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

Abstract

This work aims to enhance the quality of narrowband (0–4 kHz) voice signal in terms of frequency components, i.e. missing high‐frequency components in a range of 4–8 kHz. The proposed artificial bandwidth extension framework uses the H∞ optimisation. In this context, a signal model is used to get a better representation of wideband (0–8 kHz) information of a signal. The H∞ optimisation is used to obtain the synthesis filter for a given signal model, which is used to synthesise the high‐band (4–8 kHz) signal. The discrete Fourier transform addition is performed to add the narrowband signal and estimated high‐band signal for removing the leaked information from the synthesis filter and non‐ideal low pass filter. Gain adjustment is performed on the estimated high‐band signal to make its energy equal to the true high‐band signal. Non‐stationary characteristics of speech signals generate an assorted variety in synthesis filters and corresponding gain. For this, a deep neural network (DNN) is used to estimate the synthesis filter and gain by using the given narrowband information. The authors analyse the performances of the DNN model on two data sets. Objective and subjective analyses are carried out on these data sets.

[Retracted] Application of Artificial Neural Network Algorithm in Facial Biological Image Information Scanning and Recognition

“…The application performance of the proposed method in feature extraction of multipose face images is simulated [ 19 ]. The experiment is designed with MATLAB.…”

Application of Neural Network Algorithm Based on Principal Component Image Analysis in Band Expansion of College English Listening

“…ese methods focus on exploring the correlation between high and low frequencies in the current audio frame, focusing on the display of the static characteristics of the spectrum. Gupta et al used the hidden Markov model to simulate the time-domain dynamic evolution of the audio spectrum envelope [13] and introduced interframe correlation into spectrum envelope estimation. However, because this method only uses discrete states to simulate the time evolution of the actual audio spectrum, there is still dynamic distortion in the reconstructed audio.…”

Structural inversion of radar emitter based on stacked convolutional autoencoder and deep neural network

“…Radar emitter recognition (RER) [2] is one of the main functions of radar countermeasure systems and includes modulation type recognition, waveform recognition [3][4][5][6], and specific emitter identification (SEI) [7]. Since the deep learning method was introduced to SEI [8], the methods for the fine feature extraction of radar signals have become increasingly more abundant, and the use of feature extraction via deep learning methods is on the rise [9,10]. At present, feature extraction can be carried out from the three aspects of the time domain, frequency domain, and time-frequency domain.…”