Speech enhancement based on truncated and constrained minimum variance estimator (TCMVE) and undecimated wavelet packet non-uniform filterbanks

This paper presents a new method for speech enhancement using a truncated and constrained minimum variance estimator (TCMVE). The constraint allows for a tradeoff between residual noise and speech distortion. Mathematical analysis shows that Ephraim´s spectral domain constrained estimator (SDCE) and de Moor´s minimum variance estimator (MVE) can be treated as special cases of TCMVE. Furthermore, SDCE and MVE become equivalent for a particular choice of constraint. TCMVE was applied in multiple bandpass frequency regions using a new undecimated wavelet packet (UDWP) design to approximate a cochlear filterbank. The new algorithm was tested on noisy speech, which was then sent to an automatic speech recognizer (ASR). Both segmental SNR and recognition scores were improved over wavelet denoising for white, pink, and multi-speaker babble noise for many SNR.