A Spectrum Recovery Algorithm using Signal-to-Noise Ratio Classification for Noise Reduction

In the area of speech signal processing, real background noise is important problem for noise reduction, therefore more skillful methods are required in this area. Accordingly, this paper proposes a spectrum recovery algorithm using a signal-to-noise ratio classification method based on a classification of a voiced or unvoiced signal. Therefore, the proposed algorithm recovers a speech spectrum from a noisy speech spectrum using a time-delay neural network for noise reduction. As such, the proposed system detects the voiced and unvoiced signal, then reduces the noise spectrums for each input frame using the time-delay neural network. Based on measuring correct classification rates and spectrum recovery results, experiments confirm that the proposed algorithm is effective for speech degraded by various noises.