Word recognition in continuous speech using linear prediction analysis

A promising method of automatic word recognition in continuous speech, recently designated as word spotting, has been demonstrated. The method uses error residual ratios from LPC (Linear Predictive Coding) vocoder analysis for waveform comparison and a dynamic programming procedure for time registration between the incoming speech and a template of the key word. Using a similarity threshold, the incoming speech is compared with several templates to account for variability is spectral shape. This system can work in real time using a real time vocoder. The multiple templates are used in such a way that a small number of templates, three or four, is made to look like several hundred or more. This is accomplished by dynamically constructing a composite template from parts of each single template as part of the processing of the incoming speech. Thus, a particular composite template is constructed for each word being recognized. An accuracy of 99 percent with no false alarms was achieved using 205 key words, five different speakers, and approximately ten minutes of speech text. Performance in the presence of additive white gaussian noise of approximately 11 dB signal-to-noise ratio was 66 percent. When the speech was processed to account for the noise, results improved to 85 percent to 90 percent accuracy. Finally, a digit recognition experiment was performed using over 1200 digits spoken by ten different people with a resultant accuracy of 97 percent.