Meta-MMS-300M-ASR ( Burmese )
This model is a fine-tuned version of facebook/mms-300m for Automatic Speech Recognition (ASR) in Burmese .
Model Overview
- Architecture: Meta MMS (300M)
- Parameters: ~317M
- Task: Automatic Speech Recognition (ASR)
- Language: Burmese (
my) - Sampling Rate: 16,000 Hz (Mono)
- Tokenizer / Vocab: 64 Burmese Character Tokens (CTC Head)
Dataset Details
The model was trained on a standardized, high-quality Burmese speech corpus:
- Total Duration: ~22 Hours of audio
- Total Utterances: 24,560 WAV files (16 kHz, 16-bit PCM, Mono)
- Total Speakers: 13 Synthetic Burmese Speakers
- Speakers: ပီယ ၊ ဝါစာ ၊ သီရိ ၊ ဒီပ ၊ အက္ခရာ ၊ သဒ္ဒါ ၊ သရ ၊ တာရာ ၊ ကဝိ ၊ သုတ ၊ ပသာဒ ၊ ဂီတ ၊ နန္ဒ
- Training & Validation: 11 speakers (20,699 train utterances, 2,300 validation utterances)
- Test Set (Held-Out): 2 unseen speakers ( ဂီတ & နန္ဒ , 1,561 utterances )
- Text Preprocessing: Standardized Myanmar Unicode with unified word segmentation and punctuation removal. The held-out test split evaluates zero-shot acoustic generalization across completely unseen voices.
Training Configuration
| Parameter | Value |
|---|---|
| Effective Batch Size | 64 (8 per device x 4 gradient accumulation) |
| Base Architecture | facebook/mms-300m ( Wav2Vec 2.0 with 24 Transformer layers, 1024 hidden size, 16 attention heads ) |
| Acoustic Feature Extractor | 16 kHz 1D raw waveform input (CNN encoder frozen during training) |
| Tokenization | Burmese Unicode character CTC vocabulary |
| Peak Learning Rate | 3.0e-4 |
| Learning Rate Scheduler | Linear |
| Warmup Steps | 200 |
| Total Steps | 2,000 (~6 Epochs) |
| Gradient Checkpointing | Enabled (Input Require Grads) |
| Augmentation | None |
Evaluation Results
| Stage / Evaluation | Step | Train Loss | Val Loss | WER (%) | CER (%) | SER (%) | DER (%) | IER (%) | chrF |
|---|---|---|---|---|---|---|---|---|---|
| Baseline (Untrained) | 0 | - | - | 100.00 | 121.74 | 100.00 | 84.62 | 0.00 | 1.30 |
| Validation Set | 250 | 6.4141 | 1.0876 | 67.66 | 14.87 | 99.17 | 10.48 | 4.22 | 69.03 |
| Validation Set | 500 | 1.5876 | 0.4096 | 37.75 | 6.34 | 85.74 | 6.62 | 2.86 | 87.67 |
| Validation Set | 750 | 0.8111 | 0.2971 | 28.93 | 4.44 | 74.65 | 3.76 | 4.08 | 91.69 |
| Validation Set | 1000 | 0.5157 | 0.2717 | 26.27 | 3.84 | 70.91 | 3.76 | 3.59 | 93.22 |
| Validation Set | 1250 | 0.3567 | 0.2459 | 24.87 | 3.56 | 67.22 | 4.20 | 2.94 | 94.06 |
| Validation Set | 1500 | 0.1711 | 0.2656 | 23.40 | 3.25 | 64.09 | 2.83 | 4.11 | 94.56 |
| Validation Set | 1750 | 0.0564 | 0.2951 | 21.23 | 2.93 | 60.22 | 3.20 | 3.34 | 95.22 |
| Validation Set | 2000 | 0.0208 | 0.2972 | 20.66 | 2.85 | 59.35 | 3.12 | 3.15 | 95.39 |
| UNSEEN TEST (Final) | Final | - | - | 32.64 | 6.66 | 90.01 | 5.33 | 2.60 | 87.22 |
Usage
import torch
from transformers import pipeline
pipe = pipeline(
'automatic-speech-recognition',
model='thantzinphyo/Meta-MMS-300M-ASR',
device='cuda:0' if torch.cuda.is_available() else 'cpu',
)
output = pipe('audio.wav')
print(output['text'])
References & Citations
- Meta MMS Paper: Scaling Speech Technology to 1,000+ Languages (Pratap et al., 2023)
- Wav2Vec 2.0 Paper: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al., 2020)