Compact deep neural networks for real-time speech enhancement on resource-limited devices

Speech is easily interfered by external environment in reality, which results in the loss of important features. Deep learning has become a popular speech enhancement method because of its superior potential in solving nonlinear mapping problems for complex features. However, the deficiency of traditional deep learning methods is the weak learning capability of important information from previous time steps and long-term event dependencies between the time-series data. To overcome this problem, we propose a novel speech enhancement method based on the fused features of deep neural networks (DNNs) and gated recurrent unit (GRU). The proposed method uses GRU to reduce the number of parameters of DNNs and acquire the context information of the speech, which improves the enhanced speech quality and intelligibility. Firstly, DNN with multiple hidden layers is used to learn the mapping relationship between the logarithmic power spectrum (LPS) features of noisy speech and clean speech. Secondly, the LPS feature of the deep neural network is fused with the noisy speech as the input of GRU network to compensate the missing context information. Finally, GRU network is performed to learn the mapping relationship between LPS features and log power spectrum features of clean speech spectrum. The proposed model is experimentally compared with traditional speech enhancement models, including DNN, CNN, LSTM and GRU. Experimental results demonstrate that the PESQ, SSNR and STOI of the proposed algorithm are improved by 30.72%, 39.84% and 5.53%, respectively, compared with the noise signal under the condition of matched noise. Under the condition of unmatched noise, the PESQ and STOI of the algorithm are improved by 23.8% and 37.36%, respectively. The advantage of the proposed method is that it uses the key information of features to suppress noise in both matched and unmatched noise cases and the proposed method outperforms other common methods in speech enhancement.

In this paper, we considered the problem of the speech enhancement similar to the real-world environments where several complex noise sources simultaneously degrade the quality and intelligibility of a target speech. The existing literature on the speech enhancement principally focuses on the presence of one noise source in mixture signals. However, in real-world situations, we generally face and attempt to improve the quality and intelligibility of speech where various complex stationary and nonstationary noise sources are simultaneously mixed with the target speech. Here, we have used deep learning for speech enhancement in complex-noisy environments and used ideal binary mask (IBM) as a binary classification function by using deep neural networks (DNNs). IBM is used as a target function during training and the trained DNNs are used to estimate IBM during enhancement stage. The estimated target function is then applied to the complex-noisy mixtures to obtain the target speech. The mean square error (MSE) is used as an objective cost function at various epochs. The experimental results at different input signal-to-noise ratio (SNR) showed that DNN-based complex-noisy speech enhancement outperformed the competing methods in terms of speech quality by using perceptual evaluation of speech quality (PESQ), segmental signal-to-noise ratio (SNRSeg), log-likelihood ratio (LLR), weighted spectral slope (WSS). Moreover, short-time objective intelligibility (STOI) reinforced the better speech intelligibility.

In this study, different kinds of hybrid Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithms with forecasting models including Random Forest (RF), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM) neural networks, are developed to estimate the mean daily wind speed at the height of 2 m in Ağrı city (WS st12 ), Turkey. In these hybrid models, different layer networks of single and integrated LSTM and GRU models include general single LSTM, general single GRU, simple coupled LSTM-GRU, and novel coupled LSTM with GRU through Addition layer (i.e., LSTM + GRU model) structures are applied. The most effective parameters on the WS st12 , from a list of on-site potential meteorological parameters and wind speed values in its adjacent cities of Ağrı province from Jan 2015–Dec 2019 through the Pearson correlation coefficient method, are determined. In the hybrid CEEMDAN and DNNs-based models, State activation functions (SAF), numbers of hidden neurons (NHN), dropout rates (P-rate), and network structural architect (NSA) as the meta-parameters are tuned for lessening the impact of overfitting/underfitting dilemmas and improving modeling performance. According to the comparison plots, performance evaluation measures, and total learnable parameter (TLP), the novel developed hybrid CEEMDAN-RF-(LSTM + GRU) model is confirmed as the best approach with an R 2 of 0.86 while, in the optimal scenario using the RF model, R 2 was 0.47. Graphical Abstract Based on the graphical snapshot, this study focuses on estimating daily mean wind speed at a 2-meter height in Ağrı, Turkey, using hybrid data-driven models. The research integrates the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) algorithm with advanced forecasting techniques, including Random Forest (RF), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) neural networks. The modeling framework explores various configurations, such as standalone LSTM and GRU, coupled LSTM-GRU structures, and a novel LSTM + GRU model using an Addition layer to enhance predictive accuracy.

This chapter considers the challenges when using Recurrent Neural Networks (RNNs) for Big multimodal time series, forecasting where both the spatial and temporal information has to be used for accurate forecasting. Although RNN and its variations such as Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Simple Recurrent Unit (SRU) progressively improve the quality of training outcomes by implementing the network structures and optimisation techniques, one major limitation in such models is that most of them vectorise the input data and thus destroy a continuous spatial representation. We propose an approach termed the Tensorial Recurrent Neural Network (TRNNs) which addresses the problem on how to analyse the multimodal data as well as considers the relationship along with the time series, and shows that the TRNN outperforms other RNN models for image captioning applications.

In smart mariculture, traditional methods are not only difficult to adapt to the complex, dynamic and changeable environment in open waters, but also have many problems, such as poor accuracy, high time complexity and poor long-term prediction. To solve these deficiencies, a new water quality prediction method based on TCN (temporal convolutional network) is proposed to predict dissolved oxygen, water temperature, and pH. The TCN prediction network can extract time series features and in-depth data features by introducing dilated causal convolution, and has a good effect of long-term prediction. At the same time, it is predicted that the network can process time series data in parallel, which greatly improves the time throughput of the model. Firstly, we arrange the 23,000 sets of water quality data collected in the cages according to time. Secondly, we use the Pearson correlation coefficient method to analyze the correlation information between water quality parameters. Finally, a long-term prediction model of water quality parameters based on a time domain convolutional network is constructed by using prior information and pre-processed water quality data. Experimental results show that long-term prediction method based on TCN has higher accuracy and less time complexity, compared with RNN (recurrent neural network), SRU (simple recurrent unit), BI-SRU (bi-directional simple recurrent unit), GRU (gated recurrent unit) and LSTM (long short-term memory). The prediction accuracy can reach up to 91.91%. The time costs of training model and prediction are reduced by an average of 64.92% and 7.24%, respectively.

Deep learning (DL) has emerged as a powerful tool for intelligent cyberattack detection, especially Distributed Denial-of-Service (DDoS) in Software-Defined Networking (SDN), where rapid and accurate traffic classification is essential for ensuring security. This paper presents a comprehensive evaluation of six deep learning models (Multilayer Perceptron (MLP), one-dimensional Convolutional Neural Network (1D-CNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Recurrent Neural Network (RNN), and a proposed hybrid CNN-GRU model) for binary classification of network traffic into benign or attack classes. The experiments were conducted on an SDN traffic dataset initially exhibiting class imbalance. To address this, Synthetic Minority Over-sampling Technique (SMOTE) was applied, resulting in a balanced dataset of 24,500 samples (12,250 benign and 12,250 attacks). A robust preprocessing pipeline followed, including missing value verification (no missing values were found), feature normalization using StandardScaler to standardize numerical values, reshaping the data into 3D format to fit temporal models like CNN and GRU, and stratified train-test split (80% training, 20% testing) to maintain class distribution. The CNN-GRU model integrates a 1D convolutional layer for spatial pattern extraction and a GRU layer for temporal sequence learning, followed by dense layers with dropout regularization. The model was trained using the Adam optimizer with early stopping to prevent overfitting. Among all models, the CNN-GRU hybrid achieved perfect test performance, with 100% accuracy, 1.0000 precision, recall, and F1-score, and an ROC AUC of 1.0000. It also demonstrated exceptional generalization, achieving a mean cross-validation (CV) accuracy of 99.70% ± 0.09% and a mean AUC of 1.0000 ± 0.0000 across 5-fold stratified cross-validation. While individual models such as GRU, 1D-CNN, and LSTM also showed strong performance, the CNN-GRU hybrid consistently outperformed them in both accuracy and stability. These results validate the effectiveness of combining convolutional and recurrent architectures, augmented with data balancing via SMOTE, for highly accurate SDN-based intrusion detection.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-13754-1.

Speech enhancement (SE) aims to improve speech quality and intelligibility by removing acoustic corruption. While various SE models using audio-only (AO) based on deep learning have been developed to achieve successful enhancement for non-speech background noise, audio-visual SE (AVSE) models have been studied to effectively remove competing speech. In this paper, we propose an AVSE model that estimates spectral masks for real and imaginary components to consider phase enhancement. It is based on the U-net structure that allows the decoder to perform information restoration by leveraging intermediate information in the encoding process and avoids the gradient vanishing problem by providing paths direct to the encoder’s layers. In the proposed model, we present early fusion to process audio and video with a single encoder that effectively generates features for the fused information easy to decode for SE with reduced parameters of the encoder and decoder. Moreover, we extend the U-net using the proposed Recurrent-Neural-Network (RNN) attention (RA) blocks and the Res paths (RPs) in the skip connections and the encoder. While the RPs are introduced to resolve the semantic gap between the low-level and high-level features, the RA blocks are developed to find efficient representations with inherent frequency-specific characteristics for speech as a type of time-series data. Experimental results on the LRS2-BBC dataset demonstrated that AV models successfully removed competing speech and our proposed model efficiently estimated complex spectral masks for SE. When compared with the conventional U-net model with a comparable number of parameters, our proposed model achieved relative improvements of about 7.23%, 5.21%, and 22.9% for the signal-to-distortion ratio, perceptual evaluation of speech quality, and FLOPS, respectively.

Prediction of dissolved oxygen in a fishery pond based on gated recurrent unit (GRU)

Recently different Deep Learning (DL) models have been emerging to predict the disease amelioration. Generally deep learning is the state of art for learning the multiple representation of neuron. Discriminant DL model includes different architecture and initial architecture is Recurrent Neural Network (RNN) which learn the data labeled in the form of sequence and it has long term dependency and vanishing gradient problem. Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) is an improved version of RNN which deal with problems in RNN. Deep care LSTM is one of the model in LSTM developed by [Trang Pham et.al] for predicting diabetes disease and compared the results with markovian and support vector machine model. In deep LSTM model some problems are reported like short term trajectories, less accuracy. To overcome the problem Deep Care GRU has been proposed to identify the diabetes disease amelioration.

Energy efficiency, scalability, and reliability are increasingly important for sustainable data centers. In this paper, we focus on forecasting real-world resource usage using neural network time series models, specifically utilizing convolutional recurrent long short-term Memory (LSTM) and gated recurrent unit (GRU) architectures. In our analysis, we compare LSTM and GRU in terms of forecasting accuracy and computational complexity during model training. We demonstrate that recurrent neural networks are more accurate and robust compared to the traditional autoregressive integrated moving average (ARIMA) time series model in this complex forecasting problem. GRU achieved a 9% reduction and LSTM a 5% reduction in forecasting mean squared error (MSE) compared to ARIMA. Furthermore, the GRU architecture with a 1D convolution layer outperforms LSTM architecture in both forecast accuracy and training time. The proposed model can be effectively applied to load forecasting as part of a data center computing cluster. In this application, the proposed GRU architecture has 25% fewer trainable parameters in the recurrent layer than the commonly used LSTM.

This study investigates the application of advanced deep learning models to forecast fossil energy prices, a critical factor influencing global economic stability. Unlike previous research, this study conducts a comparative analysis of Gated Recurrent Unit (GRU), Recurrent Neural Network (RNN), Bidirectional Long Short-Term Memory (Bi-LSTM), Long Short-Term Memory (LSTM), Convolutional Neural Network (CNN), and Deep Neural Network (DNN) models. The evaluation metrics employed include Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). The results reveal that recurrent architectures, particularly GRU, LSTM, and Bi-LSTM, consistently outperform feedforward and convolutional models, demonstrating superior ability to capture temporal dependencies and nonlinear dynamics in energy markets. In contrast, the RNN and DNN show relatively weaker generalization capabilities. Additionally, visualizations of actual versus predicted prices for each model further emphasize superior forecasting accuracy of recurrent models. The results highlight the potential of deep learning in enhancing investment and policy decisions. Additionally, the results provide significant implications for policymakers and investors by emphasizing the value of accurate energy price forecasting in mitigating market volatility, improving portfolio management, and supporting evidence-based energy policies.

Hearing impaired listeners using cochlear implants (CIs) suffer from a decrease in speech intelligibility (SI) in adverse listening conditions. Time-frequency masks are often applied to perform noise suppression in an attempt to increase SI. Two important masks are the so-called ideal binary mask (IBM) with its binary weights and the ideal Wiener filter (IWF) with its continuous weights. It is unclear which of the masks has the highest potential for SI and speech quality enhancement in CI users. In this study, both approaches for SI and quality enhancement were compared. The investigations were conducted in normal-hearing (NH) subjects listening to noise vocoder CI simulations and in CI users. The potential for SI improvement was assessed in a sentence recognition task with ideal mask estimates in multitalker babble and with an interfering talker. The robustness of the approaches was evaluated with simulated estimation errors. CI users assessed the speech quality in a preference rating. The IWF outperformed the IBM in NH listeners. In contrast, no significant difference was obtained in CI users. Estimation errors degraded SI in CI users for both approaches. In terms of quality, the IWF outperformed, slightly, the IBM processed signals. The outcomes of this study suggest that the mask pattern is not that crucial for CIs. Results of speech enhancement algorithms obtained with NH subjects listening to vocoded or normally processed stimuli do not translate to CI users. This outcome means that the effect of new strategies has to be quantified with the user group considered.

In many deep learning based speech enhancement frameworks, the deep neural network (DNN) is usually trained under the mean square error (MSE) criterion to estimate the potential clean amplitudes. Yet, the MSE criterion is not directly connected with speech perception, which potentiates alternative training targets for further improvements in speech quality and intelligibility. This paper introduces maximizing the correlation between the enhanced speech and the reference, so as to further improve the speech quality and intelligibility of the enhanced speech synchronously. Based on the short-time Fourier transform amplitude spectra, the correlations are firstly calculated between the envelops along the time and frequency respectively. Then, the time and frequency envelop correlations are incorporated into the time-frequency correlation (TFC). Finally, the loss function is defined as 1 minus the TFC. In experiment, several loss functions containing the time and frequency envelop correlations are investigated. The system trained with the TFC loss outperforms the baseline trained with the MSE loss with promising improvements in both speech quality and intelligibility. Phoneme recognition results also confirm the effectiveness of maximizing the correlation for speech enhancement.

In view of the current problems of complex models and insufficient data processing in ultra-short-term prediction of photovoltaic power generation, this paper proposes a photovoltaic power ultra-short-term prediction model named HPO-KNN-SRU, based on a Simple Recurrent Unit (SRU), K-Nearest Neighbors (KNN), and Hunter–Prey Optimization (HPO). Firstly, the sliding time window is determined by using the autocorrelation function (ACF), partial correlation function (PACF), and model training. The Pearson correlation coefficient method is used to filter the principal meteorological factors that affect photovoltaic power. Then, the K-Nearest Neighbors (KNN) algorithm is utilized for effective outlier detection and processing to ensure the quality of input data for the prediction model, and the Hunter–Prey Optimization (HPO) algorithm is applied to optimize the parameters of the KNN algorithm. Finally, the efficient Simple Recurrent Unit (SRU) model is used for training and prediction, with the Hunter–Prey Optimization (HPO) algorithm applied to optimize the parameters of the SRU model. Simulation experiments and extensive ablation studies using photovoltaic data from the Desert Knowledge Australia Solar Centre (DKASC) in Alice Springs, Australia, validate the effectiveness of the integrated model, the KNN outlier handling, and the HPO algorithm. Compared to the Support Vector Regression (SVR), Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), and Simple Recurrent Unit (SRU) models, this model exhibits an average reduction of 19.63% in Mean Square Error (RMSE), 27.54% in Mean Absolute Error (MAE), and an average increase of 1.96% in coefficient of determination (R2) values.

This paper proposes the application of artificial intelligence to forecast the generation capacity of wind power plants by processing data through noise reduction and filtering. It subsequently employs Long-Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks for training, testing, and evaluation. Processing the initial data will help minimize noise and reduce the data space. The study focuses on preprocessing methods and selecting the appropriate neural network between LSTM and GRU. The initial data processing will assess the similarity through the Spearman rank correlation coefficient. The data used in the paper is taken from local wind turbines. The processed data will be input into the neural network for evaluation based on Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), Percent Error (PE), and training time. The entire network simulation and evaluation process is performed using MATLAB software. The simulation results show the feasibility and suitability of the GRU network model combined with noise filtering methods, bringing high accuracy and less training time compared to the LSTM network. Specifically, the analysis contrasting the GRU network with the optimized dataset and the LSTM with the unprocessed dataset is more effective than a difference in RMSE of 15.592 and MAPE of 611.047%. Moreover, the time difference between the GRU and LSTM networks with the same dataset has a much earlier time difference from 6 to 28 seconds.