A deep neural network approach to speech bandwidth expansion

Abstract

We propose a deep neural network (DNN) approach to speech bandwidth expansion (BWE) by estimating the spectral mapping function from narrowband (4 kHz in bandwidth) to wideband (8 kHz in bandwidth). Log-spectrum power is used as the input and output features to perform the required nonlinear transformation, and DNNs are trained to realize this high-dimensional mapping function. When evaluating the proposed approach on a large-scale 10-hour test set, we found that the DNN-expanded speech signals give excellent objective quality measures in terms of segmental signal-to-noise ratio and log-spectral distortion when compared with conventional BWE based on Gaussian mixture models (GMMs). Subjective listening tests also give a 69% preference score for DNN-expanded speech over 31% for GMM when the phase information is assumed known. For tests in real operation when the phase information is imaged from the given narrowband signal the preference comparison goes up to 84% versus 16%. A correct phase recovery can further increase the BWE performance for the proposed DNN method.

Representation Learning for Single-Channel Source Separation and Bandwidth Extension

“…Motivated by the success of GSNs and SPNs on the task of image completion [9], [21], we use both models to complete the high frequency parts of log-spectrograms, lost due to the telephone bandpass filter. In a recent work deep neural networks (DNNs) have been proposed for ABE [36]. First, unsupervised pre-training of RBMs is performed using the log-spectrum of the narrow-band signal as input.…”

“…We use popular generative models from representation learning including Gauss Bernoulli restricted Boltzmann machines (GBRBMs) [13], conditional Gauss Bernoulli restricted Boltzmann machines (CGBRBMs) [32], higher order contractive autoencoders (HCAEs) [16], SPNs, and GSNs. Furthermore, a rectifier MLP is applied to both tasks to facilitate a comparison of the results in this work to [33], [34], [35], [36].…”

A reverberation-time-aware DNN approach leveraging spatial information for microphone array dereverberation

“…It also avoids a brute-force approach that always selects the same number of frames for each microphone, which may eventually leads to a too big DNN input size as the number of microphone increases and thus causes dramatic performance deterioration [28]. The proposed approach has a low computational complexity, since only a single DNN is used.…”

High Frequency Bandwidth Extension for Audio Codec in Mobile Surveillance Device

“…In order to be compatible with existing narrow-band speech signals and find the losing high frequency signals, high-frequency bandwidth extension (BWE) is applied to extend the narrow band speech signals to wide band. The method is utilizing the features of low frequency signals to predict the ones of high frequency signals based on statistical correlation between high and low frequency signals, so as to reconstruct the lost high frequency signals [2][3][4][5][6]. And it needs no side information, is called -blind‖ bandwidth extension.…”