Modulation filtering using an optimization approach to spectrogram reconstruction

Modulations across time and frequency are known from previous studies to play a significant role for speech intelligibility. Hence, well-chosen manipulations of modulations via an accurate tool to systematically modify the modulation content of a signal might be useful for the improvement of speech intelligibility. This study investigates modulation filtering in a time-frequency representation of the signal (e.g., a spectrogram), using a novel approach for reconstructing a signal from its modified representation. It is suggested that this synthesis is regarded as an optimization problem, where the variables are the time samples of the output signal and where the cost function to minimize is the difference between the target spectrogram and the current spectrogram. This approach is made feasible, with regard to the large number of variables involved, by use of a limited-memory optimization algorithm. This study presents basic results regarding temporal modulation filtering and discusses the novel method and its possibilities of improvement.