A Speech Recognition System Based on Dynamic Characterization of Background Noise

Robustness of automatic speech recognition (ASR) systems in realistic conditions of background noise is the essential conditions for their ample diffusion. As we know, the systems which exist at present suffer, however, a notable decrease in performance in the presence of background noise. In this article we propose an ASR system based on a dynamic characterization of background noise. In particular, the system makes a dynamic choice of HMM model related to the different types of noise and corresponding to different signal to noise ratios (SNR). The system was implemented and the tests performed using the AURORA2 database. The results were compared with a nonadaptive classification system in the presence of clean conditions and 4 different types of background noise and with 6 different SNRs. The proposed ASR system was found to be particularly adapted to applications functioning in extremely noisy contexts