Effect of neural network input span on phoneme classification

Feedforward neural networks spanning input durations of 35, 65, 125, and 245 ns were trained to classify subword speech segments from continuous utterances into 46 phoneme-like classes. The performance of networks with different input spans showed that brief sounds can be reliably detected by networks with longer input spans, but results for a subset of longer-duration phonemes (diphthongs) indicate that a network needs a wide enough view of the input to capture the salient features of the output class. The best classification performance was observed using a completely connected three-layer network with a 125-ns input span. Performance averaged 56%, ranging from 31% to 82% across phoneme classes using a top-three candidates decision criterion