Identifying Key Phoneme Features

Spectrograms carry all necessary information for reliable human and computer perception of speech. This paper discusses the importance of spectrogram features used by a recognition algorithm developed by Ali et al. as they relate to human perception. Features, including MNSS, burst frequency, formant transitions, voicing onset time, and voicing/unvoicing information are defined and their importance to computer stop consonant recognition described. Confirming many previous findings, burst frequency and formant transitions were found to be most important in the perception of speech synthesized from spectrograms while other features played a secondary role. Software tools developed that should facilitate other similar investigations are described.