Temporal non-place information in the auditory-nerve firing patterns as a front-end for speech recognition in a noisy environment

A computational model, based upon the temporal discharge patterns of auditory-nerve fibers, is described and compared with the more traditional Fourier transform method of spectral analysis. The model produces a frequency-domain representation of the input signal based on the ensemble histogram of interspike intervals generated by a simulated array of auditory-nerve fibers. The nerve-fiber discharge mechanism is modeled as a multi-level crossing detector at the output of each cochlear filter. The model incorporates 85 cochlear filters, equally spaced on a log-frequency scale between 200 and 3200Hz. The level crossings are measured at positive threshold levels which are pseudo-randomly distributed. The resulting “Ensemble Interval Histogram” (EIH) spectrum has two principal properties: (1) fine spectral details, which are well preserved in the low-frequency region, are poorly delineated in the high-frequency portion of the spectrum, (2) the EIH representation withstands the addition of noise to a far higher degree than the traditional Fourier power spectrum. The capability of the EIH model to preserve relevant phonetic information in quiet and in noisy acoustic environments was measured quantitatively using the EIH as a front-end to a Dynamic Time Warping, speaker-dependent, isolated-word recognizer. The database consisted of a 39-word alpha-digits vocabulary spoken by two male and two female speakers, over a range of signal-to-noise ratios. In the noise-free case, the performance of the EIH-based system is comparable to a conventional Fourier-based front-end. In the presence of noise, however, the performance of the EIH-based system is superior. The recognition scores of the EIH-based front-end drop more slowly than those of the Fourier-based system with increases in noise level. As a consequence, the resulting EIH superiority increases as the signal-to-ratio decreases.