The high redundancy of speech sounds in sentential context.

Abstract

Readers are exceptionally good at guessing which letter comes next in printed text, even with little contextual information. This was quantified by Shannon [Bell Syst. Tech. J., (1948); Bell Syst. Tech. J., (1951)], who demonstrated that printed English text is approximately 50% redundant (predictable). Here, redundancy of American English speech was examined in a modified “guessing game” paradigm, as listeners guessed which speech sound came next throughout an entire sentence. Four TIMIT sentences were presented to 19 listeners, and guesses were made by clicking buttons on a computer screen corresponding to each consonant and vowel sound in American English. Listeners received feedback after each guess and heard each sentence (up to their most recent correct guess) as often as they wished. Numbers of guesses were at chance levels at sentence and word onsets, and were much better than chance with the addition of phonemic and sentential context. Equiprobable speech sounds predict 0% redundancy (chance performance) and entropy of 5.29 bits/sound. Listener performance demonstrates a mean sentence redundancy of 40% (3.18 bits/sound), within a narrow range of 36–43. Behavioral results will be evaluated with reference to simple phoneme frequency and transitional probabilities between phonemes. [Work supported by NIDCD.]