Efficient classification of noisy speech using neural networks

The classification of active speech vs. inactive speech in noisy speech is an important part of speech applications, typically in order to achieve a lower bit-rate. In this work, the error rates for raw classification (i.e. with no hangover mechanism) of noisy speech obtained with traditional classification algorithms are compared to the rates obtained with Neural Network classifiers, trained with different learning algorithms. The traditional classification algorithms used are the linear classifier, some Nearest Neighbor classifiers and the Quadratic Gaussian classifier. The training algorithms used for the Neural Networks classifiers are the Extended Kalman Filter and the Levenberg-Marquadt algorithm. An evaluation of the computational complexity for the different classification algorithms is presented. Our noisy speech classification experiments show that using Neural Network classifiers typically produces a more accurate and more robust classification than other traditional algorithms, while having a significantly lower computational complexity. Neural Network classifiers may therefore be a good choice for the core component of a noisy speech classifier, which would typically also include a hangover mechanism and possibly a speech enhancement algorithm.