freococo/sunday_journal_voices

license: other licensename: custom licenselink: LICENSE pretty_name: Sunday Journal Voices language:

• —my tags:
• —speech
• —audio
• —asr
• —myanmar
• —low-resource
• —fair-use
• —tiktok
• —webdataset task_categories:
• —automatic-speech-recognition
• —audio-classification ---

Sunday Journal Voices

Voices that shape our week, stories that define our time.

Sunday Journal Voices is a large-scale collection of 26,288 short audio segments (≈18 hours total) derived from public videos by Sunday Journal — a leading digital media platform in Myanmar known for its in-depth reporting and interviews.

The source channel regularly publishes:

• —News analysis and commentary on current events
• —In-depth interviews with public figures, experts, and community leaders
• —Discussions on social, economic, and cultural topics
• —On-the-ground reporting and human-interest stories

These videos capture the pulse of contemporary Myanmar — spoken in clear, formal, and articulate Burmese, making it an invaluable resource for speech technology.

❤️ Why I Built This

• —Myanmar (Burmese) is often labeled a “low-resource language” in the AI world.
• —I don’t reject that label because it’s false — I reject it because it reflects global neglect.
• —I built this dataset to show what’s possible — to give Myanmar speech the visibility, respect, and technical foundation it deserves.

I care about languages. I care about people being heard. And if AI is going to learn from voices — I want it to hear mine, ours, Myanmar’s.

If you want your voice to be heard — you must first teach the machines to listen.

🕊️ Why It Matters to Me

We will come, and we will go. But if your voice is integrated into AI technology — it will go on. Forever.

I cannot build you a pyramid like the ancient Egyptians did. But I can build something more accessible, more global: A living archive — of your beautiful, strong, and clear voices.

Maybe, just maybe — AI will speak our beautiful Myanmar language through your voice. And I believe it will. I truly do. 🙂

🔍 What's Included

• —26,288 audio-text chunks
• —~18 hours of real Burmese speech (17h 44m 31s)
• —Auto-transcribed captions with timestamps
• —Rich video metadata (title, views, likes, hashtags)
• —WebDataset-ready .tar.gz for streaming & training

📂 Dataset Structure

This dataset is packaged as a single .tar.gz archive containing paired audio and metadata files in WebDataset format.

Each audio-text pair consists of:

• —.mp3 — a short audio chunk extracted from a video
• —.json — aligned metadata including the transcript and contextual information

All files are named using UUIDs:

a3f1d9e671a44b88.mp3
a3f1d9e671a44b88.json

Each .json file contains the following fields:

All audio-text pairs are organized flatly inside the .tar.gz archive.

The dataset can be streamed or loaded using PyTorch WebDataset, Hugging Face 🤗 Datasets, or custom Python loaders.

🚀 How to Use

This dataset is stored as a .tar.gz archive and is compatible with both 🤗 Hugging Face Datasets and 🧠 WebDataset training pipelines.

✅ Load using Hugging Face Datasets (streaming)

from datasets import load_dataset

# The library automatically finds the .tar.gz shard in the repo
ds = load_dataset(
    "freococo/sunday_journal_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("---")

🙏 Special Thanks

This dataset would not exist without the incredible journalistic and creative efforts of:

• —📰 Sunday Journal and its editorial team
• —🎤 The journalists, reporters, and on-screen hosts
• —🗣️ The interview subjects who shared their knowledge, stories, and perspectives
• —🎥 The producers, editors, and behind-the-scenes teams who bring these stories to light

These individuals are not just content creators — they are chroniclers of our time.

Their voices, analysis, and dedication have inspired a new generation of AI practitioners, researchers, and developers — not just in Myanmar, but around the world.

They prove that data isn't just numbers — it's memory. And in the hands of the right generation, it's power.

Thank you for giving us your voices. Now, they may echo in the machines we build — not to replace you, but to remember you.

🫡🇲🇲🧠📣

⚠️ Limitations

While this dataset offers significant access to formal Burmese speech, it is not without its imperfections.

Here are key limitations you should be aware of:

• —Auto-caption errors: All transcripts were likely generated using an automated system. Based on manual inspection of similar datasets, a portion of segments may contain minor to moderate transcription errors, including:
• —Missing particles or suffixes
• —Spelling inconsistencies
• —Punctuation hallucinations
• —Overlapping speaker boundaries

• —No human corrections (yet): No post-processing or human-in-the-loop editing was performed. This dataset reflects the raw, real-world performance of automated systems.

• —Audio quality: While generally clear, some clips may include:
• —Background noise from on-location reporting
• —Music overlays or sound effects
• —Interjections or non-speech sounds

• —Underrepresentation of regional dialects: Most speakers use standard, formal Burmese, with minimal representation of ethnic accents or regional variants.

✅ But Here's the Strength

Even if you filtered out noisy or imperfect lines, you'd still be left with over 18,000 to 20,000 high-quality speech chunks.

That alone makes this one of the largest publicly available Burmese speech datasets of its kind.

It is not perfect — but it is powerful. It is not corrected — but it is real. And real voices have always mattered more than perfect ones.

📌 Recommendation

If you plan to use this dataset for training high-performance ASR models or speech analytics:

It is strongly recommended to include human-in-the-loop correction, quality filtering, or semi-supervised fine-tuning with aligned text.

📄 License

This dataset is released under a Fair Use / Research-Only License.

It is intended for:

• —✅ Non-commercial research
• —✅ Educational use
• —✅ Language preservation
• —✅ Open AI development for Burmese (Myanmar) speech

All content was sourced from Sunday Journal's public channels. For any commercial inquiries, please contact the original content owner directly.

For full details, see the LICENSE file.

📚 Citation

@misc{freococo2025sundayjournalvoices,
  title        = {Sunday Journal Voices: A WebDataset for Burmese ASR and Speech Research},
  author       = {freococo},
  year         = {2025},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/freococo/sunday_journal_voices}
}