Perceptual Factor Analysis for Speech Enhancement

1 Perceptual Factor Analysis for Speech Enhancement Chuan-Wei Ting and Jen-Tzung Chien Department of Computer Science and Information Engineering, National Cheng Kung University, Tainan, Taiwan, ROC {motor,chien}@chien.csie.ncku.edu.tw Abstract This paper presents a new speech enhancement approach originated from factor analysis (FA) framework. FA is a data analysis model where the relevant common factors can be extracted from observations. A factor loading matrix is found and a resulting model error is introduced for each observation. Interestingly, FA is a subspace approach properly representing the noisy speech. This approach partitions the space of noisy speech into a principal subspace containing clean speech and a complimentary (minor) subspace containing the residual speech and noise. We show that FA is a generalized data model compared to signal subspace approach. To perform FA speech enhancement, we present a perceptual optimization procedure that minimizes the signal distortion subject to the energies of residual speech and noise under a specified level. Importantly, we present a hypothesis testing approach to optimally perform subspace decomposition. In the experiments, we implement perceptual FA speech enhancement using Aurora2 corpus. We find that proposed approach achieves desirable speech recognition rates especially when signal-to-noise ratio is lower than 5 dB. 1. Introduction Automatic speech recognition (ASR) systems have been employed to many real-world applications. However, ASR systems are always degraded in presence of different noises in practical situations. To provide good speech quality for ASR systems, the speech enhancement is an important preprocessing procedure for noisy speech recognition. In the past decade, the researchers on speech enhancement for robust ASR have been attracting many people working on this issue. Spectral subtraction algorithm [2] is one of the most popular methods for speech enhancement. This algorithm has the drawbacks of producing speech distortion and “musical noise”. The method in [11] was proposed to overcome “musical noise” problem by using human auditory models where the perceptual effect of “musical noise” was reduced under predefined threshold. Below the masking threshold, the residual noise becomes inaudible by human ear. Other researchers presented subspace approaches to balance the trade off between speech distortion and residual noise [5] [8]. The general concept of subspace approaches is originated from that the noisy speech signal can be projected onto two subspaces; one is the signal subspace in which clean speech signal and few noises are included, and the other is the noise subspace that only contains noise information. In [4], Ephraim and Van Trees proposed signal subspace approach to find optimal estimator or filter by 2 minimizing the speech distortion subject to the constraint of residual noise kept under a threshold. This work decomposed the noisy signal into signal subspace and noise subspace by using Karhunen-Loève transform (KLT). The noisy speech signal was accordingly enhanced by using inverse KLT. Rezayee and Gazor [8] used a diagonal matrix instead of the identity matrix for finding the linear time domain constrained estimator of clean speech. Hu and Loizou [5] estimated the optimal filter by using common matrix diagonalizing the covariance matrices of the …

Content truncated to 500 words