For practical hardware implementations of isolated‐word recognition systems, it is important to understand how the feature set chosen for recognition affects the overall performance of the recognizer. In particular, we would like to determine whether hardware implementations could be simplified by reducing computation and memory requirements without significantly degrading overall system performance. The effects of system bandwidth (both in training and testing the recognizer) on the performance must also be considered since the conditions under which the system is used may be different than those under which it was trained. Finally, we must take account of the effects of finite word‐length implementations, on both the computation of features and of distances, for the system to properly operate. In this paper we present the results of a study to determine the effects on recognition error rate of varying the basic analysis parameters of a linear predictive coding (LPC) model of speech. The results showed that system performance was best with an analysis parameter set equivalent to what is currently being used in the computer simulations, and that variations in parameter values that reduced computation also degraded performance, whereas variations in parameter values that increased computation did not lead to improved performance.

A digital-based isolated word recognition system has been implemented in a module of dedicated hardware that uses a microprocessor and programmable digital signal processing circuitry. The recognizer is based upon the minimum prediction residual principle of Itakura. The recognition algorithm has been developed and tested on a general-purpose minicomputer and array processor, where it has been shown to be suitable for several recognition tasks. The recognition hardware consists of an Intel 8086 16-bit microprocessor operating in parallel with a digital speech processing peripheral (DSPP) tailored to the algorithm. The microprocessor performs the supervisory and decision operations; the DSPP performs the 200,000T + 4,500N multiply-add operations (and associated data transfers) associated with the recognition of a word of duration T sec from an N word vocabulary with I template per word. The recognizer is compact (board area of 250 sq. in.) and inexpensive (commercial component cost of about $1200 for 40 word templates).

Speech recognition systems based on LPC features sets have been applied successfully to a number of speech recognition tasks including an airlines information system, a directory assistance system, a voice repertory dialer, and a connected digit recognizer. All these systems have been based on a recognition model, originally proposed by Itakura, with a fixed analysis parameter set. To get an appreciation for the robustness of this feature set, an experimental investigation was undertaken to vary each of the parameters of the analysis system, and to measure the effect of recognition accuracy. The parameters chosen for study included—N, the size of the analysis frame; L, the shift between frames; p the number of LPC poles; α, the preemphasis factor, and LP, the analog filter cutoff. Four talkers were used in this study. Each talker recited the vocabulary (the digits, the letters of the alphabet and three command words) seven times for training, and ten times for subsequent testing. The speech was stored in digital form at a sampling rate of 20 000 Hz. Speaker trained templates were obtained from the training data. The results of this study indicate that the recognition accuracy remained high over a wide range of variation of analysis parameters.

No abstract

No abstract