A computational model of sound recognition used to analyze the capacity and adaptability in learning vowel classes

Sound recognition is likely to initiate early in auditory processing and use stored representations (spectrotemporal templates) to compare against spectral information from auditory brainstem responses over time. A computational model of sound recognition is developed using neurobiologically plausible operations. The adaptability and number of templates required for the computational model to correctly recognize 10 Klatt-synthesized vowels is determined to be around 1250 templates when trained with random fundamental frequencies from the male pitch range and randomized variation of the first three formants of each vowel. To investigate the ability to adapt to noise and other unheard vowel utterances, test sets with 1000 randomly generated Klatt vowels in babble at signal-to-noise ratios (SNRs) of 20 dB, 10 dB, 5 dB, 0 dB, and 􀀀5 dB are generated. The vowel recognition rates at each SNR are 99.7%, 99.6%, 97.0%, 77.6%, and 54.0%, respectively. Also, a test set of four vowel recordings from four speakers is tested with no noise, giving 100% recognition rate. These data suggest that storage of auditory representations for speech at the spectrotemporal resolution of the auditory nerve over a typical range of spoken pitch does not require excessive memory resources or computing to implement on parallel computer systems.