deepsphinx

deepsphinx copied to clipboard

Introduction

This repo contains an end-to-end speech recognition system implented using tensorflow. It uses bidirectional LSTMs in the encoding layer with attention in the decoding layer. The LSTMs in the encoding layer are strided and in every layer, the time dimension is reduced by 2.

Model

Default configuration contains 3 bidirectional layers for encoding and decoding layer:

Current results

I am getting round 15% WER in dev93 by training in si284 set.