A method for enhancing speech and warning signals based on parallel convolutional neural networks in a noisy environment.

BACKGROUND: Digital hearing aids are based on technology that amplifies sound and removes noise according to the frequency of hearing loss in hearing loss patients. However, within the noise removed is a warning sound that alert the listener; the listener may be exposed to danger because the warning sound is not recognized. OBJECTIVE: In this paper, a deep learning model was used to improve these limits and propose a method to distinguish the warning sound in speech signals mixed with noise. In addition, the improved speech and warning sound were derived by removing noise present in the classification sound signals. METHODS: To classify the sound dataset, an adaptive convolution filter that changes according to two signals is proposed. The proposed convolution filter is applied to the PCNNs model to analyze the characteristics of the time and frequency domains of the dataset and classify the presence or absence of warning sound. In addition, the CEDN model was used to improve the intelligibility of the warning and the speech in the signal based on the warning sound classification from the proposed PCNNs model. RESULTS: Experimental results show that the PCNNs model using the proposed multiplicative filters is efficient for analyzing sound signals with complex frequencies. In addition, the CEDN model was used to improve the intelligibility of the warning and the speech in the signal based on the warning sound classification from the proposed PCNNs model. CONVLUSION: We confirmed that the PCNN model with the proposed filter showed the highest training rate, lowest error rate, and the most stable results. In addition, the CEDN model confirmed that speech and warning sounds were recognized, but it was confirmed that there was a limitation in clearly recognizing speech as the noise ratio increased.

Noise reduction algorithm with the soft thresholding based on the Shannon entropy and bone-conduction speech cross- correlation bands.

BACKGROUND: The conventional methods of speech enhancement, noise reduction, and voice activity detection are based on the suppression of noise or non-speech components of the target air-conduction signals. However, air-conduced speech is hard to differentiate from babble or white noise signals. OBJECTIVE: To overcome this problem, the proposed algorithm uses the bone-conduction speech signals and soft thresholding based on the Shannon entropy principle and cross-correlation of air- and bone-conduction signals. METHODS: A new algorithm for speech detection and noise reduction is proposed, which makes use of the Shannon entropy principle and cross-correlation with the bone-conduction speech signals to threshold the wavelet packet coefficients of the noisy speech. RESULTS: The proposed method can be get efficient result by objective quality measure that are PESQ, RMSE, Correlation, SNR. CONCLUSION: Each threshold is generated by the entropy and cross-correlation approaches in the decomposed bands using the wavelet packet decomposition. As a result, the noise is reduced by the proposed method using the MATLAB simulation. To verify the method feasibility, we compared the air- and bone-conduction speech signals and their spectra by the proposed method. As a result, high performance of the proposed method is confirmed, which makes it quite instrumental to future applications in communication devices, noisy environment, construction, and military operations.

Implementation of automatic external defibrillator using real time ventricular fibrillation detecting algorithm based on time domain analysis.

The increase in mortality associated with arrhythmia is an inevitable problem of modern society such as westernized eating habits and an increase in stress due to industrialization, and the related social costs are increasing. In this regard, the supply of automatic external defibrillator (AED) used outside hospitals is increasing mainly in public institutions, and AED is a medical practice performed by non-medical personnel. Therefore, studies on arrhythmia detection algorithm to make accurate clinical judgment for proper use are increasing. In this paper, we propose a time domain analysis method to detect arrhythmia in real time and implement AED by porting it to programmable gate array and digital signal processor. The analysis of the phase domain improves the detection rate of R-peak using the differentiated electrocardiogram (ECG) waveform rather than the existing ECG waveform and makes it easy to distinguish the normal ECG from the arrhythmia signal in the phase domain. The proposed algorithm was verified by simulation using Labview and ModelSim, and it was verified that the proposed algorithm works effectively by performing animal experiments using the implemented AED.

Mastication noise reduction method for fully implantable hearing aid using piezo-electric sensor.

BACKGROUND: Fully implantable hearing devices (FIHDs) can be affected by generated biomechanical noise such as mastication noise. OBJECTIVE: To reduce the mastication noise using a piezo-electric sensor, the mastication noise is measured with the piezo-electric sensor, and noise reduction is practiced by the energy difference. METHODS: For the experiment on mastication noise, a skull model was designed using artificial skull model and a piezo-electric sensor that can measure the vibration signals better than other sensors. A 1 kHz pure-tone sound through a standard speaker was applied to the model while the lower jawbone of the model was moved in a masticatory fashion. RESULTS: The correlation coefficients and signal-to-noise ratio (SNR) before and after application of the proposed method were compared. It was found that the signal-to-noise ratio and correlation coefficients increased by 4.48 dB and 0.45, respectively. CONCLUSION: The mastication noise is measured by piezo-electric sensor as the mastication noise that occurred during vibration. In addition, the noise was reduced by using the proposed method in conjunction with MATLAB. In order to confirm the performance of the proposed method, the correlation coefficients and signal-to-noise ratio before and after signal processing were calculated. In the future, an implantable microphone for real-time processing will be developed.

Voice activity detection algorithm using perceptual wavelet entropy neighbor slope.

This paper presents a voice activity detection (VAD) approach using a perceptual wavelet entropy neighbor slope (PWENS) in a low signal-to-noise (SNR) environment and with a variety of noise types. The basis for our study is to use acoustic features that have large entropy variance for each wavelet critical band. The speech signal is decomposed by the proposed perceptual wavelet packet decomposition (PWPD), and the VAD function is extracted by PWENS. Finally, VAD is decided by the proposed VAD decision rule using two memory buffers. In order to evaluate the performance of the VAD decision, many speech samples and a variety of SNR conditions were used in the experiment. The performance of the VAD decision is confirmed using objective indexes such as a graph of the VAD decision and the relative error rate.

Individual tooth region segmentation using modified watershed algorithm with morphological characteristic.

In this paper, a new method for individual tooth segmentation was proposed. The proposed method is composed of enhancement and extraction of boundary and seed of watershed algorithm using trisection areas by morphological characteristic of teeth. The watershed algorithm is one of the conventional methods for tooth segmentation; however, the method has some problems. First, molar region detection ratio is reduced because of oral structure features that is low intensities in molar region. Second, inaccurate segmentation occurs in incisor region owing to specular reflection. To solve the problems, the trisection method using morphological characteristic was proposed, where three tooth areas are made using ratio of entire tooth to each tooth. Moreover, the enhancement is to improve the intensity of molar using the proposed method. In addition, boundary and seed of watershed are extracted using trisection areas applied other parameters each area. Finally, individual tooth segmentation was performed using extracted boundary and seed. Furthermore, the proposed method was compared with conventional methods to confirm its efficiency. As a result, the proposed method was demonstrated to have higher detection ratio, better over segmentation, and overlap segmentation than conventional methods.