HMM adaptation using vector taylor series for noisy speech recognition

By A. Acero, Li Deng, Trausti T. Kristjansson, Jerry Zhang. 2000

326 citations recorded by the source corpus.

In this paper we address the problem of robustness of speech recognition systems in noisy environments. The goal is to estimate the parameters of a HMM that is matched to a noisy environment, given a HMM trained with clean speech and knowledge of the acoustical environment. We propose a method based on truncated vector Taylor series that approximates the performance of a system trained with that corrupted speech. We also provide insight on the approximations used in the model of the environment and compare them with the lognormal approximation in PMC.

Subfields: Artificial Intelligence, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Signal Processing

Citations in the corpus

Feature Learning in Deep Neural Networks - A Study on Speech Recognition Tasks (2013, methodology): “Other techniques such as Vector Taylor Series (VTS) adaptation are designed to address the mismatch caused by environmental noise and channel distortion [5].”
An investigation of deep neural networks for noise robust speech recognition (2013, methodology): “In methods such as Vector Taylor Series (VTS) adaptation [22], an estimated noise model is used to adapt the Gaussian parameters of the recognizer based on a physical model that defines how noise corrupts clean speech.”
An Overview of Noise-Robust Automatic Speech Recognition (2014, background): “In standard VTS, the delta and delta-delta model parameters are updated [185] with the continuous time approximation [190], which makes the assumption that the dynamic cepstral coef ﬁ cients are the time derivatives of the static cepstral coef-ﬁ cients.”
Porting concepts from DNNs back to GMMs (2013, methodology): “Since the deep GMMs retain the maximum-likelihood trained Gaussians as first layer, model-based noise robust techniques such as VTS [7], SPLICE [8] or joint uncertainty decoding [9] should still be applicable.”
Model-Based Approaches to Handling Uncertainty (2011, methodology): “Rather than estimating the “true” distribution, a simple approximation is to assume that the distribution of the corrupted speech is Gaussian in nature [17, 2].”
Extended VTS for noise-robust speech recognition (2009, unknown): “(4) The corrupted static mean and covariance become [5] μ s y = f (μ s x , μ s n , μ s h ) ; (5a) Σ s y = diag " JxΣ s x J T x + JnΣ s n J T n " .”
Extended VTS for Noise-Robust Speech Recognition (2011, methodology): “There are a range of approximations that can be used to estimate the model parameters given the mismatch function [1], [2].”
Nonlinear Compensation Using the Gauss–Newton Method for Noise-Robust Speech Recognition (2012, methodology): “Proper approximations must be taken to estimate the corrupted speech distributions.”
High-performance robust speech recognition using stereo training data (2001, methodology): “That is, it partitions the acoustic space in terms of the noisy speech, in contrast to some earlier algorithms (such as FCDCN [ 1 ]) that partitioned the acoustic space in terms of the clean speech cepstral vector x. One main advantage of this new partitioning is that it obtains a more uniform and d”
Exemplar-Based Sparse Representations for Noise Robust Automatic Speech Recognition (2011, methodology): “In popular model-based compensation techniques such as parallel model combination (PMC) and vector Taylor series (VTS) approaches [3], [36], [37], the acoustic model is updated in each frame to account for the noise estimated to be present in that frame.”
Machine Learning Paradigms for Speech Recognition: An Overview (2013, methodology): “A set of generative learning approaches in ASR following this philosophy are variably called " parallel model combination " [26], vector Taylor series (VTS) method [27], [28], and Algonquin [29].”