Using neural network and principal component analysis to study vowel recognition with temporal envelope cues

Previous studies in normal-hearing and cochlear-implant subjects have shown high levels of speech recognition with primarily temporal envelope cues. The present study used principal component analysis (PCA) to extract important features in temporal envelopes and then constructed a 3-layer feedforward artificial neural network to study their role in vowel recognition. Twelve vowels by 30 speakers in a /hVd/ context served as the test material. Temporal envelopes from 1 to 8 spectral bands were extracted and subjected to PCA with 15 principle components. Similar to previous perceptual data, the present study showed that 63% correct vowel recognition was achieved with only 4-band envelope cues. The principle components responsible for this high level of vowel recognition included phonemic transition cues and steady-state amplitude cues. The present result can be applied to the development of novel algorithms to improve performance for automatic speech recognition and auditory prosthetic devices.