Compensation for Nonlinear Distortion in Noise for Robust Speech Recognition.

摘要

The performance, reliability, and ubiquity of automatic speech recognition systems has flourished in recent years due to steadily increasing computational power and technological innovations such as hidden Markov models, weighted finite-state transducers, and deep learning methods. One problem which plagues speech recognition systems, especially those that operate offline and have been trained on specific in-domain data, is the deleterious effect of noise on the accuracy of speech recognition. Historically, robust speech recognition research has focused on traditional noise types such as additive noise, linear filtering, and reverberation. This thesis describes the effects of nonlinear dynamic range compression on automatic speech recognition and develops a number of novel techniques for characterizing and counteracting it. Dynamic range compression is any function which reduces the dynamic range of an input signal. Dynamic range compression is a widely-used tool in audio engineering and is almost always a component of a practical telecommunications system. Despite its ubiquity, this thesis is the first work to comprehensively study and address the effect of dynamic range compression on speech recognition.;More specifically, this thesis treats the problem of dynamic range compression in three ways: (1) blind amplitude normalization methods, which counteract dynamic range compression when its parameter values allow the function to be mathematically inverted, (2) blind amplitude reconstruction techniques, i.e., declipping, which attempt to reconstruct clipped segments of the speech signal that are lost through non-invertible dynamic range compression, and (3) matched-training techniques, which attempt to select the pre-trained acoustic model with the closest set of compression parameters. All three of these methods rely on robust estimation of the dynamic range compression distortion parameters. Novel algorithms for the blind prediction of these parameters are also introduced. The algorithms' quality is evaluated in terms of the degree to which they decrease speech recognition word error rate, as well as in terms of the degree to which they increase a given speech signal's signal-to-noise ratio. In all evaluations, the possibility of independent additive noise following the application of dynamic range compression is assumed.