Abstract
An empirical comparison of three adaptive algorithms for speech enhancement in noisy reverberant conditions is presented. The subband least mean square (LMS) [E. Toner and D. R. Campbell, Speech Commun. 12, 253–259 (1993)] and minimum entropy noise reduction schemes [M. Girolami, Electron. Lett. 33(17), 1437–1438 (1997)] are compared with a reference wideband LMS approach [S. D. Stearns and R. A. David, Signal Processing Algorithms (Prentice–Hall, Englewood Cliffs, NJ, 1988)]. All three methods were applied to the enhancement of speech corrupted with speech-shaped noise. The schemes were tested using simulated anechoic and real-room reverberant (T60=0.3 s) environments. The anechoic and real-room recordings were binaural including head shadow and diffraction effects. The subband LMS adaptive processing scheme uses the LMS adaptive noise cancellation algorithm in frequency-limited subbands. The inputs from each microphone are split into 16 contiguous subbands using a cochlear distribution according the function provided by Greenwood [J. Acoust. Soc. Am. 87, 2592–2605 (1990)]. Each frequency-limited subband is processed using a LMS adaptive noise cancellation filter operating in an intermittent or continuous mode depending on input signal characteristics. The neural network approach is motivated by temporally sensitive hebbian super-synapses, which may form sparse representations based on commonly occurring spatio-temporal input patterns. The results for real-room signals demonstrate statistically equivalent performance of the subband LMS and wideband neural method, both of which prove superior to the wideband LMS algorithm.
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call