Preprocessing and neural classification of English stop consonants [b, d, g, p, t, k]

Neural networks are accepted as powerful learning tools in pattern recognition in which they proved their performance. Nevertheless, many problems like phoneme classification with a multi-speaker continuous speech database are hard even for neural networks. The authors´ aim is to propose an artificial neural network architecture that detects acoustic features in speech signals and classifies them correctly. They reached this goal with English stop consonants [b, d, g, p, t, k] extracted from the general multi-speaker database (TIMlT) by modifying some parameter values in the preprocessing algorithm and by using a modified TDNN (time delay neural network) architecture. The net performed a good classification giving as testing recognition percentage the following results: 92.9 for [b], 91.8 for [d], 92.4 for [g], 80.3 for [p], 90.2 for [t], 91.2 for [k]