First International Conference on Spoken Language Processing (ICSLP 90)
Surrounding noise seriously affects the performance of speech recognition which can operate under usual noise environment. When we try to make an audio equipment or television set which can be controlled by speech, the sound radiated by the equipment itself also affects the performance of speech recognition. In this study, we propose the method to reduce the latter type noise, a priori known noise. The exponential step normalized LMS algorithm is used for the adaptation of FIR digital filter. Also the method of speech candidate detection is presented using the characteristics of adaptive filter. The experiments are carried out under usual room condition with the several broadcasted sound as noise. The total reduction of noise varies from 15dB to 25dB due to the difference of noise type. There is the correlation between the spectrum of noise and the obtained filter coefficients on adaptive filter.
Bibliographic reference. Usagawa, Tsuyoshi / Morita, Yuji / Ebata, Masanao (1990): "Remote control system using speech-reduction of known noise", In ICSLP-1990, 289-292.