Results Obtained from a Vowel Recognition Computer Program

Abstract

As a first step toward a general speech recognition computer program, a program has been developed to recognize ten Eng.ish vowels in isolated words of the form /b/—bowel—/t/. Input to the computer was real time spectral data obtained through the input system described at these meetings in May, 1958, by Forgie and Hughes. The program was developed from the study of spectral data from five speakers. The program was tried on new speakers, the errors analyzed, and the program modified. The program, as finally evolved by this process, first determines the rough location of the first two formants. The remaining confusions are resolved by combinations of the following, as required: (1) position of F3, (2) the slopes of F1 and F2, (3) an estimate of pitch, and (4) the energy in the frequency band between F1 and F2. Whenever the over-all power is above an arbitrary threshold, each 6-msec scan of the filtered data is classified as one of the ten possible vowels. A tally at the end of the word determines the final decision. The over-all recognition score for 21 subjects (11 male and 10 female) is 89%. Modifications of the program took place on the first 16 subjects only. The score for the last five speakers was 82%. (Lincoln Laboratory is operated by Massachusetts Institute of Technology with joint support of U. S. Army, Navy, and Air Force.)