Optimizing Algorithms for Phrase Recognition

The paper presents a simple algorithm to cull short phrases of interest spoken in a long recorded speech. Listening to a long and boring recording of a suspect who is being wiretapped may irk intelligence investigator if it does not contain any speech segment of interest. It would be much useful if first a preliminary test is run to see if the recorded speech does contain phrases of interest which could help convict a suspect. If so the entire recording can then be reexamined. Vowels spoken in an utterance can be reasonably identified by locating the vowels’ fist two formants i.e., F1 and F2 onto vowel loops similar to the ones drawn by Peterson and Barney for English vowels’ sounds [1]. Using Pratt or a similar type of software can determine the formants of vowels. Two algorithms were designed to map the vowels’ formants spoken in an utterance onto the vowel loops and identify which vowel sounds they represented [2]. The algorithm-1 used the calculated value of F1 and determined in which vowel-loop F2 lied, and algorithm-2 used the calculated value of F2 and determined in which vowelloop F1 lied. In some cases the algoirithm-1 did a better job than algorithm-2 in terms of computational time while in other cases the reverse was true. This paper essentially extends the idea of mapping (F1,F2) onto the vowel loops by another algorithm, called algorithm-3, which is faster than the previous two. The algorithm-3 requires conversion of vowel loops in a particular format to act as our data bank; let us call it conversion table because it will convert any point F1,F2 that is covered by the vowel loops into a corresponding vowel symbol. The calculated values of F1 and F2 of vowels are discretized in suitable steps, and a corresponding vowel symbol is determined from the conversion table. The vowel symbol at the intersection of F1 and F2 is selected.