myanmar_asr_classical

Classical Burmese (Myanmar) speech recognition model — 4-layer Bidirectional LSTM with projection + CTC.

16.7M parameters. 10.51% val SER / 14.05% newdata SER.

Part of the freococo Burmese ASR family, alongside the modern Conformer model.

Why classical

Not every deployment needs a Conformer. This model is:

• Smaller — 16.7M vs 23.8M parameters
• Simpler — pure BLSTM + CTC, no attention, no transformer decoder
• Proven — classical RNN architecture, battle-tested for low-resource ASR
• Comparable — 14.05% vs 12.30% on the same 19k-utterance independent crawl

Roughly 1.7 points behind the Conformer on harder data — a fair trade-off for a lighter, more interpretable model.

Comparison to the modern model

Model	Arch	Params	Val SER	Newdata SER
myanmar_asr (modern)	12-layer Conformer + joint CTC/Attn	23.8M	9.42%	12.30%
myanmar_asr_classical	4-layer BLSTMP + CTC	16.7M	10.51%	14.05%

Same training corpus (1.66M utterances), same schedule (25 epochs, cosine, peak LR 6e-4), same tokenizer, same evaluation protocol. Direct architecture comparison with no confounds.

Per-file analysis on the 19k test set:

• Both models transcribe 32.4% of utterances exactly
• The modern model is uniquely perfect on 8.5%
• The classical model is uniquely perfect on 6.4%
• Per-file SER correlation: 0.939 (they fail on the same chunks)

Architecture

Audio (16 kHz mono)
  → Log-mel spectrogram (80 bins, 400/160)
  → CMVN
  → 2× (Conv2d ×2 + BatchNorm + ReLU) + MaxPool2d  → 4× subsampling
  → 4 × (BiLSTM(hidden=320) + Linear(640→512) + LayerNorm + residual)
  → Linear(512 → 2566)
  → CTC log-softmax

Parameters	16,676,038
Frontend	Conv2d ×4 + MaxPool ×2
Encoder	4-layer BLSTMP (hidden 320, proj 512)
Output head	CTC
Vocabulary	2,566 syllable tokens
Input	16 kHz mono, 80-dim log-mel
Output	Space-separated Burmese syllables

Training

Corpus	~1.66M utterances
Epochs	25
Optimizer steps	178,325
Optimizer	AdamW, cosine decay, warmup 500
Peak LR	6e-4
Batch	180 s / 90 clips per micro-step, grad_accum=4
Hardware	1× NVIDIA RTX 3060 (12 GB)
Wall time	~40 hours

Usage

Install

pip install -r requirements.txt

CLI

python transcribe.py input.wav
python transcribe.py input.wav --device cuda
python transcribe.py *.wav --output results.txt

Python API

from transcribe import BurmeseASR

asr = BurmeseASR(device="cuda")
text = asr.transcribe("input.wav")
print(text)

Example

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

Metric

Syllable Error Rate (SER). Syllables are the natural phonological unit of Burmese. WER requires word segmentation (non-canonical for Burmese); CER is biased by variable code-point counts per syllable. SER gives a consistent, well-defined error rate.

Files

File	Purpose
model.safetensors	Model weights (67 MB)
model.py	Standalone FastASR + BLSTMP architecture
transcribe.py	Inference script (CLI + Python API)
config.json	Architecture + training spec
vocab.json	Token → id mapping
cmvn.json	Mel-spectrogram mean / std
preprocessor_config.json	Feature extraction parameters
example.mp3	Sample input
requirements.txt	Python dependencies

Model family

Repo	Arch	SER (newdata)
myanmar_asr	Conformer	12.30%
myanmar_asr_google_openslr80	Conformer + fine-tune	—
myanmar_asr_classical (this)	BLSTMP	14.05%

Dedication & Acknowledgements

AI Engineering Partners

• DeepSeek AI: Primary co-engineering partner through every stage.
• Gemini AI (Google): Secondary review partner.

Data Contributors

• National Unity Government (NUG) of Myanmar
• PVTV, MRTV, MRTV-4 broadcast archives
• Myanmar Celebrity Voices, Media Queen Entertainment
• Sunday Journal, Khit Thit Media news teams
• Google Myanmar ASR corpus, Myanmar Bible Speech corpus
• Original Myanmar Voices contributors
• Independent Myanmar journalists, broadcasters, and creators

Statement of Independence

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

All training data is publicly available. Weights, code, and configs are released under Apache 2.0 without restriction.

Any incorrect or unexpected output is the sole responsibility of freococo — not the original creators of the source data. All credit for the data belongs to them. All errors in the model belong to me.

Citation

@misc{myanmar_asr_classical_2026,
  title  = {myanmar_asr_classical: Classical BLSTMP-CTC Burmese ASR},
  author = {freococo},
  year   = {2026},
  url    = {https://huggingface.co/freococo/myanmar_asr_classical}
}