Listen, Attend and Spell - Githubissues

kweonwooj / papers

summary of ML papers I've read

319 stars 34 forks source link

Listen, Attend and Spell #58

Open kweonwooj opened 6 years ago

kweonwooj commented 6 years ago

Abstract

Propose End-to-End Encoder-Decoder model (Listen, Attend and Spell (LAS)) for Speech Recognition task
Transcribing Speech Utterances to Characters Task
14.1% WER w/o dictionary or language model, and 10.3% with language model / SOTA CLDNN-HMM model achieves 8.0%

Details

Proposing novel approach of "Pyramidal RNN" to an area where HMM or DNN was majorly used
LAS Component
- Listener : pyramidal BiDi LSTM
- Attention
- Speller : 2 Layer LSTM + MLP + Softmax
Result
Attention Viz

Personal Thoughts

Classic example of RNN paper in speech recognition task in early days
Worth implementing for Kaggle Speech Recognition Task

Link : https://arxiv.org/pdf/1508.01211.pdf Authors : Chan et al. 2015