Table III.

Summary of the significant state-of-the-art DNN speech recognition models

Architecture	Dataset	Error rate
RNN [126] - FIT, Czech Republic, Johns Hopkins University, 2011	Penn Corpus (natural language modeling)	123*
Autoencoder/DBN [127] - Collaboration, 2012	English Broadcast News Speech Corpora (spoken word recognition)	15.5%**
LSTM [129] - Google, 2014	Google Voice Search Task (spoken word recognition)	10.7%**
Deep LSTM [130] - National Chiao Tung University, 2016	CHiME 3 Challenge (spoken word recognition)	8.1%**
CNN-BLSTM [131] - Microsoft, 2017	Switchboard (spoken word recognition)	5.1%
Attention (LAS) & LSTM [132] - Google, 2018	In-house google dictation (spoken word recognition)	4.1%
Attention & LSTM with pretraining [133] - Collaboration, 2018	LibriSpeech (spoken word recognition)	3.54%

(*perpeplexity-size of model needed for optimal next word prediction with 10K classes, **word error rate)