Algorithms for keyword spotting with application to speech recognition

Abstract

The speech recognition system is very useful for the interaction between human and newlinemachine. Language is one of the barriers that create a hindrance to human to human newlineinteractions. In the scenario of arm conflict or natural disasters we need to communicate with newlinespeaker of less prevalent languages. Hence, it is very important and useful to develop a newlinespeech recognition system for low resource language like Gujarati. Various applications of newlinelocal language speech recognition are agriculture, automatic telephone system, voice newlineoperated services. The creation of language and acoustic re-sources, for any given spoken newlinelanguage, are typically a costly task. For example, a large amount of time and money is newlinerequired for the proper creation of annotated speech corpora for Automatic Speech newlineRecognition (ASR) and domain-specific text corpora for Language Modelling (LM). Speech newlinecorpora/corpus is database of speech audio files and text transcriptions of these audio files newlinein a format that can be used to create Acoustic Models. For proper working of the system, it newlineis required to identify the spoken words from the given speech inputs, i.e. Keyword spotting newlineplays a crucial role. In this thesis, our work focuses on in-ear microphone compared to newlineconventional microphone system to minimize the effects of background noise. In addition to newlinethat, we also implement endpoint detection algorithms and tested algorithms to separate the newlinekeywords from the silences and other unwanted noises. For feature extraction, we use Real newlineCepstral Coefficients (RC) and Mel Frequency Cepstral Coefficients (MFCC). We also newlineconfigured two and three layers of neural networks and tested for word recognition. For newlineGujarati speech database generation, various factors are considered such as, speakers of newlinevarious ages (e.g. Child, young, old), gender (e.g., Male, female), accent (kathiyawadi, newlinesortie, ahmedawadi). In future, our keyword spotting algorithm can be used, to drive a newlinerobotic arm hence the speech database has a vocabulary consisting of ten isolated Gujarati newlinewords