Model-based Optimization of a Low-dimensional Modulation Filter Bank for DRR and T60 Estimation
Semih Ağcaer, Rainer Martin · 2019
Amplitude Modulation Spectrum (AMS) features can be implemented as a cascade of two filter banks whereas the filter bandwidths can be optimized for a particular application. In this work we train AMS-based features using a combination of a model-based optimization (MBO) approach and feature selection for full-band DRR and full-band T60estimation. MBO replaces the computational complex data-based cost function by approximating a less complex surrogate model and thus reduces the time needed for training. We evaluate our approach on the publicly available ACE challenge corpus and achieve with only five features the best RMSE in the DRR estimation task using the single microphone configuration and upper mid-range performance for T60estimation. The computational complexity of our algorithm is much lower than all other submitted algorithms.