myanmar_asr_google_openslr80

Fine-tuned variant of freococo/myanmar_asr specialized for the Google Myanmar ASR Dataset (OpenSLR-80).

Same architecture, same code, same tokenizer — only the weights differ.

Results

Model	SER on Google val (200 utts)
Base myanmar_asr	8.70%
This model (fine-tuned)	6.97%

−1.73 point improvement from 1,800 fine-tune samples in ~16 minutes on a free Colab T4.

Metric Notes

• Syllable Error Rate (SER) computed under batch-1 inference — matches how transcribe.py processes single files.
• Batch-8 padded inference reports 5.24%, reflecting the training distribution. Batch-1 is the honest deployment number.

Fine-Tuning Details

Parameter	Value
Base checkpoint	freococo/myanmar_asr (val SER 9.42%)
Fine-tune dataset	freococo/google_myanmar_asr_voices
Train / val split	1,800 / 200 (10%, seed 42)
Peak LR	1e-5 (AdamW, cosine decay)
T_max	100 epochs
Early stopping	Patience 10 on val SER
Epochs run	42 (best at epoch 32)
Hardware	1× Google Colab T4 (15 GB)
Wall time	~16 min

Reproduce

A complete Colab notebook that runs the fine-tune end-to-end is attached to this repository (finetune_google_openslr80.ipynb).

Key hyperparameters:

FINETUNE_LR = 1e-5
EPOCHS      = 100  # schedule length; early stopping ends the run
PATIENCE    = 10   # stop after 10 epochs without SER improvement

Usage

Identical to the base model — same files, same CLI, same API.

1. Installation

pip install -r requirements.txt

2. Command-Line Inference

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

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

3. Python API

from transcribe import BurmeseASR

asr = BurmeseASR(model_dir="path/to/myanmar_asr_google_openslr80", device="cuda")
text = asr.transcribe("input.wav")
print(text)

Example Output

Input : example.mp3 (3.13 s Google Myanmar clip)
Output: ဆို တော့ တယ် လီ ဖုန်း အော် ပ ရေ တာ ဖြစ် လာ ရင် ကော ဝန် ဆောင် မှု ပိုင်း က ကောင်း နိုင် ပါ့ မ လား

When to Use This Model

• Use this model for clean, standard read speech resembling OpenSLR-80 / Google Myanmar.
• Use the base model (freococo/myanmar_asr) for broader domain coverage — noisy news, conversations, and mixed sources.
• Fine-tuning on 1,800 clean samples specializes the model to this domain; it is intentionally focused rather than universally superior across all acoustic environments.

Fine-Tune on Your Own Domain

This repository demonstrates a fast adaptation recipe: any target domain with ~2,000 samples can be specialized in ~15 minutes on a free GPU:

1. Prepare a manifest of (audio_file, syllabized_transcript) pairs.
2. Load the base weights from freococo/myanmar_asr.
3. Fine-tune with LR 1e-5, cosine schedule, and early stopping on validation SER.
4. The Colab notebook finetune_google_openslr80.ipynb in this repo can be directly adapted.

Release Files

File	Purpose
model.safetensors	Fine-tuned FP32 model weights (~95 MB)
model.py	Standalone ConformerASR architecture
transcribe.py	Production inference script (CLI + Python API)
config.json	Architecture + fine-tune metadata
vocab.json	Token-to-ID mapping (2,566 tokens)
cmvn.json	Mel-spectrogram normalization stats
preprocessor_config.json	Feature extraction parameters
example.mp3	Verification sample audio
requirements.txt	Python runtime dependencies
finetune_google_openslr80.ipynb	Colab notebook for reproducing this fine-tune

Credits

• Base model: freococo/myanmar_asr
• Fine-tune dataset: freococo/google_myanmar_asr_voices
• Original corpus: OpenSLR-80 — Burmese Speech Corpus (Google, LREC 2020)
• Architecture & Training: freococo
• AI Engineering Collaboration: DeepSeek AI & Gemini AI

Citation

@misc{myanmar_asr_google_openslr80_2026,
  title  = {myanmar_asr_google_openslr80: Fine-tuned Burmese ASR for OpenSLR-80},
  author = {freococo},
  year   = {2026},
  url    = {https://huggingface.co/freococo/myanmar_asr_google_openslr80}
}