Advances in Very Deep Convolutional Neural Networks for LVCSR

TL;DR

This paper introduces a new very deep CNN design for LVCSR that enables efficient sequence training and deployment, matching state-of-the-art performance on large datasets by removing timepadding and timepooling.

Contribution

The paper proposes a CNN architecture without timepadding and timepooling, facilitating efficient sequence evaluation and batch normalization, leading to improved large-scale speech recognition performance.

Findings

Achieved 9.4% WER on Hub5 test set with a single model.

Enabled efficient sequence training and deployment of CNNs.

Matched previous best results with a simpler, more scalable model.

Abstract

Very deep CNNs with small 3x3 kernels have recently been shown to achieve very strong performance as acoustic models in hybrid NN-HMM speech recognition systems. In this paper we investigate how to efficiently scale these models to larger datasets. Specifically, we address the design choice of pooling and padding along the time dimension which renders convolutional evaluation of sequences highly inefficient. We propose a new CNN design without timepadding and without timepooling, which is slightly suboptimal for accuracy, but has two significant advantages: it enables sequence training and deployment by allowing efficient convolutional evaluation of full utterances, and, it allows for batch normalization to be straightforwardly adopted to CNNs on sequence data. Through batch normalization, we recover the lost peformance from removing the time-pooling, while keeping the benefit of efficient convolutional evaluation. We demonstrate the performance of our models both on larger scale data than before, and after sequence training. Our very deep CNN model sequence trained on the 2000h switchboard dataset obtains 9.4 word error rate on the Hub5 test-set, matching with a single model the performance of the 2015 IBM system combination, which was the previous best published result.