Improved noise robust automatic speech recognition system with spectral subtraction and minimum statistics algorithm implemented in FPGA
John William Orillo, Roderick Yap, Edwin Sybingco · 2012
In this study, spectral subtraction speech enhancement is integrated to a two word vocabulary speech recognition system to effectively reduce the effects of background noise and increase the recognition rate. The whole system was implemented in FPGA and was modelled in MATLAB. The preprocessing subsystem contains the spectral subtraction algorithm and acoustic front end speech enhancements while the speech recognition subsystem contains the HMM and Viterbi search algorithms. 10 dirty speech samples of word `stop' and `clockwise' (sampled at 84 dB) were tested in the speech recognition prototype with varying background noise from 44.6 to 85.4 dB and noise floor (β) from 0.01 to 1. At the end of the testing, the system was able to recognize the two words (stop and clockwise) efficiently with accuracy rate of above 80% until a background noise of 68.6 dB. The best average recognition rate (from 44.6 to 85.4 dB background noise) of 48.5% on the other hand was recorded at 0.01 noise floor. The system without spectral subtraction enhancement was noticed to function efficiently only at 56.6 dB.