A high-quality digital filterbank for speech recognition which runs in real time on a standard microprocessor

Asynchronous digital filterbank or LPC output is widely used for high-quality research and commercial speech recognition applications. Special-purpose hardware is usually applied. In this paper, a decimation/DFT filterbank system is described which may be implemented in real time on a standard microprocessor. Key features of the approach are a dithering technique, which allows output sampling at a rate lower than would be theoretically predicted as adequate, and the Winograd Fourier transform algorithm (WFTA) implementation of the DFT. The contributions of this paper are the dithering idea, the demonstration that it is feasible to implement a high-quality digital filter-bank on a standard microprocessor, and the discussion of starting-phase error. The nonlinear effects of dithering are discussed in detail, and it is shown that the dithering technique significantly reduces starting phase error, which can be important in very close discrete utterance recognition (DUR) situations.