shard_writer.close()
        shard_writer = wds.ShardWriter(
            os.path.join(OUTPUT_DIR, f"{TAR_PREFIX}%06d.tar"),
            maxcount=TAR_SIZE,
            start_shard=START_SHARD + shard_count,
            maxsize=1 << 31
        )
        shard_count += 1

    # Load media file bytes
    with open(file_path, "rb") as f:
        media_bytes = f.read()

    # Generate metadata dictionary (customize fields as needed)
    metadata = {
        "file_name": row["file_name"],
        "original_file": row.get("original_file", ""),
        "broadcast_date": row.get("broadcast_date", ""),
        "url": row.get("url", ""),
        "duration": row.get("duration", ""),
    }

    uid = str(uuid.uuid4())[:16]  # Unique sample key

    # Write sample
    sample = {
        "__key__": uid,
        "mp3": media_bytes,             # Change key name if you're using another file type
        "json": json.dumps(metadata)
    }

    shard_writer.write(sample)
    sample_count += 1

    # Optionally delete chunk file after writing
    os.remove(file_path)

# === Finalize last shard
if shard_writer:
    shard_writer.close()

print(f"\n✅ Done! Total TARs written: {shard_count}")
print(f"🗃️  Files deleted from: {CHUNKS_DIR}")

Citation

If you use this dataset, please cite it as:

@dataset{freococo_voa_myanmar_asr_audio_1,
  author       = {freococo},
  title        = {VOA Burmese – ASR-Ready WebDataset (Part 1)},
  year         = {2025},
  url          = {https://huggingface.co/datasets/freococo/voa_myanmar_asr_audio_1},
  note         = {Segmented WebDataset format with full metadata and UUID keys}
}