Title: Perceptual confusions for temporally smoothed envelope of syllables in noise
The goal of the present study was to investigate the nature of the loss of the temporal envelope cues produced by temporally smoothing the CV envelopes with LPF of different cutoff frequencies under conditions of different SNRs in NH listeners. Stimuli were generated from 16 consonant-vowel (CV) sounds of the Linguistic Data Consortium. For each CV sound, the time-intensity envelope of speech was extracted from each of 26 critical bands by the Hilbert transformation. Temporal smearing of these processed signals was produced by applying low-pass filters with one of five cutoff frequencies in Hz (4, 8, 16, 32, No filtering). Confusion matrices for 16 CV sounds were measured as a function of SNR in dB (–12, –6, 0, 6, 12, Q). The results showed that the grand mean error was lowered with increase in temporal modulation across SNR with higher weight on SNR than on LPF. The results also showed that CVs /fa/, /va/, /THa/, and /tha/ were the most difficult sounds (H group), seven CVs /ka/, /ma/, /na/, /sa/, /sha/, /jza/, and /za/ were the easiest sounds (L group), and five CVs /ba/, /da/, /ga/, /pa/, and /ta/ were the most influenced by LPF coutoff (T group). The major competitors between the unprocessed and the processed stimuli were similar for H and L groups, but different for T group. The Phoneme Discrimination Threshold for the spoken CV confirmed that H group required higher SNR than T group than L group. The conditional entropy for the articulatory feature system showed that duration was the most robust channel, whereas affrication was the least robust channel. The information transmitted for nasality and place were similarly transmitted, but that for voicing was relatively low. The percent for the variance accounted for by the feature system was between 80 and 90% in quiet.