The Technique of English Word Syllable Division in Speech Synthesis Based on Neural Network

With the increasing maturity of speech synthesis technology, on the one hand, it has been more and more widely used in people’s lives; on the other hand, it also brings more and more convenience to people. The requirements for speech synthesis systems are getting higher and higher. Therefore, advanced technology is used to improve and update the accent recognition system. This paper mainly introduces the word stress annotation technology combined with neural network speech synthesis technology. In Chinese speech synthesis, prosodic structure prediction has a great influence on naturalness. The purpose of this paper is to accurately predict the prosodic structure, which has become an important problem to be solved in speech synthesis. Experimental data show that the average error of samples in the network training process is lel/85, and the minimum value of the training error after 500 steps is 0.00013127, so the final sample average error is lel = 85 ∗ 0.0013127 = 0.112 < 0.5, and use the deep neural network (DNN) to train different parameters to obtain the conversion model, and then synthesize these conversion models, and finally achieve the effect of improving the synthesized sound quality.

AlertNet: Deep convolutional-recurrent neural network model for driving alertness detection

Drowsy driving is one of the major problems which has led to many road accidents. Electroencephalography (EEG) is one of the most reliable sources to detect sleep on-set while driving as there is the direct involvement of biological signals. The present work focuses on detecting driver’s alertness using the deep neural network architecture, which is built using ResNets and encoder-decoder based sequence to sequence models with attention decoder. The ResNets with the skip connections allow training the network deeper with a reduced loss function and training error. The model is built to reduce the complex computations required for feature extraction. The ResNets also help in retaining the features from the previous layer and do not require different filters for frequency and time-invariant features. The output of ResNets, the features are input to encoder-decoder based sequence to sequence models, built using Bi-directional long-short memories. Sequence to Sequence model learns the complex features of the signal and analyze the output of past and future states simultaneously for classification of drowsy/sleepstage-1 and alert stages. Also, to overcome the unequal distribution (class-imbalance) data problem present in the datasets, the proposed loss functions help in achieving the identical error for both majority and minority classes during the raining of the network for each sleep stage. The model provides an overall-accuracy of 87.92% and 87.05%, a macro-F1-core of 78.06%, and 79.66% and Cohen's-kappa score of 0.78 and 0.79 for the Sleep-EDF 2013 and 2018 data sets respectively.

A Sparsity-Driven Backpropagation-Less Learning Framework Using Populations of Spiking Growth Transform Neurons

Growth-transform (GT) neurons and their population models allow for independent control over the spiking statistics and the transient population dynamics while optimizing a physically plausible distributed energy functional involving continuous-valued neural variables. In this paper we describe a backpropagation-less learning approach to train a network of spiking GT neurons by enforcing sparsity constraints on the overall network spiking activity. The key features of the model and the proposed learning framework are: (a) spike responses are generated as a result of constraint violation and hence can be viewed as Lagrangian parameters; (b) the optimal parameters for a given task can be learned using neurally relevant local learning rules and in an online manner; (c) the network optimizes itself to encode the solution with as few spikes as possible (sparsity); (d) the network optimizes itself to operate at a solution with the maximum dynamic range and away from saturation; and (e) the framework is flexible enough to incorporate additional structural and connectivity constraints on the network. As a result, the proposed formulation is attractive for designing neuromorphic tinyML systems that are constrained in energy, resources, and network structure. In this paper, we show how the approach could be used for unsupervised and supervised learning such that minimizing a training error is equivalent to minimizing the overall spiking activity across the network. We then build on this framework to implement three different multi-layer spiking network architectures with progressively increasing flexibility in training and consequently, sparsity. We demonstrate the applicability of the proposed algorithm for resource-efficient learning using a publicly available machine olfaction dataset with unique challenges like sensor drift and a wide range of stimulus concentrations. In all of these case studies we show that a GT network trained using the proposed learning approach is able to minimize the network-level spiking activity while producing classification accuracy that are comparable to standard approaches on the same dataset.

Classification of Teachers and Lecturers Engagement on Webinar during the Pandemic using the Utilization of Big Data

The Pandemic had a big impact on education in Indonesia and also in the world. In early 2020, during this pandemic, face-to-face meetings have turned into virtual or online meetings for both the learning process and seminars or workshops. The rapid development of technology supports this change in the world of education, this can be seen from the number of online seminars conducted to improve the competence of lecturers or teachers. The development of this online seminar allows the circulation of information that is increasingly large, fast, and almost unlimited by time and space. This causes a large amount of information to be scattered in the virtual world in various fields. With this very fast information technology, trillions of bytes of data are created every day from various sources such as on social media, especially those related to applications that are often used in website-based seminar media. This is called unstructured big data. In this study, big data will be implemented to classify educators' engagement of online seminar participants during an early pandemic. The activity stages in big data management and data processing support are acquired, accessed, analytical, and applied. The method for this study is the Adaptive Neuro-Fuzzy Inference System (ANFIS) to classify the engagement of teachers and lecturers an online seminar. The results of the training error obtained from ANFIS are 0.273482 with the ANFIS structure 4-12-12-12-1 or 4 inputs, 12 hidden layers, and 1 output.

MO475PREDICTION OF CARDIORENAL SYNDROME BY ARTIFITIAL INTELLIGENCE

Background and Aims Potassium excretion is a secretory phenomenon and levels are often abnormal in patients with heart failure. An abnormal sodium serum level is the most common electrolyte disorder and independent predictor of readmission for heart failure and post discharge death. Since different factors could affect balance of Potassium in Cardiorenal syndrome, in this study soft computing was used to predict most important factors for the detection of the severity of systolic heart failure by ejection fraction (EF), and a subclinical phase of the cardiorenal disease by EPI creatinine-cystatin C formula (Chronic Kidney Disease Epidemiology Collaboration). Method The balance of potassium in Cardiorenal syndrome is analyzed by soft computing approach namely adaptive neuro fuzzy inference system or ANFIS. Results The clinical group consisted of 79 patients, 40 of whom were men (50.63%) and 39 of whom were women (49.37%), in the average age of 70.72 ± 9.26 years. After comparing serum electrolytes (Na+, K+) did not differ significantly in the clinical group from those of the control group. The tested biomarkers showed significantly higher values in the clinical group than in the control group: BNP (p&lt;0.001), cystatin C (p&lt;0.001). Figures 1 and 2 shows flowcharts of the used inputs and outputs and how they are implemented in the ANFIS networks. There are two ANFIS networks since there are two outputs. ANFIS networks shold determine which input has the strongest influence on the given outputs nased on root mean squre errors or prediciton accuracy.Based on the training error (trn) one can determine the inputs influence on the given output. Checking error (chk) is used to track the results validity. In other words the checking errors could track training error. It was found that BNP (pg/mL) has the most influence on the - EPI creatinine-cystatin C formula. Serum sodium (Na) has the most influence on the ejection fraction (EF). Conclusion Serum sodium-potassium disturbances are associated with advanced heart failure and reduced prognosis. ANFIS is suitable for nonlinear systems with highly redundant data. Although there are encouraging advances around this unsolved clinical problem, further investigation should consider the progressive inclusion of patients with advanced renal impairment to allow a better understanding of cardiorenal syndrome. The result of our research shows that if the values of BNP and Na significantly deviate from normal values, it is expected that EPI creatinine-cystatin C formula and EF indicate impaired organ function and that such patients are candidates for hospital treatment.

Understanding double descent through the lens of principal component regression

Several recent papers have studied the double descent phenomenon: a classic U-shaped empirical risk curve when the number of parameters is smaller or equal to the number of data points, followed by a decrease in empirical risk (referred to as “second descent”) as the number of features is increased past the interpolation threshold (the minimum number of parameters needed to have 0 training error). In a similar vein as several recent papers on double descent, we concentrate here on the special case of over-parameterized linear regression, one of the simplest model classes that exhibit double descent, with the aim of better understanding the nature of the solution in the second descent and how it relates to solutions in the first descent. In this paper, we show that the final second-descent model (obtained using all features) is equivalent to the model estimated using principal component (PC) regression when all PCs of training data are included. It follows that many properties of double descent can be understood through the relatively simple and well-characterized lens of PC regression. In particular, we will identify a set of conditions that will guarantee final second-descent performance to be better than the best first-descent performance: it is the scenario in which PC regression using all features does not suffer from over-fitting and can be guaranteed to outperform any other first-descent model (any linear regression model using no more features than training data points). We will also discuss how this work relates to transfer learning, semi-supervised learning, few-shot learning, as well as theoretical concepts in neuroscience.