Speaker-independent connected digit recognition algorithm for a single chip signal processor

Abstract

An algorithm was designed for a fixed point arithmetic signal processor chip to perform real-time speaker-independent English digit recognition. Each word was represented by a single 10-state Markov model, the states of which were 9-way mixtures of 24-dimensional Gaussian densities of cepstral features. Algorithms for feature extraction include autocorrelation, linear predictive coding, and the computation of both cepstra and differential cepstra. Algorithms for pattern matching include a Laplacian distance measure, viterbi decoder, best choice, and partial traceback. To achieve real-time operation, single precision arithmetic was employed for the Laplacian distance metric, which is the bottleneck in the recognizer. Memory storage was minimized by quantizing model parameters to 10 bits and dynamically pruning a tree of word candidates. Recognition accuracy of about 98% per word was obtained; this is approximately the same as that obtained with a floating point simulator as tested on a connected digits NIST database.