A Voice Analog

Abstract

An artificial voice was constructed using novel and simple techniques. An outgrowth of an experiment in speech recognition, this voice analog was used to ascertain the efficiency of the information extraction techniques used in a speech recognition machine. The extracted information curves were reconstructed to full speech by means of this voice analog. Four voice-characterized noise generators were connected to four voltage-controlled bandpass filters. Each noise generator and filter pair operated within an assigned major subband. Four slowly varying frequency-describing information curves were connected, one to each of the filters. Each filter then passed a narrow minor subband of frequencies determined by the voltage of the frequency-describing curve at the same instant. Three additional power-describing information curves modulated the four filter outputs. The resulting curves were mixed, amplified, and connected to a loudspeaker. The voice analog speech was harsh and estimates of its intelligibility varied greatly. Plosive consonants were good and consonants in general were better than vowels. The complete system occupied only a few square feet of bench space. Although the voice analog was primarily an experiment which provided the desired data and was not fully refined at the time of its construction, its compactness and simplicity suggest the desirability of further development.