Whisper-Base-ASR / README.md
myandev's picture thantzinphyo's picture
Duplicate from thantzinphyo/Whisper-Base-ASR
de9f257
|
Raw History Blame Contribute Delete
4.13 kB
metadata
language:
  - my
license: apache-2.0
tags:
  - audio
  - automatic-speech-recognition
  - speech
  - whisper
  - burmese
datasets:
  - thantzinphyo/Burmese-Daily-Dialogue-Corpus
metrics:
  - cer
  - wer
  - chrf
pipeline_tag: automatic-speech-recognition
base_model:
  - openai/whisper-base
model-index:
  - name: Whisper-Base-ASR
    results:
      - task:
          type: automatic-speech-recognition
          name: Automatic Speech Recognition
        dataset:
          name: Burmese Daily Dialogue Corpus (BDDC)
          type: thantzinphyo/Burmese-Daily-Dialogue-Corpus
        metrics:
          - name: Validation CER
            type: cer
            value: 4.43
          - name: Validation WER
            type: wer
            value: 26.55
          - name: Validation chrF
            type: chrf
            value: 91.94
          - name: Test CER (Unseen Speakers)
            type: cer
            value: 8.55
          - name: Test WER (Unseen Speakers)
            type: wer
            value: 36.29
          - name: Test chrF (Unseen Speakers)
            type: chrf
            value: 83.71

Whisper-Base-ASR ( Burmese )

This model is a fine-tuned version of openai/whisper-base for Automatic Speech Recognition (ASR) in Burmese .

Model Overview

  • Architecture: Whisper (Base)
  • Parameters: ~74M
  • Task: Automatic Speech Recognition (ASR)
  • Language: Burmese (my)
  • Sampling Rate: 16,000 Hz (Mono)

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: แ€•แ€ฎแ€š แŠ แ€แ€ซแ€…แ€ฌ แŠ แ€žแ€ฎแ€›แ€ญ แŠ แ€’แ€ฎแ€• แŠ แ€กแ€€แ€นแ€แ€›แ€ฌ แŠ แ€žแ€’แ€นแ€’แ€ซ แŠ แ€žแ€› แŠ แ€แ€ฌแ€›แ€ฌ แŠ แ€€แ€แ€ญ แŠ แ€žแ€ฏแ€ แŠ แ€•แ€žแ€ฌแ€’ แŠ แ€‚แ€ฎแ€ แŠ แ€”แ€”แ€นแ€’
  • Training & Validation: 11 speakers (20,699 train utterances, 2,300 validation utterances)
  • Test Set (Held-Out): 2 unseen speakers ( แ€‚แ€ฎแ€ & แ€”แ€”แ€นแ€’ , 1,561 utterances )
  • Text Preprocessing: Standardized Myanmar Unicode (\u1000-\u109F) 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 (16 per device x 4 gradient accumulation)
Peak Learning Rate 1.0e-4
Learning Rate Scheduler Cosine
Warmup Steps 200
Total Steps 2,000 (~12.37 Epochs)
Mixed Precision BF16
Weight Decay 0.01
Gradient Checkpointing Enabled
Augmentation None

Evaluation Results

Stage / Evaluation Step Train Loss Val Loss WER (%) CER (%) SER (%) DER (%) IER (%) chrF
Baseline (Untrained) 0 - - 753.85 437.94 100.00 7.69 653.85 0.00
Validation Set 250 0.6374 0.0725 47.99 10.33 93.57 10.36 2.65 81.45
Validation Set 500 0.2351 0.0453 37.15 6.92 87.52 6.12 2.81 86.66
Validation Set 750 0.1369 0.0417 31.71 5.70 81.61 3.60 4.19 89.08
Validation Set 1000 0.0482 0.0416 28.87 5.00 76.74 4.22 3.29 91.07
Validation Set 1250 0.0252 0.0435 28.18 4.79 76.74 3.72 3.31 91.23
Validation Set 1500 0.0039 0.0476 28.06 4.69 78.48 3.81 2.94 91.40
Validation Set 1750 0.0007 0.0502 26.91 4.47 76.30 3.26 3.44 91.79
Validation Set 2000 0.0004 0.0507 26.55 4.43 75.65 3.31 3.37 91.94
UNSEEN TEST (Final) Final - - 36.29 8.55 94.30 5.40 3.22 83.71

Usage

import torch
from transformers import pipeline

pipe = pipeline(
    'automatic-speech-recognition',
    model='thantzinphyo/Whisper-Base-ASR',
    device='cuda:0' if torch.cuda.is_available() else 'cpu',
)

output = pipe('audio.wav', generate_kwargs={'language': 'burmese', 'task': 'transcribe'})
print(output['text'])

License

Apache-2.0