Skip to content

Latest commit

 

History

History
executable file
·
45 lines (28 loc) · 1.42 KB

README.md

File metadata and controls

executable file
·
45 lines (28 loc) · 1.42 KB
Raw

Conformer: Convolution-augmented Transformer for Speech Recognition

Reference: https://arxiv.org/abs/2005.08100

Conformer Architecture

Example Model YAML Config

Go to config.yml

Usage

Training, see python examples/conformer/train_*.py --help

Testing, see python examples/conformer/test_*.py --help

TFLite Conversion, see python examples/conformer/tflite_*.py --help

Conformer Subwords - Results on LibriSpeech

Summary

  • Number of subwords: 1031
  • Maxium length of a subword: 4
  • Subwords corpus: all training sets, dev sets and test-clean
  • Number of parameters: 10,341,639
  • Positional Encoding Type: sinusoid concatenation
  • Trained on: 4 RTX 2080Ti 11G

Pretrained and Config, go to drive

Transducer Loss

conformer_subword

Error Rates

Test-clean Test batch size WER (%) CER (%)
Greedy 1 6.37933683 2.4757576
Greedy V2 1 7.86670732 2.82563138
Test-other Test batch size WER (%) CER (%)
Greedy 1 15.7308521 7.67273521