Performance Analysis of Hybrid Automatic Continuous Speech Recognition Framework for Kannada Dialect

This paper demonstrates the execution investigation of the automatic continuous speech recognition system for Kannada language using hybrid modelling techniques. The well-known modelling techniques, in particular, deep neural system (DNN), hidden Markov display (HMM), subspace Gaussian mixture model (SGMM) and maximum mutual information (MMI) have been combined to form the hybrid modelling for speech recognition. The persistent Kannada speech information is gathered from 1600 speakers (960 males and 640 females) of the age bunch in the scope of 8 years-80 years. The speech information is acquired from different geographical regions of the Karnataka state under certifiable condition. It comprises of 20,000 words that spread 30 locale. The point of this paper is to examine the execution of hybrid modelling techniques with regards to Kannada speech recognition. Kaldi toolbox is utilized for the implementation of this system, in which Mel frequency cepstral coefficient (MFCC) is used as a feature extraction procedure. The word error rate (WRR) is the error metric used to determine the efficiency of the automatic speech recognition (ASR) system. The experimental results demonstrate that the recognition rate got through the combination of DNN and HMM is better over other hybrid ASR modelling strategies.