Robust Speech Recognition for Unknown Communication Channels Pont-à-Mousson, France April 17-18, 1997

ASR in Highly Non-Stationary Environments Using Adaptive Noise Cancelling Techniques

R. Martinez, A. Alvarez, V. Nieto, V. Rodellar, P. Gomez

Departamento de Arquitectura y Tecnologia de Sistemas Informaticos, Universidad Politdcnica de Madrid, Madrid, Spain

Through this paper the results of applying three different Adaptive Noise Cancelling Algorithms for Speech Enhancement in Robust Speech Recognition Command-Driven Interfaces are presented and discussed. Time-domain methods related with Recursive Least Squares Lattice-Ladder Filters with a posteriori updating show the best cancelling behaviour, although at a higher computational cost. The influence of microphonic separation is also evaluated. The best results are obtained for cardioid microphones with a separation of around 20 cm. taking advantage of their angular properties to prevent speech crosstalk in the reference microphone, thus reducing the computational cost required to acceptable limits.

Full Paper

Bibliographic reference.  Martinez, R. / Alvarez, A. / Nieto, V. / Rodellar, V. / Gomez, P. (1997): "ASR in highly non-stationary environments using adaptive noise cancelling techniques", In RSR-1997, 181-184.