<p>Robust feature extraction techniques are essential for consistently capturing emotions across diverse languages and datasets. This study performs a systematic comparative evaluation of deep learning feature embeddings from multiple self-supervised learning (SSL) models (Wav2vec 2.0 Large, Hidden-Unit BERT (HuBERT) Large, HuBERT XLarge, and Data2vec Large), handcrafted acoustic features such as Mel-Frequency Cepstral Coefficients, Chromagram, and Mel Spectrograms, and their fusion, to assess their effectiveness in speech emotion recognition on two language-specific datasets. The methodology is experimented on the RAVDESS (English) and EMODB (German) datasets using two machine learning classifiers: Support Vector Machine (SVM) and Logistic Regression (LogR). Our findings indicate that combining SSL embeddings with handcrafted acoustic features significantly improves emotion recognition performance compared to using each feature set independently in all models and datasets, regardless of the classifier used. LogR surpasses SVM classifiers in both datasets, achieving the highest average recall of 87.01% and 90.67% on RAVDESS using HuBERT XLarge features fused with handcrafted acoustic features, and on EMODB using HuBERT Large features fused with handcrafted acoustic features, respectively, in 5-fold cross-validation. The high recall rates demonstrate the effectiveness of our approach and underscore SSL models’ ability to capture both linguistic and prosodic features for speech emotion recognition. Despite the additional computational overhead associated with SSL in extracting features, the substantial improvement in classification performance more than compensates for these costs.</p>