Feature Selection for DNN-based Bandwidth Extension
* Presenting author
Artificial bandwidth extension (BWE) is still an important topic, especially in the automotive domain where consumers experience a dramatic degradation in voice quality when a wideband call suddenly falls back to 8-kHz GSM. This happens e.g. due to poor network coverage in the countryside. The aim of BWE is to bridge the perceived voice quality gap by reconstructing the wideband signal. In this work, we take a Deep Neural Network (DNN) – based approach. We address the problem of selecting a robust feature set from a larger pool of time- and frequency-domain features.This is achieved in a bottom-up fashion. Starting with Mel Frequency Cepstral Coefficients (MFCC) as a basic feature set, we conduct a sequence of experiments to evaluate the performance improvement that can be achieved by adding a feature from the pool. This is carried out for all features and the one with the highest improvement is selected. The final feature set is obtained by iteratively repeating this procedure until the achievable improvement drops below a threshold. A focus lies on the robustness of frequency-domain features in comparison with time-domain features regarding background noise and channel characteristics.