Artificial Neural Networks and Speech Recognition

The ability to solve some classification problems with relative ease and without the need to formalize the statistical properties of the problem have made neural networks very popular in the field of speech processing. In this chapter, the usefulness of a neural network using autoregressive backpropagation and time-delay components (AR-TDNN) is illustrated. Combined with HMMs, the AR-TDNN is incorporated in a flexible hybrid structure which attempts to improve the performance of complex phonetic feature (nasality) detection and classification.