Data-driven environmental compensation for speech recognition: A unified approach

Abstract

Environmental robustness for automatic speech recognition systems based on parameter modi®cation can be accomplished in two complementary ways. One approach is to modify the incoming features of environmentally-degraded speech to more closely resemble the features of the (normally undegraded) speech used to train the classi®er. The other approach is to modifying the internal statistical representations of speech features used by the classi®er to more closely resemble the features representing degraded speech in a particular target environment. This paper attempts to unify these two approaches to robust speech recognition by presenting several techniques that share the same basic assumptions and internal structure while diering in whether they modify the features of incoming speech or whether they modify the statistics of the classi®er itself. We present the multivaRiate gAussian-based cepsTral normaliZation (RATZ) family of algorithms which modify incoming cepstral features, along with the STAR (STAtistical Reestimation) family of algorithms, which modify the internal statistics of the classi®er. Both types of algorithms are data driven, in that they make use of a certain amount of adaptation data for learning compensation parameters. The algorithms were evaluated using the SPHINX-II speech recognition system on subsets of the Wall Street Journal database. While all algorithms demonstrated improved recognition accuracy compared to previous algorithms, the STAR family of algorithms tended to provide lower error rates than the RATZ family of algorithms as the SNR was decreased. Ó 1998 Elsevier Science B.V. All rights reserved. Re Âsume ÂDans les syst emes de reconnaissance de la parole, la robustesse a l'environnement par adaptation des param etres peutêtre obtenue de deux fac Ëons compl ementaires. Une premi ere approche consiste a modi®er les param etres acoustiques de la parole d egrad ee par l'environnement de fac Ëon a ressembler aux param etres de la parole (habituellement non d egrad ee) qui a et e utilis ee lors de l'entraõnement. La deuxi eme solution est de modi®er les param etres statistiques internes au reconnaisseur de fac Ëon a mieux repr esenter les caract eristiques de la parole d egrad ee dans un environnement cible particulier. Le pr esent papier tente d'uni®er ces deux approches de reconnaissance robuste de la parole en pr esentant plusieurs techniques qui partagent les mêmes hypoth eses de base et la même structure, tout en di erant dans le choix de savoir si elles modi®ent les param etres d'entr ee ou les param etres statistiques du reconnaisseur. Nous pr esentons ici la famille d'algorithmes bas es sur la normalisation cepstrale gaussienne multi-variable (RATZ) qui modi®ent les caract eristiques cepstrales d'entr ee, ainsi que les algorithmes STAR (re-estimation statistique), qui modi®ent les param etres internes du reconnaisseur. Les deux types d'algorithmes sont bas es sur les donn ees et utilisent une certaine quantit e de donn ee d'adaptation pour estimer les param etres de compen...

Speech Enhancement Based on Feature Reconstruction for Automatic Speech Recognition System with Unknown Structure

“…This represents one of the most commonly used techniques for additive noise suppression and removal of channel distortion respectively. We also evaluated a feature compensation method, Vector Taylor Series (VTS) [5] for performance comparison where the noisy speech GMM is adaptively estimated using the Expectation-Maximization (EM) algorithm over each test utterance [5]. The Advanced Front-End (AFE) algorithm developed by ETSI was also evaluated as one state-of-the-art method, which contains an iterative Wiener filter and blind equalization [12].…”

“…In addition, the acoustic model employed by the feature reconstruction methods also should match the acoustic model (i.e., Hidden Markov Model) of the speech recognizer in terms of the training database and the recording condition to provide the best speech recognition performance. Many feature reconstruction methods employ acoustic model such as Gaussian Mixture Model (GMM) [5] [6].…”

Features and Model Adaptation Techniques for Robust Speech Recognition: A Review

“…The use of a single vector r can only compensate for convolutional noise in the feature domain. In [50], a method called multivaRiate gAussian-based cepsTral normaliZation (RATZ) is proposed to use multiple correction vectors. In RATZ, the clean feature space is modeled by a GMM.…”

“…The STAR algorithm of Moreno [50] is closely related to the RATZ feature compensation algorithm described in section 3.2. Feature compensation methods usually have a model adaptation counterpart.…”

“…However, unlike RATZ which uses a separate GMM for the prior distribution of clean speech, STAR utilizes the HMM. STAR estimates the correcting terms, μ k and ∑ k , for the 256 Gaussians using the same way as RATZ, and then compensates the clean mean and variance vectors to approximate the noisy speech distribution [50]. As these Gaussians are shared by all HMM models, once they are compensated, all the HMM states are adapted.…”