README.md

metadata

language:
  - my
license: apache-2.0
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - wav2vec2
  - wav2vec2-xls-r
  - burmese
datasets:
  - thantzinphyo/Burmese-Daily-Dialogue-Corpus
metrics:
  - cer
  - wer
  - chrf
pipeline_tag: automatic-speech-recognition
model-index:
  - name: Wav2Vec2-XLS-R-300M-ASR
    results:
      - task:
          type: automatic-speech-recognition
          name: Automatic Speech Recognition
        metrics:
          - name: Validation CER
            type: cer
            value: 3.53
          - name: Validation WER
            type: wer
            value: 24.25
          - name: Validation chrF
            type: chrf
            value: 94.97
          - name: Test CER (Unseen Speakers)
            type: cer
            value: 6.58
          - name: Test WER (Unseen Speakers)
            type: wer
            value: 33.24
          - name: Test chrF (Unseen Speakers)
            type: chrf
            value: 87.69
        dataset:
          name: Burmese Daily Dialogue Corpus (BDDC)
          type: thantzinphyo/Burmese-Daily-Dialogue-Corpus
base_model:
  - facebook/wav2vec2-xls-r-300m

Wav2Vec2-XLS-R-300M-ASR ( Burmese )

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m for Automatic Speech Recognition (ASR) in Burmese .

Model Overview

• Architecture: Wav2Vec2 XLS-R (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	32 (4 per device x 8 gradient accumulation)
Peak Learning Rate	3.0e-4
Learning Rate Scheduler	Linear
Warmup Steps	200
Total Steps	2,000 (~6 Epochs)
Mixed Precision	BF16
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	113.39	100.00	85.06	0.00	0.55
Validation Set	250	13.0478	1.1601	72.71	17.40	99.78	13.98	3.03	66.26
Validation Set	500	3.2769	0.4160	39.69	7.33	87.70	6.37	3.16	87.39
Validation Set	750	1.6812	0.3104	33.10	5.68	83.43	3.66	4.38	91.12
Validation Set	1000	1.1205	0.2926	29.82	4.68	78.35	5.50	2.43	92.75
Validation Set	1250	0.7550	0.2558	27.96	4.08	75.09	5.15	2.70	93.58
Validation Set	1500	0.3550	0.2688	25.96	3.81	71.17	4.16	3.21	94.43
Validation Set	1750	0.1300	0.3087	25.18	3.67	70.57	4.77	2.47	94.70
Validation Set	2000	0.0465	0.3125	24.25	3.53	69.13	3.77	3.07	94.97
UNSEEN TEST (Final)	Final	-	-	33.24	6.58	91.35	5.87	2.57	87.69

Usage

import torch
from transformers import pipeline

pipe = pipeline(
    'automatic-speech-recognition',
    model='thantzinphyo/Wav2Vec2-XLS-R-300M-ASR',
    device='cuda:0' if torch.cuda.is_available() else 'cpu',
)

output = pipe('audio.wav')
print(output['text'])