DNN-based speech mask estimation for eigenvector beamforming

This paper presents a novel approach for estimating autoregressive (AR) model parameters using deep neural network (DNN) in the AR-Wiener filtering speech enhancement. Unlike conventional DNN that predicts one kind of target, the DNN used in this paper is trained to predict the AR model parameters of speech and noise simultaneously at offline stage. We train this network by minimizing the Euclidean distance between the output of DNN and the AR model parameters of clean speech and noise. At online stage, the acoustic features are first extracted from noisy speech as the input of the DNN. Then, AR model parameters of speech and noise are estimated by the DNN simultaneously. Finally, the Wiener filter is constructed by the AR model parameters of speech and noise. However, the AR model parameters only models the spectral shape not the spectral details, there are still some residual noise between the harmonics. In order to solve this problem, we introduce the speech-presence probability (SPP), that is, in the test stage, the SPP is estimated and is used to update the Wiener filter. The experimental results show that our approach has higher performance compared with some existing approaches.

In this work, we propose ShallowDeepNet, a novel system architecture that includes a shallow and a deep neural network. The shallow neural network has the duty of data preprocessing and generating adversarial samples. The deep neural network has the duty of understanding data and information as well as detecting adversarial samples. The deep neural network gets its weights from transfer learning, adversarial training, and noise training. The system is examined on the biometric (fingerprint and iris) and the pharmaceutical data (pill image). According to the simulation results, the system is capable of improving the detection accuracy of the biometric data from 1.31% to 80.65% when the adversarial data is used and to 93.4% when the adversarial data as well as the noisy data are given to the network. The system performance on the pill image data is increased from 34.55% to 96.03% and then to 98.2%, respectively. Training on different types of noise can benefit us in detecting samples from unknown and unseen adversarial attacks. Meanwhile, the system training on the adversarial data as well as noisy data occurs only once. In fact, retraining the system may improve the performance further. Furthermore, training the system on new types of attacks and noise can help in enhancing the system performance.

Ensemble of deep neural networks using acoustic environment classification for statistical model-based voice activity detection

Determining the structure of a protein given its sequence is a challenging problem. Deep learning is a rapidly evolving field which excels at problems where there are complex relationships between input features and desired outputs. Deep Neural Networks have become popular for solving problems in protein science. Various deep neural network architectures have been proposed including deep feed-forward neural networks, recurrent neural networks and more recently neural Turing machines and memory networks. This article provides a short review of deep learning applied to protein prediction problems.

Handwritten keyword spotting using deep neural networks and certainty prediction

Deep Neural Networks (DNNs) have made significant improvements to reach the desired accuracy to be employed in a wide variety of Machine Learning (ML) applications. Recently the Google Brain's team demonstrated the ability of Capsule Networks (CapsNets) to encode and learn spatial correlations between different input features, thereby obtaining superior learning capabilities compared to traditional (i.e., non-capsule based) DNNs. However, designing CapsNets using conventional methods is a tedious job and incurs significant training effort. Recent studies have shown that powerful methods to automatically select the best/optimal DNN model configuration for a given set of applications and a training dataset are based on the Neural Architecture Search (NAS) algorithms. Moreover, due to their extreme computational and memory requirements, DNNs are employed using the specialized hardware accelerators in IoT-Edge/CPS devices. In this paper, we propose NASCaps, an automated framework for the hardware-aware NAS of different types of DNNs, covering both traditional convolutional DNNs and CapsNets. We study the efficacy of deploying a multi-objective Genetic Algorithm (e.g., based on the NSGA-II algorithm). The proposed framework can jointly optimize the network accuracy and the corresponding hardware efficiency, expressed in terms of energy, memory, and latency of a given hardware accelerator executing the DNN inference. Besides supporting the traditional DNN layers, our framework is the first to model and supports the specialized capsule layers and dynamic routing in the NAS-flow. We evaluate our framework on different datasets, generating different network configurations, and demonstrate the tradeoffs between the different output metrics. We will open-source the complete framework and configurations of the Pareto-optimal architectures at https://github.com/ehw-fit/nascaps.

Speech enhancement for non-stationary noise environments

The natural history of classification systems for the organic mental disorders from ICD-8 through ICD-10 and DSM-III is presented. Out of a total of 31 cases the 6 with a main diagnosis of the F0-section of ICD-10 are described and field trial results analysed in terms of classification errors, goodness of fit, difficulty, and interrater-reliability. Specific categories (instead of "collective categories") were used more frequently and fewer classification errors made on the basis of ICD-10 as compared with ICD-9. The authors suggest that an operationalized differentiation of dementia as one of the core categories would further increase the reliability and clinical practicability of this section.

Like commonly used electromyography (EMG), mechanomyogram (MMG) is another type of important signal that represents muscle activities in the form of muscle vibrations. A number of studies have showed that both EMG and MMG have great prospect in the application of prosthetic control system. Most previous studies using either EMG or MMG signals mainly focus on improving the control performance of prosthetic system or investigating the signal's metrics. The differences of these two types of signals associated with the muscle activities remain unclear in classifying multiple motion classes. In addition, limb position variation may be one of important factors which would cause classification performance degradation. In this study, we analyzed and compared the metrics of these two kinds of biomedical signals by motion classification and position classification in terms of classification error. The results showed that the similar motion classification performance could be achieved using both EMG and MMG, having the classification errors of 7.4% and 9.9%, respectively. When it comes to position classification, MMG achieved significantly lower classification error of 0.7% versus 16.5% when using EMG.

Prediction of intraoperative complexity from preoperative patient data for laparoscopic cholecystectomy

Ultra wide-band (UWB) radio uses pulses of very short duration, thereby spreading the energy of the radio signal very thinly over several Gigahertz bandwidth. The power spectral density (PSD) of UWB signals consists of a continuous component and a discrete component. The discrete component presents greater interference to narrow-band communication systems than the continuous component. Multi-band UWB, recently proposed to IEEE 802.15.3a, divides the working spectrum into sub-bands, in each of which one waveform is used for the impulse. In this paper, we devise an efficient mechanism to suppress the discrete PSD component of multi-band UWB.

Phase unwrapping based packet loss concealment using deep neural networks

A wavelet packet deep learning model for Energy-Based structural collapse assessment under Earthquake-Fire Scenarios: Framework and hybrid simulation

The speech enhancement effect of traditional deep learning algorithms is not ideal under low signal-to-noise ratios (SNR). Skip connections-deep neural network (Skip-DNN) improves the traditional deep neural network (DNN) by adding skip connections between each layer of the neural network to solve the degradation problem of DNN. In this paper, the Multiresolution Cochleagram (MRCG) features in the gammachirp transform domain are denoised to obtain the improved MRCG (I-MRCG). The noise reduction method adopts the Minimum Mean-Square Error Short-Time Spectral Amplitude Estimator (MMSE-STSA) and takes I-MRCG as the input feature and Skip-DNN as the training network to improve the speech enhancement effect of the model. This paper also proposes an improved source-to-distortion ratio (SDR) loss function. When the loss function uses the improved SDR, it will improve the performance of Skip-DNN speech enhancement model. The experiments in this paper are performed on the Edinburgh dataset. When using I-MRCG as the input feature of Skip-DNN, the average perceptual evaluation of speech quality (PESQ) is 2.9137, and the average short-time objective intelligibility (STOI) is 0.8515. Compared with MRCG as Skip-DNN input features, the improvements are 0.91% and 0.71%, respectively. When the improved SDR is used as the loss function of the speech model, the average PESQ is 2.9699 and the average STOI is 0.8547. Compared with other loss functions, the improved SDR has a better enhancement effect when used as the loss function of the speech enhancement model.

This paper presents a front-end speech enhancement approach to robust speech recognition in automotive environments. It combines hybrid voice activity detection (VAD), relative transfer function (RT-F) based generalized sidelobe cancelation, and single-channel post filtering to enhance the speech signal of interest, thereby improving the robustness of speech recognition. First, we choose four typical driving scenarios, which include most of the noise types in automobiles to record training data. The recorded data is then used to train deep neural network models (DNNs) for both speech and noise. The trained DNNs are subsequently used to estimate the speech presence probability on a frame-by-frame basis. This speech presence probability is then combined with the output of an energy-based VAD to form a hybrid VAD, which serves as the basis for the rest components of the speech enhancement system, including RTF estimation, adaptive beamforming, and post-filtering. Experiments are conducted in real automotive environments. The results show that the developed method can significantly improve the performance of both VAD and automatic speech recognition (ASR).