Spire Based Speaker-Independent Continuous Speech Recognition Using Mixed Feature Sets

摘要

A system was developed to investigate continuous speech recognition. The system incorporates multiple features and dynamic programming to recognize continuous inputs of the spoken digits (zero through nine). The fundamental design concept extends from previous successful recognition research efforts involving both isolated and continuous speech using multiple feature sets, multiple template sets, and dynamic programming. Among the features used in the investigation are wide band spectrogram, narrow band spectrogram, linear predictive coding (LPC) coefficients, LPC spectrum, frication frequency, and format tracks. An advanced speech research tool called SPIRE provided the computational functions needed to extract the raw features.