DSP seminar

DATE: Tuesday, February 3
TIME: 4 p.m.
LOCATION: Beckman Room 5602

SPEAKER: Jun Huang

TITLE: Robust Speech Recognition in White Noise and Colored Noise

                                ABSTRACT

Automatic speech recognition in noisy environment is an active research
subject nowadays. This talk will first review some conventional robust speech
recognition methods such as feature transformation and model
modification. Then an energy-constrained signal subspace (ECSS)
method is proposed for speech enhancement and recognition in additive
noise. The key idea is to decompose the vector space of a noisy signal
into a pure noise subspace and a signal-plus-noise subspace. An energy
constraint is introduced to improve the estimation of noise-like speech
segments. Colored noise is modelled by an autoregressive (AR) process
A prewhitening filter is constructed from the estimated AR parameters using
modified covariance method. Experimental results on white and colored noises
shows that this method can improve the SNR by 2 dB ~ 6 dB and the word
recognition acuracy (WRA) by 13.7% ~ 55.9% under various SNR conditions.
Possibilities of future improvement will also be briefly discussed.