GMM based Bayesian approach to speech enhancement in signal / transform domain

An iterative longest matching segment approach to speech enhancement with additive noise and channel distortion

“…23 The remainder of this paper is organized as follows. Section 2 outlines the assumptions made 24 in this research for modeling noisy speech, and the key idea of the proposed approach for speech 25 estimation. Section 3 introduces the first part of the proposed algorithm, including corpus-based 26 modeling of speech with noise and channel distortion, and the longest matching segment algorithm 27 for speech estimation.…”

“…Suppose Fig. 1 shows, on the top, the noisy signal 24 PSD y k,t for a specific frequency bin k sampled at consecutive discrete frame times t, consisting of 25 the clean signal PSD x k,t and some unknown noise PSD n k,t . Below the noisy signal, Fig.…”

“…For speech enhancement, while Wiener filtering based on the matched corpus speech frame, as 23 described in (Ming et al, 2011), can be used to suppress the additive noise, it is not effective for 24 recovering speech from channel distortion. Therefore, in our speech enhancement experiments with 25 the Aurora 4 database, we reconstruct the waveform for each underlying speech frame by using 26 the magnitude spectrum of the matched corpus speech frame.…”

“…Data-driven speech models, built on the training data of real speech, represent a different 33 way of imposing prior or constraint on the speech to be estimated. Common speech models include 34 vector-quantization (VQ) codebooks (e.g., Naidu and Srinivasan, 2012;Srinivasan et al, 2006), 35 Gaussian mixture models (GMM) (e.g., Kundu et al, 2008), hidden Markov models (HMM) (e.g., 36 Ephraim et al, 1989; Sameti and Deng, 2002;Zhao and Kleijn, 2007), and inventory-based models 37 terms of improved speech recognition and speech enhancement performance.…”

“…As a result, they require specific knowledge about the noise 23 for noise removal and hence speech recovery. The typical constraint or prior for the underlying 24 speech is the probability distribution of the speech short-time discrete Fourier transform (DFT) 25 coefficients or spectral amplitudes (e.g., Cohen, 2005;Ephraim and Malah, 1984;Lotter and Vary, 26 2005; Martin, 2002;Martin and Breithaupt, 2003). The common methods for noise estimation 27 include prediction by using neighboring measurements without significant speech content based on 28 voice activity detection, minimum statistics, time-recursive averaging, MMSE-based high-resolution 29 noise DFT estimation and their combination (e.g., Cohen Sohn and Kim, 1999).…”