Speech recognition-a time-frequency subspace filtering based approach

In conventional speech recognition systems, the speech waveform is represented by a time sequence of linear time invariant (LTI) filters, typical filter descriptions including cepstral, LPC, and filterbank coefficients. The authors investigate an alternative approach, in which a conventional LTI filter sequence is replaced by a single time-frequency subspace filter. It is shown that the speech signal is typically concentrated within a low dimensional time-frequency subspace and that the location of this subspace can be used to discriminate between different speech sounds