A Simple Cepstral Domain DNN Approach to Artificial Speech Bandwidth Extension

High‐band feature extraction for artificial bandwidth extension using deep neural network andH∞optimisation

“…In this work, a segmental signal‐to‐noise ratio (segSNR) [57], log spectral distance (LSD) [44], NB mean opinion score listening quality objective (MOS‐LQO) [57, 58], and WB MOS‐LQO [59, 60] as the objective measures are taken for examining the quality of artificially extended speech signals. Next, we convert the IIR filter into an approximate FIR filter by using the Taylor series truncation method.…”

“…Here, the proposed approach is compared with other existing approaches by maintaining the same experimental conditions such as LPF, HPF, the dimension of HB feature, DNN architecture (seven hidden layers and 256 neurons in each hidden layer), data set, and NB signal processing. Two recently reported current works such as the modulation technique [14] with a slight modification and a cepstral domain approach [44] are included for comparison. The gain for the modulation technique is calculated by following [15] and the cepstrum feature is used for representing the NB information as well as the HB spectral envelope information.…”

“…The gain for the modulation technique is calculated by following [15] and the cepstrum feature is used for representing the NB information as well as the HB spectral envelope information. The NB feature and HB feature for the cepstral domain approach contain the NB magnitude spectrum and cepstral coefficients representing the HB magnitude spectrum [44], respectively. Objective measures are arranged in Table 4 for the proposed approach and the existing methods using the same DNN model.…”

“…The current approach directly obtains an IIR synthesis filter, which contains needed HB signal information. Moreover, the NB signal and the estimated HB signal are added using the discrete Fourier transform (DFT) addition and gain adjustment techniques with slight modification in the current approach [14, 15, 27, 44]. The benefit of utilising the DFT addition is in removal of the leaked information in these signals from the synthesis filter and LPF.…”

Sinusoidal-Based Lowband Synthesis for Artificial Speech Bandwidth Extension

“…Furthermore, DNNs have been employed for estimating the UB spectral envelope directly (regression) [25]- [28]. Opposed to the source-filter model, UB spectral magnitudes and UB phases can be estimated right away using sum-product networks (SPMs) [29], DNNs [30], [31], or recurrent neural networks (RNNs) [32], which can then be transformed back to the time domain by an overlap-add (OLA) structure. In several studies, an increased speech quality when using ABE solutions was shown [18], [33].…”