Epoch-Based Spectrum Estimation for Speech

An implicit assumption when using the discrete Fourier transform for spectrum estimation is that the time signal is periodic. This assumption clashes with the quasi-periodicity of voiced speech when the traditional short-time Fourier transform (STFT) is applied to it. This causes distortion and leads to a performance handicap in downstream processing. This work proposes a remedy to this by using epochs in the signal to determine better frame boundaries for the Fourier transform. The epochs are the estimated glottal closure instants in voiced speech and significant peaks in the unvoiced speech signal. The resulting coefficients are compared to the traditional STFT coefficients using copy-synthesis. An improvement of 15 dB signal-to-noise ratio and a PESQ score of 2.5 to 3.5 is achieved for copy-synthesis using 20 mel-filters. The results demonstrate that there is a great potential in improving down stream speech processing applications using this approach to spectrum estimation.