Speech emotion recognition has drawn extensive attention in recent years. We propose deep learning (DL)-based speech emotion recognition using synthetic bone-conducted (BC) speech. In our proposed model, air-conducted(AC) speech is transformed to BC speech using an infinite impulse response (IIR) filter. Data augmentation techniques are utilized and the parameters of convolutional neural network (CNN) models are modified to enhance the accuracy of the proposed model. Simulation results demonstrate that the proposed model outperforms the existing models in terms of recognition accuracy for BC speech. The accuracy of the proposed model is 72.50% for BC speech, whereas that of the existing model is 69.83% for AC speech. This is because BC speech can enhance low-frequency components, which is important for recognizing emotions.

A distributed estimation algorithm is discussed referring to the single-input multiple-output (SIMO) channel model over a wireless sensor network (WSN) to estimate the transmitted data signal blindly. We find a sensor node in the best local sensor network by minimizing the total average deviation from the minimum signal power in the WSN. The signal at the sensor node is used as the input signal for the distributed blind equalizer. We assume two classes of channels in the SIMO channel model, namely, Class 1 of the common channel conditions with different noise variances and Class 2 of the different channel conditions with common noise variances, based on raised cosine channel models. Computer simulations confirm the excellent performance of the proposed method through the evaluation of mean square error and symbol error rate.

Underwater acoustic (UWA) communication is an essential technology for networking underwater sensors and drones to support maritime activities. However, real-time data transfer among mobile nodes faces technical difficulties due to the severe nature of the UWA channel, which has large delay and Doppler spreads. Furthermore, the performance of UWA communication may be severely impaired by marked channel changes between line-of-sight (LoS) and non-line-of-sight (NLoS) environments. The goal of this work is to explore the UWA channel under LoS and NLoS conditions and evaluate their effect on UWA communication in terms of communication quality. Specifically, we measured the delay and Doppler spreads of the UWA channel by switching LoS/NLoS conditions. We also investigated the relationship between the signal-to-noise ratio (SNR) and bit error rate (BER) for UWA communication using Doppler-resilient orthogonal signal division multiplexing (D-OSDM). Experimental results showed that the delay spread of the UWA channel and received signal power markedly changes depending on LoS/NLoS conditions. Specifically, the delay spread in the LoS environment was about 10 ms, while that in the NLoS environment was much larger or difficult to observe owing to signal power attenuation. Furthermore, SNR in the NLoS environment was approximately 15 dB less than that in the LoS environment, resulting in increased BER. The obtained results suggest that adaptive signal power control is necessary to realize stable UWA communication in shallow water areas where there is a mixture of LoS and NLoS condition.

Pneumonia is a major cause of death in Japan, with aspiration pneumonia being the most common cause. However, no quantitative index exists for non-invasive swallowing assessment. Non-invasive, quantitative methods of assessing swallowing have been studied using pharyngeal ultrasound videos, surface electromyography, and swallowing sounds. However, each signal has limitations regarding the temporal and spatial ranges of swallowing movements that can be measured. This study develops a signal processing method to capture detailed swallowing behaviour based on the information obtained from each signal by recording three types of signals and comparing them on the same time axis.

The evaluation of swallowing via swallowing sounds for preventing aspiration pneumonia has been reported previously; however, the distinction between masticatory and swallowing sounds has not been studied. Therefore, this study aimed at distinguishing mastication and swallowing based on the sounds produced during eating. Daubechies wavelet analysis of the sounds generated when eating samples of almonds and white rice of different hardness and viscosity values was performed. Different features were found to be present in the masticatory and swallowing sounds. During chewing sounds, the frequency changes to a higher frequency. This feature was found in 78% of the masticatory sounds obtained in the experiments. Swallowing sounds, on the other hand, change to a lower frequency. This feature was found in all swallowing sounds obtained in the experiment. Based on this feature, distinguishing between masticatory and swallowing sounds may be possible.

Timbre conversion of musical instrument sounds, utilizing deep neural networks (DNNs), has been extensively researched and continues to generate significant interest in the development of more advanced techniques. We propose a novel algorithm for timbre conversion that utilizes a variational autoencoder. However, this system must be capable of predicting the amplitude spectrogram from the melfrequency cepstrum coefficient (MFCC). This research aims to build a DNN-based decoder that utilizes the MFCC and time-frame-wise total amplitude as inputs to predict the amplitude spectrogram. Experiments conducted using a musical instrument sound dataset show that a decoder incorporating bidirectional long short-term memory yields accurate predictions of amplitude spectrograms.

In asset management businesses, such as portfolio management, it is common to operate in the medium to long term due to the increase in operational burden and transaction costs. However, to compose a longer-term model, the number of usable learning data decreases due to the larger observation interval of the data; hence, the model performance declines. To solve this problem, in this study, a data augmentation was conducted by the combined use of data of multiple time scales, and confirm its effectiveness to keep a better generalization ability of trained models even if the target task of machine-learning methods is longer time scale. In addition, portfolio management was conducted using the constructed model.