Machine Learning and AI with Simple Python and Matlab Scripts

9CNNs for Speech Recognition

9.1 Introduction

In this chapter, we introduce a simple CNN implementation for speech recognition, which is the ability for a computer program to recognize speech or recorded audio in any format. In typical applications, recognized and identified speech will most likely be converted to another format such as text or a digital record, thereby significantly increasing the influence of speech recognition in the daily usage of various intelligent systems. Many modern devices have built‐in speech recognition capabilities through virtual assistant applications such as Alexa from Amazon and Siri from Apple [81]. Other popular speech recognition applications include customer service systems used by many commercial entities ranging from car rental companies to credit card services, automatic generation of subtitles, machine translation tasks, assistance for the disabled and background noise elimination in loud environments (e.g. noise cancellation devices used in helicopters or submarine vessels).

A spectrogram [82] is a diagram that plots the spectrum of frequencies in an audio sample over time. The spectrogram is calculated by using the Fourier transform to identify the frequencies present in an audio clip in a given time window [83]. The pixel width of a spectrogram represents the number of time divisions in an audio file. For each time division, the Fourier transform is calculated, producing a column of pixels that show the spectrum of frequencies during ...