Artificial bandwidth extension using H∞ sampled-data control theory

Abstract Artificial bandwidth extension of a speech signal is a way to improve speech quality and intelligibility in narrowband telephonic communication. Artificial bandwidth extension techniques extend the bandwidth of narrowband signals using only narrowband information available at the receiver end. This work proposes a new bandwidth extension technique based on the H ∞ sampled-data control theory and deep neural network (DNN) regression approach for recovering the missing high-frequency components of the speech signal. The H ∞ sampled-data control theory helps in designing of a synthesis filter by optimally utilizing the inter-sample information of a signal and a signal model. The obtained synthesis filter is further used to recover the high-frequency information of the signal. The non-stationary (time-varying) characteristic of speech signals mandates numerous synthesis filters for reconstructing the whole speech signal. Hence, a DNN model is used for estimating the synthesis filter information and a gain factor for specified narrowband information of an unseen signal. Objective analysis is done on the TIMIT and RSR15 datasets. Subjective analysis is done on the RSR15 dataset.