The IBM 2016 English Conversational Telephone Speech Recognition System

TL;DR

This paper presents an advanced English conversational telephone speech recognition system that achieves a new record low word error rate of 6.6% on the Switchboard dataset by combining multiple acoustic models and sophisticated language models.

Contribution

It introduces a novel combination of acoustic models including recurrent, convolutional, and LSTM networks, along with improved language modeling techniques, to significantly reduce word error rates.

Findings

Achieved a 6.6% word error rate on Switchboard

Demonstrated the effectiveness of model fusion and hierarchical neural LMs

Set a new benchmark for conversational speech recognition

Abstract

We describe a collection of acoustic and language modeling techniques that lowered the word error rate of our English conversational telephone LVCSR system to a record 6.6% on the Switchboard subset of the Hub5 2000 evaluation testset. On the acoustic side, we use a score fusion of three strong models: recurrent nets with maxout activations, very deep convolutional nets with 3x3 kernels, and bidirectional long short-term memory nets which operate on FMLLR and i-vector features. On the language modeling side, we use an updated model "M" and hierarchical neural network LMs.