freococo/myanmar_asr

Myanmar ASR MODEL

An end-to-end Burmese (Myanmar) speech recognition model. Trained from scratch on 1.66M utterances of Myanmar speech.

23.8M parameters. 9.42% SER on held-out data.

Highlights

• Small & fast — 23.8M parameters. Runs efficiently on CPU and consumer GPUs.
• End-to-end — raw audio in, Burmese syllables out.
• Trained from scratch — no fine-tuning from an English or multilingual base.
• Robust — trained on mixed-domain audio: news, conversation, reading, and aggressively augmented variants.
• Open — Apache 2.0, weights + code + configs all published.

Model Architecture

Feature	Specification
Parameters	23,753,868
Architecture	12-layer Conformer encoder, joint CTC + attention
Encoder width	256
Attention heads	4 (RoPE positional encoding)
Conv kernel	31
Vocabulary	2,566 Burmese syllable tokens
Input	16 kHz mono, 80-dim log-mel spectrogram
Output	Space-separated Burmese syllables

Training Details

Parameter	Value
Corpus	~1.66M utterances of Burmese speech
Epochs	25
Optimizer steps	178,325
Optimizer	AdamW, cosine decay, warmup 1000
Batch size	180 s / 90 clips per micro-step, grad_accum=4
Peak LR	4e-4
Hardware	1× NVIDIA RTX 3060 (12 GB)
Wall time	~42 hours

Results

Split	Size	SER
Held-out validation	4,997	9.42%
Independent web crawl	18,896	12.30%

Metric — Syllable Error Rate (SER): Syllables are the natural phonological unit of Burmese. WER is not well-defined for Burmese (no orthographic word boundaries); CER is biased by variable code-point counts per syllable. SER gives a consistent, linguistically grounded error rate.

Usage

1. Installation

pip install -r requirements.txt

2. Command-Line Inference

# Transcribe a single file on CPU or GPU
python transcribe.py example.mp3 --device cpu
python transcribe.py input.wav --device cuda

# Batch transcribe multiple files to a text file
python transcribe.py *.wav --output results.txt

3. Python API

from transcribe import BurmeseASR

asr = BurmeseASR(device="cuda") # or device="cpu"
text = asr.transcribe("example.mp3")
print(text)

Example Output

Input : example.mp3 (3.13 s Burmese news clip)
Output: စစ် ကောင် စီ အ တွက် လေ ယာဉ် ဆီ ရောင်း ချ ဖြန့် ဖြူး ပေး နေ တဲ့ လူ

Release Files

File	Purpose
model.safetensors	Bit-identical FP32 model weights (95 MB)
model.py	Standalone PyTorch Conformer ASR architecture
transcribe.py	Production inference script (CLI + Python API)
config.json	Model configuration and hyperparameter specifications
vocab.json	Token-to-ID mapping (2,566 tokens)
cmvn.json	Mel-spectrogram normalization statistics (mean/std)
preprocessor_config.json	Audio feature extraction parameters
example.mp3	Verification sample audio
requirements.txt	Python runtime dependencies

Notes

• Audio preprocessing: Audio must be 16 kHz mono. transcribe.py handles resampling and downmixing automatically for common formats (wav, mp3, flac, m4a).
• Tokenization: The output consists of space-separated Burmese syllables matching the training vocabulary.
• Decoding: The joint CTC/Attention architecture includes the attention decoder in the weights, while the CTC head is used for fast streaming/greedy transcription.

Dedication & Acknowledgements

This model is dedicated to the preservation and accessibility of the Burmese language for future generations.

AI Engineering Partners

• DeepSeek AI: Primary co-engineering partner through every stage — architecture design, memory optimization, training pipeline, debugging, and evaluation. Lead technical collaborator for this project.
• Gemini AI (Google): Secondary review partner for code audits, metric analysis, and independent verification of results.

Data Contributors This model would not exist without the voices and archives of:

• National Unity Government (NUG) of Myanmar
• PVTV broadcasters and journalists
• Myanmar Celebrity Voices contributors
• Media Queen Entertainment archives
• Sunday Journal news team
• Khit Thit Media news team
• MRTV and MRTV-4 broadcast archives
• Google Myanmar ASR open corpus
• Myanmar Bible Speech corpus contributors
• Original Myanmar Voices open contributors
• Independent Myanmar journalists, broadcasters, and creators whose voices form the backbone of this language-preservation effort

To the voices of a nation — thank you. This model exists because you spoke, and your speech deserves to be understood.

Statement of Independence

This model was built independently — no funding, no sponsorship, no instructions from any organisation, government, company, or individual. It is a gift to the Burmese language community.

Every dataset used in this project is publicly available through this account.

The weights, code, and configurations are released under Apache 2.0. Use them however you want. No restrictions. The only goal is benefit for Myanmar and its language.

If this model produces incorrect or unexpected output, that responsibility is mine alone — freococo. Not the original creators of the source data, whose voices and archives made this work possible. All credit for the data belongs to them. All errors in the model belong to me.

Citation

@misc{myanmar_asr_2026,
  title  = {myanmar_asr: End-to-end Burmese Speech Recognition},
  author = {freococo},
  year   = {2026},
  url    = {https://huggingface.co/freococo/myanmar_asr}
}