Speech Emotion Recognition (SER) is a critical component of affective computing with broad applications in human–computer interaction, mental health monitoring, and intelligent multimedia systems. However, SER remains challenging due to the emotional ambiguity, lack of labeled data, class imbalance, and speaker variability. This study presents an effective SER framework that integrates contrastive representation learning, optimized spectrogram-based data augmentation, and selective synthetic data generation by using TimeGAN to enhance emotion classification performance. Contrastive learning enables the model to better discriminate acoustically similar emotions while Optuna automatically tunes augmentation strategies such as noise injection, time shifting, and time-frequency masking. Unlike existing approaches that apply synthetic generation uniformly across all classes, the proposed method targets only confusing or under-represented emotion classes to preserve the inter-class separability. A CNN-BiLSTM architecture is used to extract spectral and temporal information of the speech. The framework is evaluated with benchmark SER datasets—EMO-DB and RAVDESS—under speaker independent protocols. Experimental results demonstrate improved accuracy, robustness, and generalization under limited and imbalanced data conditions, supported by confusion matrices, UMAP, and t-SNE visualizations.

Deepfake audio created with sophisticated speech synthesis and voice cloning technologies is a threat to the credibility of digital communication. Its realism has raised serious concerns in different applications such as digital forensics, cybersecurity, media authentication and voice-based security systems. However, deepfake audio detection still remains difficult. Synthetic speech tends to have subtle artifacts that can mimic the natural vocal pattern very closely. Variations in speakers, recording conditions and background noise make the task more complex. In addition, dataset imbalance and low diversity in training samples could lead to low robustness in the model. To overcome these limitations, the present study aims to propose a framework of transfer learning-based methods based on a combination of fine-tuned pre-trained models, as well as systematic data augmentation. Augmentation methods are introduced to increase the variability and mimic real acoustic conditions. This approach supports the learning of more stable and generalizable representations for both genuine and manipulated speech. The framework employs three DL models: ResNet50 to capture global spectro-temporal structures, VGGish to extract mid-level semantic audio embeddings and YAMNet to identify fine-grained temporal irregularities associated with synthetic speech artifacts. Features from these models are fused through concatenation to construct a unified hybrid feature space. A feature selection stage then reduces redundancy before classification using a lightweight model. Experimental results demonstrate the superiority of the proposed hybrid approach and achieved an accuracy of 99.7%. This performance significantly outperformed individual baseline models and achieved strong generalization across diverse acoustic conditions.

Human activity recognition (HAR) is a method to predict human activities from sensor signals using machine learning (ML) techniques. HAR systems have several applications in various domains, including medicine, surveillance, behavioral monitoring, and posture analysis. Extraction of suitable information from sensor data is an important part of the HAR process to recognize activities accurately. Several research studies on HAR have utilized Mel frequency cepstral coefficients (MFCCs) because of their effectiveness in capturing the periodic pattern of sensor signals. However, existing MFCC-based approaches often fail to capture sufficient temporal variability, which limits their ability to distinguish between complex or imbalanced activity classes robustly. To address this gap, this study proposes a feature fusion strategy that merges time-based and MFCC features (MFCCT) to enhance activity representation. The merged features were fed to a convolutional neural network (CNN) integrated with long short-term memory (LSTM)—DeepConvLSTM to construct the HAR model. The MFCCT features with DeepConvLSTM achieved better performance as compared to MFCCs and time-based features on PAMAP2, UCI-HAR, and WISDM by obtaining an accuracy of 97%, 98%, and 97%, respectively. In addition, DeepConvLSTM outperformed the deep learning (DL) algorithms that have recently been employed in HAR. These results confirm that the proposed hybrid features are not only practical but also generalizable, making them applicable across diverse HAR datasets for accurate activity classification.