Noise robust exemplar-based connected digit recognition

Abstract

This paper proposes a noise robust exemplar-based speech recognition system where noisy speech is modeled as a linear combination of a set of speech and noise exemplars. The method works by finding a small number of labeled exemplars in a very large collection of speech and noise exemplars that jointly approximate the observed speech signal. We represent the exemplars using melenergies, which allows modeling the summation of speech and noise, and estimate the activations of the exemplars by minimizing the generalized Kullback-Leibler divergence between the observations and the model. The activations of the speech exemplars are directly being used for recognition. This approach proves to be promising, achieving up to 55.8% accuracy at signal-to-noise ratio −5 dB on the AURORA-2 connected digit recognition task.

OpenBliSSART: Design and evaluation of a research toolkit for Blind Source Separation in Audio Recognition Tasks

“…Furthermore, applications of NMF in audio processing are not limited to BSS, as there is a growing number of studies showing the advantage of NMF-based audio feature extraction, especially in noisy conditions [6][7][8]. On the other hand, with the increasing amount of computational power available today even on mobile devices, we are moving towards the point where NMF-based algorithms are ready to be used in real-life applications.…”

“…Consequently, it receives increasing attention at the moment [6]. Thereby different sets of components are selected as an in-or overcomplete feature basis, for which the time-varying activation matrix (i. e., the second NMF factor) is computed by supervised NMF.…”

“…The aforementioned NMF and NMD algorithms can be run on magnitude, power, and Mel-scale spectra; besides, the 'sliding window' NMF from [6] is implemented that transforms the original spectrogram V to a matrix V such that every column of V is the row-wise concatenation of a sequence of short-time spectra (in the form of row vectors):…”

Exemplar-based joint channel and noise compensation

“…models which describe the magnitude spectra of complex sounds as being composed of a purely additive (no negative components) combinations of spectral atoms, have proven to be adept at separating the target speech from interfering sounds such as noise [1,2], other speakers [3,4], music [5,6,7] and even reverberation [8]. For noise-robust automatic speech recognition (ASR), such compositional models really excel when the atoms also have some temporal extent [9,10].…”

“…The techniques needed to build a noise-robust system based on the compositional model have been explored in previous work. In [2], we proposed the use of speech and noise exemplars as dictionary atoms and showed that it is possible to directly map the sparse weights of the clean speech atoms to state posteriors. In a feature enhancement method [9], the compositional model is used to obtain clean speech and noise estimates which are in turn used to define a Wiener filter.…”