Figure 2. The effects of noise on speech identification and neural coding are compared between intact speech and four speech vocoders that differed in their composition of acoustic ENV and TFS. A, Mean consonant identification scores across listeners (with SEM bars) are shown as a function of signal-to-noise ratio, SNR, with chance level (1/16) indicated by the dashed line. As noise level increased (SNR decreased), identification became more difficult in all cases; however, the relative performance across vocoders differed for positive and negative SNRs. B, C, Neural coding of TFS (ρTFS) and phonemic ENV (ρENV), where the neural cross-correlation coefficients were computed between model spike train responses to the noisy vocoded speech and the intact speech in quiet. Mean ρTFS and ρENV values across AN fibers [all eight CFs ≤ 2.5 kHz for TFS (Johnson, 1980); all 10 CFs ≤8 kHz for ENV] are plotted with SEM bars. Recovered phonemic envelope coding [e.g., from the periodicity envelope (PDENV), broadband TFS (BBTFS), and narrowband TFS (NBTFS) speech vocoders] is represented by dashed curves in C, whereas true phonemic envelope coding is represented by solid curves. Black dashed line at ρENV = 0.1 in C represents the ENV noise floor; the TFS noise floor was negligible (ρTFS = 0.01).