Unsupervised spectral subtraction for noise-robust ASR

Abstract

This paper proposes a simple, computationally efficient 2-mixture model approach to discriminate between speech and background noise at the magnitude spectrogram level. It is directly derived from observations on real data, and can be used in a fully unsupervised manner, with the EM algorithm. In this paper, the 2-mixture model is used in an "Unsupervised Spectral Subtraction" scheme that can be applied as a pre-processing step for any acoustic feature extraction scheme, such as MFCCs or PLP. The goal is to improve noise-robustness of the acoustic features. Experimental results on both OGI Numbers 95 and Aurora 2 tasks yielded a major improvement on all noise conditions, while retaining a similar performance on clean conditions.

Switching Linear Dynamical Systems for Noise Robust Speech Recognition

“…Each test utterance was corrupted with additive noise independently sampled from a Gaussian with zero mean and covariance σ 2 . The features for the HMM were computed using Unsupervised Spectral Subtraction (USS) [17], thereby providing filtered features to the HMM recogniser. The setup used for the feature-based HMM was the same as that used to obtain the baseline performance on the AURORA task [13], namely 18 states, left-right transition matrix, a mixture of three Gaussians per state and 39 MFCC features, including first and second temporal derivatives as well as energy.…”

“…Whilst successful under controlled conditions, this standard approach is often particularly fragile in the presence of noise [17]. This important issue is commonly addressed by a preprocessing step which attempts to remove noise, see for example [17], [7], [12], [25] and [1].…”

A Fast Parts-Based Approach to Speaker Verification Using Boosted Slice Classifiers

“…Since the proposed approach involves working with specific frequency points in the spectrum, it might be directly coupled with a suitable time-frequency masking framework aimed at noise removal [32] or signal separation [33]. Finally, since the approach is data-driven, it could be applied to other related tasks like phoneme recognition.…”

SNR features for automatic speech recognition

“…Ris and Dupont [5] present a survey of methods to measure noise, favouring the low-energy envelope tracking approach of Martin [6]. Lathoud et al [7] present a statistical spectral model that yields both noise and speech estimates.…”

“…For instance, histogram normalisation, a logical progression of CMN/CVN to higher order moments, has been successfully combined with spectral compensation techniques by Segura et al [8]. Lathoud et al [7], who describe their technique as "Unsupervised" spectral subtraction (USS), also report good results in combination with cepstral normalisation.…”