Bone-conducted speech is less susceptible to ambient noise interference, but it suffers from poor speech quality due to the limited bandwidth. In this letter, we propose a U-Net-like network for the restoration of bone-conducted speech in the time domain. The proposed network consists of residual-connected one-dimensional convolutions and shifted window-based attention modules, which can model long-term dependencies crucial in speech processing. We find that the prevalent time-domain${{l}_{1}}$loss may be insufficient for the generation of high-frequency information absent in bone-conducted speech. To address this issue, we propose to utilize the generalized energy distance loss based on multi-scale Mel spectrograms as the objective function. Experimental results on the ESMB dataset validate the efficacy of our proposed method in restoration of bone-conducted speech. The proposed approach significantly outperforms two recent time-domain benchmarks, DPT-EGNet and EBEN, in terms of PESQ and STOI metrics.

Bone-conducted speech is not susceptible to background noise but suffers from poor speech quality and intelligibility due to the limited bandwidth. This paper proposes a two-stage approach to restore the quality of bone-conducted speech, namely, bandwidth extension and speech vocoder. In the first stage, a deep neural network is trained to learn mappings from a low-resolution representation of the bone-conducted speech, i.e., log Mel-scale spectrogram, to that of the air-conducted speech, which extends the bandwidth of the bone-conducted speech. In the second stage, a speech vocoder is employed to transform the extended log Mel-scale spectrogram of the bone-conducted speech back to time-domain waveforms. Due to the many-to-many correspondence between the air-conducted and bone-conducted speech, supervised learning may not be the best training protocol for the bone-conducted/air-conducted feature mapping. We thus propose to leverage adversarial training to further improve the bandwidth extension performance in the first stage. The two stages are decoupled and can be trained independently. The vocoder is trained on a large multi-speaker dataset and can generalize well to unknown speakers. Also, the vocoder can help to remedy the spectral artifacts introduced in the bandwidth extension stage. Objective and subjective evaluations on ESMB dataset show that the proposed two-stage system substantially outperforms existing bone-conducted speech enhancement systems.

Although much progress has been made in synthetic speech detection, there lacks comprehensive analysis of the essential differences between spoofed and genuine speech. We here utilize supervised contrastive loss originated from contrastive learning as an analytical tool to characterize the class similarity structure of ASVspoof 2019 logical access (LA) dataset, which shows that an ideal back-end classifier for synthetic speech detection should have the ability to capture long-term dependencies. Recently, Transformer has been found to have an excellent ability in learning long-term dependencies of input data. We hence propose a back-end classifier based on Transformer Encoder for synthetic speech detection. Convolution blocks are added before the Transformer Encoder, which leverages inductive biases to improve the generalization ability. Compared to two-dimensional convolution, one-dimensional convolution makes better architectural assumptions about the input speech features, which helps with modeling long-term dependencies and decreases the risk of overfitting. The proposed Transformer combined with one-dimensional convolution has fewer parameters than most existing back-end classifiers, and achieves an equal error rate of 1.06% and a minimum tandem detection cost function metric of 0.0345 when evaluated on ASVspoof 2019 LA dataset, which is one of the best models reported in the literature.