freococo/mrtv_news_voices

🗣️ Overview

MRTV Voices is a large-scale Burmese speech dataset built from publicly available news broadcasts and programs aired on Myanma Radio and Television (MRTV) — the official state-run media channel of Myanmar.

🎙️ It contains over 130,000 short audio clips (≈117 hours) with aligned transcripts derived from auto-generated subtitles.

This dataset captures:

• Formal Burmese used in government bulletins and official reports
• Clear pronunciation, enunciation, and pacing — ideal for ASR training
• Diverse speech contexts: news reading, field reporting, interviews, and announcements

Whether you're working on:

• Automatic Speech Recognition (ASR)
• Language model fine-tuning
• Media analysis
• or Myanmar speech research

...MRTV Voices offers one of the most structured and scalable resources for developing Burmese AI systems.

📂 Dataset Structure

This dataset is packaged as a single .tar.gz archive (mrtv_voices_000001.tar.gz) using the WebDataset format, which is optimized for high-throughput streaming and model training.

Each audio-text pair includes:

• 🎧 .mp3 — a short audio chunk (typically 1–6 seconds)
• 📄 .json — a rich metadata file containing the transcript and contextual info

All files are named using random UUIDs (e.g., d84f1e09f1ab4a79.mp3), and grouped flatly inside the archive.

📦 Inside each .json file:

Field	Description
file_name	Chunk filename (e.g., abc123.mp3)
original_file	The original MRTV video filename
transcript	Subtitle text in Burmese
duration	Audio duration (in seconds)
video_url	Link to original video
language	Always "my" (Myanmar)
title	Title of the broadcast or clip
description	Additional context or social caption
view_count	Views at the time of scraping
like_count	Likes at the time of scraping
comment_count	Comment count
repost_count	Repost/share count
channel	Channel name (e.g., MRTV)
upload_date	Date in YYYYMMDD format
hashtags	List of hashtags (if any)
thumbnail	Thumbnail image URL
source	Always: Myanma Radio and Television

---

The dataset can be loaded using:

• 🤗 Hugging Face Datasets (load_dataset(..., streaming=True))
• 🧠 WebDataset for PyTorch
• Custom Python scripts

🚀 How to Use

You can stream this dataset directly using 🤗 datasets. Each sample contains both the transcript and full metadata.

from datasets import load_dataset

# The library automatically finds the .tar.gz shard in the repo
ds = load_dataset(
    "freococo/mrtv_news_voices",
    split="train",
    streaming=True
)

# Iterate through the first 5 samples
for sample in ds.take(5):
    # The transcript is a top-level feature for easy access
    print(f"🎙️ Transcript: {sample['txt']}")

    # The audio data is in the 'mp3' column
    audio_data = sample['mp3']
    audio_array = audio_data['array']
    sampling_rate = audio_data['sampling_rate']
    print(f"🎧 Audio loaded with shape {audio_array.shape} and rate {sampling_rate} Hz")

    # The 'json' column is ALREADY a Python dictionary
    metadata = sample['json']
    print(f"📺 Channel: {metadata.get('channel')}")
    print(f"🎥 Video URL: {metadata.get('video_url')}")
    print("---")

⚠️ Limitations

While this dataset offers broad coverage of Burmese speech and media clips, it is important to understand its limitations:

• Propagandistic content: MRTV is a state-run broadcaster currently operated by the military regime. Some content may contain nationalist, anti-democracy, anti-Western, or racially biased perspectives.
• Zero human correction: All transcriptions were scraped directly from TikTok auto-captions, with no manual editing. Around 25–35% may contain spelling errors, punctuation artifacts, or missing words.
• Censorship bias: Editorial control by the regime means some topics (e.g. dissent, protest) are absent or misrepresented.
• Speaker diversity: Most clips are from formal news or PR settings. Street speech, dialects, and youth slang are underrepresented.
• ASR challenges: Some captions have poor alignment or are missegmented, which could affect training quality unless filtered or fine-tuned.

---

✅ Why It’s Still Useful

Despite the caveats:

Even if you discarded every noisy sample, you’d still have over 100,000 high-quality Burmese audio-text pairs — suitable for ASR training, forced alignment, or language model pretraining.

This is one of the largest open Burmese speech datasets ever released — with real voices, real pronunciation, and full metadata.

Use with care. Filter where needed. And always document your preprocessing steps when publishing derived models.

📜 License and Intent

This dataset is released under the Universal Public Domain Dedication (CC0 1.0 Universal) license.

All content was collected from the publicly accessible state-owned Myanmar Radio and Television (MRTV) channel on TikTok.

---

⚠️ Important Context

• MRTV is currently operated by the Myanmar military regime.
• Videos were produced using public funds, state media resources, and government-owned facilities.
• As such, the content belongs to the people of Myanmar — and should be available for open use, research, and analysis.

---

✅ You May:

• Use, modify, and redistribute this dataset.
• Train open-source or commercial AI models.
• Build tools that help preserve the Burmese language.
• Analyze, audit, or critique the media for public accountability.

---

❌ You May NOT:

• Misrepresent the source of the data.
• Use the voices or transcriptions to impersonate real people or endorse ideologies.
• Claim ownership of the original MRTV media content.

---

This dataset is shared in the public interest, for language accessibility, open research, and education.

It does not endorse any specific political regime, message, or actor.

Use responsibly, document your pipeline, and help make Burmese speech technology accessible to all.

💬 Author’s Note

This dataset was compiled, cleaned, and published by freococo, a Myanmar-born open-source advocate and AI researcher.

I receive no financial compensation for this work.

I built this dataset because:

• 🧠 Burmese (Myanmar) is still underrepresented in speech and language technology.
• 📣 Our voices deserve to be part of global AI — not left behind.
• 🕊️ Language preservation is resistance against cultural erasure.

---

MRTV is a controversial source — I understand that. But speech is speech.
And if we're training machines to listen, they must hear everyone.

This dataset includes laughter, announcements, news delivery, monologues, interviews, and everyday public speech from Myanmar.

Even if it's imperfect or biased — it’s real.
It reflects how Burmese is spoken in state media, with clear pronunciation, varied cadence, and diverse speaker profiles.

Use this to build tools, evaluate speech systems, or simply understand how Burmese sounds in a formal media setting.

This is not just about data — it’s about visibility, dignity, and linguistic equality.

🫡🇲🇲

📚 Citation

@misc{freococo2025mrtvvoices,
  title        = {MRTV Voices: A Burmese Speech Dataset from State Broadcasts},
  author       = {freococo},
  year         = {2025},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/freococo/mrtv_voices}
}