The flexible vibrational sensor (FVS) has the potential to become a popular wearable communication device because of its natural noise shielding characteristics and soft materials. However, FVS speech faces a severe loss of frequency components. To improve speech quality, a time-domain neural network model based on the dual-path transformer combined with equalization-generation components prediction (DPT-EGNet) is proposed. More specifically, the DPT-EGNet consists of five modules, namely the pre-processing module, dual-path transformer module, equalization module, generation module, and post-processing module. The dual-path transformer module is leveraged to extract the local and global contextual relationship of long-term speech sequences, which is extremely beneficial for inferring the missing components. The equalization and generation modules are designed according to the characteristics of FVS speech, which further improve the speech quality by simulating the inversion process of the speech distortion. The experimental results demonstrate that the proposed model effectively improves the quality of FVS speech; the average perceptual evaluation of speech quality (PESQ), short-time objective intelligibility (STOI), and composite measure for overall speech quality (COVL) scores of three males and three females are relatively increased by 64.19%, 29.63%, and 101.37%, which is superior to other baseline models developed in different domains. The proposed model also has significantly lower complexity than the others.

Even though audio replay detection has improved in recent years, its performance is known to severely deteriorate with the existence of strong background noises. Given the fact that different frames of an utterance have different impacts on the performance of spoofing detection, this paper introduces attention-based long short-term memory (LSTM) to extract representative frames for spoofing detection in noisy environments. With this attention mechanism, the specific and representative frame-level features will be automatically selected by adjusting their weights in the framework of attention-based LSTM. The experiments, conducted using the ASVspoof 2017 dataset version 2.0, show that the equal error rate (EER) of the proposed approach was about 13% lower than the constant Q cepstral coefficients-Gaussian mixture model (CQCC-GMM) baseline in noisy environments with four different signal-to-noise ratios (SNR). Meanwhile, the proposed algorithm also improved the performance of traditional LSTM on audio replay detection systems in noisy environments. Experiments using bagging with different frame lengths were also conducted to further improve the proposed approach.

In recent years, neural network-based voice conversion methods have been rapidly developed, and many different models and neural networks have been applied in parallel voice conversion. However, the over-smoothing of parametric methods [e.g., bidirectional long short-term memory (BLSTM)] and the waveform collapse of neural vocoders (e.g., WaveNet) still have negative impacts on the quality of the converted voices. To overcome this problem, we propose a BLSTM and WaveNet-based voice conversion method cooperated with waveform collapse suppression by post-processing. This method firstly uses BLSTM to convert the acoustic features between parallel speakers, and then synthesizes pre-converted voice with WaveNet. Subsequently, several alternative iterations of BLSTM post-processing is performed, and the final converted voice is generated by WaveNet. The proposed method can directly generate converted audio waveforms and avoid the waveform-collapsed speech caused by a single WaveNet generation effectively. The experimental results indicate that acoustic features trained by using the BLSTM network could achieve better results than conventional baselines. From our experiments on VCC2018, the usage of WaveNet could alleviate the problem of over-smoothing, which contributes to improving the similarity and naturalness of the final results of voice conversion.

Non-acoustic sensors are widely used in speech signal processing tasks, and their immunity to the background acoustic noise shows great benefits to traditional speech enhancement. To avoid using acoustic speech disturbed by strong noise, spectra mapping from throat microphone (TM) speech to acoustic microphone (AM) speech has been studied. However, there is a distinguished difference between the spectra of the two kinds of speech, and the mapping relationship is different in the low-band and high-band spectra, which limits the performance of the traditional full-band spectra mapping model. In this paper, to improve the perceived quality and intelligibility of TM speech, we investigate the low-band and high-band spectral structure between TM and AM speech, respectively, and propose a spectra-based band-division mapping framework for TM speech enhancement based on the investigation. In the framework, the low-band target spectra of AM speech are mapped based on the equalization method, and the high-band spectra are mapped from the full-band TM speech spectra, which are lack of high-frequency components. The overall framework can be viewed as a combination of spectra equalization in the low band and spectra generation in the high band. Both the objective and subjective evaluation results show clear advantages over the existing TM speech enhancement method based on the full-band spectra mapping.