Voiceless stop consonant identification using LPC spectra

Three types of LPC spectrum-based systems for discriminating among the voiceless stop consonants /p,t,k/ were tested on a large multi-speaker corpus. The systems differed in the number of short-time spectra used in recognition and in the decision strategies employed. The first type of system used a single spectrum computed at the point of consonant release. Systems of the second type used a sequence of three spectra computed using nonoverlapping windows spanning a 75 msec interval around the stop release. The third type of system used 12 spectra computed over an 80 msec interval. Classification was performed with k-nearest-neighbor and minimum vector quantization distortion classification rules. Recognition accuracies of 90.92%, 90-92% and 91- 95% were obtained for the three types of systems.