Gradient Based Spectral Peak Location for Noise Robust Speech Recognition

In this paper a gradient-based algorithm for finding spectral peak locations is presented. The algorithm makes use of gradient and acceleration locations in the spectrogram for locating the peaks. Use of frequency gradients and accelerations locate peaks. The results are then interpolated to yield a smooth peak envelope. The method is evaluated in the aurora framework. A first pass locates all spectral peaks and automatically eliminates low magnitude, high frequency peaks that are likely to contain more noise than speech information. The second pass widens the spectral peaks with spectral information. This widening is expected to increase automatic speech recognition based on the peak envelope discrete cosine transformation feature representation.