Shingo Uenohara

A Study on the Optimum Number of Training Data in Snore Activity Detection Using SVM

Health promotion and maintenance is becoming increasingly important and depends on three elements: nutrition, exercise, and rest (sleep). In the present study, focusing on sleep, we develop a smartphone-based system based on snore activity detection to investigate day-to-day variations in the sleep state, which does not require dedicated hardware. Here, we analyze the number of training data required for snore activity detection using a support vector machine (SVM), and we consider ways to improve detection performance. The sound pressure level and mel-frequency cepstrum coefficients are calculated from sleep sound data obtained using a smartphone. Snore activity detection is performed by machine learning using an SVM with a linear kernel, the SVM is trained by labeled acoustic features, and the trained SVM models are used to detect snore activity. In general, the accuracy of the generated models increases with the increasing number of training data in the learning algorithm, which in turn increases the computational cost, therefore, a balance between accuracy and cost efficiency is much required. We investigate the relation between the detection rate and the number of training data in snore activity detection, and we propose the optimum number of data required for learning.

Reducing Computational Complexity of Multichannel Nonnegative Matrix Factorization Using Initial Value Setting for Speech Recognition

In this paper, we propose efficient the number of computational iteration method of MNMF for speech recognition. The proposed method initializes estimates MNMF algorithm with the estimated spatial correlation matrix reduces the number of iteration of updates algorithm. The experiment result shows that our method reduced the computational complexity of MNMF.

Evaluating Classification Methods in Snore Activity Detection

Health improvement and maintenance is becoming increasingly important and depends on three elements: nutrition, exercise, and rest (sleep). In the present work, we focus on sleep and develop a smartphone-based system based on snore activity detection to investigate day-to-day variations in the sleep states, which requires no dedicated hardware. Snore activity detection is performed using classification methods to detect the snore activity using acoustic features. As acoustic features, the sound pressure level and mel-frequency cepstrum coefficients are calculated from the sleep sound data obtained using a smartphone. In this study, we evaluated the performance of three classification methods, support vector machine, multi-kernel learning using support vector machine and deep learning in snore activity detection.

Evaluation of parallel measurement methods for multichannel impulse responses

This study evaluates the multiple maximum length sequence (MMLS) and multiple exponential sweep (MES) methods for performing parallel measurements of multi-channel acoustic systems. The parallel measurements of multiple loudspeakers can be efficiently performed using the MMLS or MES method. Experiments were conducted to evaluate the accuracy, measurement duration, and computing time for the two methods. The results show that the MMLS method was the same as the MES method. Although the measurement duration was shorter for the MMLS method compared with the MES method, the computing time was longer.

Parameters of Noise Suppression Based on Wavelet Transform for Phonocardiographic Signals

While heart disease is one of the three major diseases, only well-qualified doctors can evaluate phonocardiographic signals. This calls for an easily available system that can automatically diagnose phonocardiographic signals. When recording in a room, suppression is required as these signals are heavily contaminated by noise from various sources such as air conditioners and fans. Wavelet transform is one method for denoising phonocardiographic signals, but appropriate parameters are required. In this study, we investigated both normal and abnormal phonocardiographic signals to determine the appropriate use of single and multilevel thresholds and the best types of wavelet functions. The experiment results show that the most appropriate wavelet function is Symlet14 and multilevel thresholding is best for low SNRs.

Identification of phonocardiographic signals using noise suppression based on wavelet transform

While heart disease is one of the three major diseases, only well-qualified doctors can evaluate phonocardiographic signals. However, phonocardiographic signals are not always used in healthcare because only a few professionals are experts in evaluating phonocardiographic signals. We need to develop an easily available system that can automatically diagnose phonocardiographic signals. In previous study, phonocardiographic signals are typically analyzed using wavelet transform to match and extract the characteristics of known normal and abnormal phonocardiographic signals. However, everyday noises such as lung and breath sounds, environmental noises, and blood flow noises may contaminate these signals and hinder analysis. We have previously proposed noise suppression using wavelet transform for phonocardiographic signals. In this study, we compare the proposed method and spectral subtraction to identify the kind of phonocardiographic signals. The experiment results show that the proposed method provides be...

Comparison of the multiple maximum length sequences and multiple exponential sweeps method for parallel measurements of impulse responses

The measurement of room impulse responses is a central problem in audio signal processing, particularly for spatial audio rendering and sound field reproduction applications. The calibration of modern rendering systems requires the knowledge of the room impulse responses between the loudspeakers and several possible listener positions, in order to compensate for the characteristics of the loudspeakers and the room. The parallel measurement of multiple loudspeakers can be efficiently performed using the multiple maximum length sequence (MMLS) or the multiple exponential sweep (MES) method. This study evaluates the MMLS and MES methods for performing parallel measurements of multi-channel acoustic systems. Moreover, we apply noise reduction to improve their measurement accuracies. Experiments were conducted to evaluate the accuracy, measurement duration, and computing time for the two methods. The results show that the MMLS method was as accurate as the MES method. Although the measurement duration was shor...

Multi-channel non-negative matrix factorization with binary mask initialization for automatic speech recognition

Non-negative Matrix Factorization (NMF) factorizes a non-negative matrix into two non-negative matrices. In the field of acoustics, multichannel expansion has been proposed to consider spatial information for sound source separation. Conventional multi-channel NMF has a difficulty in an initial-value dependency of the separation performance due to local minima. This paper proposes initial value settings by using binary masking based sound source separation whose masks on the time frequency domain are calculated from the time difference of arrival of each source. The proposed method calculates initial spatial correlation matrices using separated sources by binary masking. The music separation experiments confirmed that the separation performance of the proposed method was better than that of the conventional method. In addition, we evaluated initial value settings by using binary masking for automatic speech recognition (ASR) tasks in noisy environments. The ASR experiments confirmed that appropriate initi...

Performance improvement of snore activity detection in a noisy environment

Health promotion and maintenance is becoming increasingly important and depends on three elements: nutrition, exercise, and rest (sleep). In the present work, focusing on sleep, we develop a smartphone-based system based on snore activity detection to investigate day-to-day variations in the sleep state, which does not require dedicated hardware. Snore activity detection is performed by machine learning using a support vector machine (SVM) with a linear kernel; the SVM is trained by labeled acoustic features, and the trained SVM models are used to detect snore activity. As acoustic features, The sound pressure level and mel-frequency cepstrum coefficients are calculated from sleep sound data obtained using a smartphone. In this paper, we investigated the effects of adding sleep environment noise recorded before sleep to the training set in snore activity detection, and we considered ways to improve detection performance. Performance comparison among the conventional method of SVM and the proposed method w...

Snore activity detection using smartphone sensors

In this paper, we analyze the effects of ambient noise on snore activity detection, and consider ways to improve detection performance. A smartphone is used to obtain sleep sound data, from which the acoustic features of sound pressure level (SPL) and Mel-frequency cepstrum coefficients (MFCC) are calculated. Snore activity detection is performed by machine learning using a support vector machine (SVM) with a linear kernel. The SVM is trained by the labeled acoustic features, and the trained SVM models are used to detect snore activity. Adding ambient noise recorded before sleep to the training set is expected to improve detection performance. Experimental results showed that an improvement in detection performance from F-measure of 0.75 to 0.81 using SPL, from F-measure of 0.62 to 0.62 using MFCC, from F-measure of 0.69 to 0.74 using SPL and MFCC on average.