Speech enhancement in vehicular environments as a front end for robust speech recogniser

摘要

The term “Speech Enhancement” refers to improving quality and intelligibility of the degraded speech. The current work focuses on three speech enhancement techniques. (1) Speech enhancement based subspace modelling using EVD (Eigen value decomposition). (2) Speech enhancement based subspace modelling using KLT (Karhunen Loeve transform). (3) Subspace modelling using KLT based on phoneme class analysis. Generally, subspace modelling involves decomposing noisy signal into two subspaces namely signal subspace (contains signal components + few noise components) and noise subspace (only noise components). In this regard, EVD based enhancement cancels noise subspace completely whereas, KLT based enhancement cancels both noise subspace and noise components in signal subspace. The phoneme based enhanced speech is obtained by suppressing the noise component at phone level. The comparison of these three enhancement techniques is evaluated for its quality using objective measures such as Perceptual Evaluation of Speech Quality (PESQ), for the speech affected by vehicular noises such as car, flight, train, white and babble, at -10 to 16 dB SNR levels in steps of 2 dB. This enhanced speech is then assessed using an Automatic Speech Recognition (ASR) system supported by hidden Markov model (HMM) technique.