YoungloLee/tf2-speech-recognition-transformer

Tensorflow 2 Speech Recognition Code (Transformer)

tf2-speech-recognition-transformer

Tensorflow 2 Speech Recognition Code (Transformer)

Developers

• Younglo Lee (yllee@ispl.korea.ac.kr)

Contents

• Contents
• Speech Recognition
• Features
• Prerequisites
• Examples
• References

Speech Recognition

• Tensorflow 2 implementation for korean speech recognition.
• Dataset can be downloaded in http://www.aihub.or.kr/aidata/105/download and https://github.com/clovaai/ClovaCall

Features

• Transformer speech recognition model
• Stacked frames with overlap (a.k.a. Lower Frame Rate, refer to https://arxiv.org/abs/2002.02562)
• Korean syllable tokenization

Prerequisites

• Python 3.x
• Tensorflow 2.2.0
• Librosa 0.6.3
• etc.

Examples

• Tensorboard examples (curriculum learning with increasing data length from 500 -> 1000 -> all)

- Attention weight examples (4 heads)

References and Resources

• https://github.com/Rayhane-mamah/Tacotron-2
• https://github.com/clovaai/ClovaCall
• https://github.com/sooftware/KoSpeech (for train/test split csv)
• https://www.tensorflow.org/tutorials/text/transformer