Description:
<jats:p>A feedforward connectionist network trained by backpropagation was used to detect 15 speech features. The network was trained over 240 sentences (40 men and 40 women), and tested over 200 sentences (10 men and 10 women), all part of the MIT Ice Cream database. Network input consisted of a smoothed spectral vector at 15-ms-intervals, plus two coefficients of amplitude and spectral change. The network achieves a signal detection discrimination level (a-prime) of 0.87 compared to a level of 0.76 for a ten-nearest-neighbor system. Almost identical training and test performances indicates excellent generalization to new speakers and text. Processing costs are mainly signal processing and network training; detection itself can be done in real time. Performance is much better for broad features like sonorance, which occur frequently, than for infrequent features like sibilance, partly because of their low frequency and partly because of other characteristics. [Work supported by USWest.]</jats:p>