Quantization of automatic speech recognition networks

Abstract

Recently, end-to-end neural networks based speech recognition systems have received great interests in the speech community. Although these systems offer a wide range of advantages such as high performance, they often rely on a neural network with a large number of parameters, i.e. large memory footprint and large decoding time. The main goal of this thesis is investigation of quantized neural networks for end-to-end speech recognition systems. Different quantization methods like post training quantization, scalar quantization, iterative product quantization and quantization aware training are investigated in this thesis on ASR model. By doing so, the final system is 3.7 times smaller than the baseline system, faster in decoding with a speedup factor of 1.19 and with almost same level of recognition performance.