Exploiting Combination of Spectral Features in Light Weight Neural Network for ABE on Limited Data

This paper proposes an approach to extend the bandwidth of 4 KHz speech signals upto 8 KHz using a simple feed-forward neural network (FFNN). We employ the commonly used Short-Time Fourier Transform (STFT) values as features and boost its resolution by incorporating Mel Frequency Cepstral Coefficients (MFCCs). Apart from this feature-level fusion, a layer-level fusion is also examined to demonstrate the expansion in bandwidth. This is an advantage over traditional artificial bandwidth extension (ABE) methods using neural nets, which are heavily data-driven and require high model complexity. The proposed approach exploits the combination of relevant spectral features to avoid the requirement of highly complex architectures. This is because the combined features capture more intricate features relating the 4 KHz signal to the 8 KHz one. This facilitates the deployment of the technique on resourceconstrained devices, even with limited data. These reconstructed high pass (8 KHz) audio signals are perceptually better than the original 4 KHz bandwidth signals. Additionally, the performance of the proposed approach exceeds that of state-of-the-art methods in the literature. This approach has potential applications in enhancing the quality by extending the bandwidth of audio signals in telecommunication systems and also finds application in cochlear implant (CI) users.