Analysis and Optimization of Telephone Speech Command Recognition System Performance in Noisy Environment

Abstract

This paper deals with the analysis and optimization of a speech command recognition system (SCRS) trained on Czech telephone database Speechdat(E) for use in a selected noisy environment. The SCRS is based on hidden Markov models of context dependent phones (triphones) and mel-frequency cepstral coefficients analysis of speech (MFCC). The main aim is to analyze and to search for the optimal settings of SCRS with respect to additive noise robustness without use of additional techniques for additive noise reduction. The analysis is pointed to the appropriate setting of MFCC computation, the silence model adjustment and grammar selection possibilities. It is shown, that the correct performance of SCRS strictly depends on an appropriate adjustment of the silence model. The ability of the silence model adaptation is confirmed. When SNR is higher than 15 dB the suitable performance of SCRS can be guarantied without any modification of the triphones speech models by: 1. the optimal setting of MFCC computation, 2. the proper silence model adaptation. The assumption of a speech command recognition system use in an environment where SNR is higher than 15 dB is fulfilled in many applications.