The effect of speaking rate and vowel context on the perception of consonants in babble noise

In this paper, we study human perception of consonants in the presence of additive babble noise at two speaking rates. In addition, we work on a model that attempts to replicate these human results through a phoneme recognition model. Consonant-Vowel-Consonant (CVC) stimuli comprising of a set of 13 consonants and 3 vowels (/a/, /i/, /u/) were recorded in a sound proof booth by two talkers at two different speaking rates (fast and slow). Noisy stimuli were generated by adding babble noise at different levels to the quiet recordings. These stimuli were used to conduct perceptual experiments in which 52 listeners were asked to listen and repeat back the CVC phrases presented in babble noise under 3 SNR conditions and both speaking rates. The data was transcribed by two trained linguists. The results were analyzed by SNR, vowel, and speaking rate. Rate did not have an effect on the perception of consonants in quiet conditions. With the exception of /CuC/ stimuli, speaking rate had a pronounced effect, with slow speech being more intelligible than fast speech in the presence of noise. /CaC/ stimuli were, on average, more robust than other stimuli in all conditions. In addition, syllable initial consonants were better identified than final consonants, especially in noise. The effect of rate was more pronounced in voiced syllable final consonants. The stimuli collected was run through a phoneme recognition model based on GMMs. The machine effectively sees ”noisy” data even though we present it with clean stimuli due to mismatched train and test conditions (train on TIMIT, test on sound booth stimuli). The machine consonant recognition accuracies and confusions are similar to that of humans for 0dB stimuli. The machine also mimicks the effect of rate on humans for the /i/ and /u/ context while it does not for the /a/ context..