An Investigation of Signal Processors for Recognition

Currently, in uncorrupted acoustic environments, a state-of-the-art, continuous speech, 5000word vocabulary speech recognizer can achieve a recognition accuracy of 95%. In a telephone channel acoustic environment, the same recognition system can double in error rate, which is unacceptable for most applications. It is the goal of this project to quantitatively measure how well a speech recognition system recognizes natural number utterances (e.g. \four thousand lira") over a telephone line. The performance of recognizers will be quanti ed using word error, which is a measure of the di erence between the actual transcription of a speech waveform and its recognized transcription. This project will compare two signal processing algorithms used in recognition of telephone speech. One standard, widely accepted signal processing algorithm, cepstral mean subtraction, motivated by computational e ciency and high performance, was evaluated. Another newer signal processing algorithm, RASTA, motivated primarily by auditory physiology was compared to cepstral mean subtraction in terms of word error. The two algorithms performed comparably in tests. CMS produced a 22% word error and RASTA produced a 25% word error on this di cult task. This project is signi cant from a biomedical point of view because telephone speech recognizers can be used to aid the handicapped. For example, the deaf could use a speech recognition system to aid them in their telephone communication with hearing people.