Maximum a posteriori Speech Enhancement Based on Double Spectrum

While the acoustic frequency domain has been widely used for speech enhancement, usage of the modulation domain is less common. In this paper, we investigate single-channel speech enhancement in the recently proposed Double Spectrum (DS) framework and provide insights on the statistical properties of speech and noise in the DS domain. Relying on our statistical analysis in the DS, we derive a maximum a posteriori estimator of speech in the DS domain. By means of experiments, we evaluate the speech enhancement performance of the proposed method and relevant benchmarks in the acoustic frequency and modulation domains and show that the proposed method achieves a good balance between noise attenuation and speech distortion for various SNRs and noise types.