Non-linear speech analysis using recurrent radial basis function networks

摘要

This paper presents a recurrent radial basis function network as a one step ahead predictive speech signal filter. The resulting non-linear estimation of the signal state space allows accurate prediction using only three delayed samples of clean speech and in noisy speech six samples allow this performance to be maintained. The prediction residual can be used as a powerful speech pitch detector and the nonlinear network shows significant improvement over conventional auto-regressive filters, allowing post-processors to make more accurate estimations of pitch pulse position, the pitch, and the regions of voiced speech. This represents a new form of preprocessing for pitch tracking of real speech in a noisy environment.