<p>This research investigates the generalization of speech emotion recognition models across multiple datasets and languages, with a focus on the role of feature scaling in enhancing the performance of within-corpus evaluations. We evaluated one dimensional convolutional neural network (1DCNN), XGBoost, and a stacked ensemble across six datasets without augmentation, using four feature scaling algorithms (StandardScaler, RobustScaler, MinMaxScaler, and Normalizer), which yielded significant findings. The results indicate that the RAVDESS (English) benchmark dataset maintained consistent accuracy across different scaling algorithms, achieving state-of-the-art performance. In contrast, the results of the EMODB (German) varied depending on the scaling algorithm used; the XGBoost algorithm performed better when combined with Normalizer-based feature scaling. Generally, 1DCNN outperformed XGBoost, and the ensemble model consistently outperformed individual baselines, achieving average accuracy gains of up to 3.56% in EMODB and 2.40% in RAVDESS. Improvements were also observed in Bangla datasets (1.53% in BanglaSER, 3.12% in KBES (KUET Bangla Emotional Speech), and 2.25% in SUBESCO), as well as in the custom Malayalam dataset (3.60%). Complementary statistical analyses of skewness, kurtosis, and outlier counts revealed substantial distributional differences across corpora, shaping how algorithms responded to scaling. Cross-corpus and cross-lingual analyses highlighted the limitations of conventional models. Fine-tuned transfer learning substantially improved robustness, achieving gains of 28–35% in unweighted average recall across datasets. These findings highlight that while scaling effects in within-corpus tasks depend on dataset distributions and algorithm–data interactions, effective cross-corpus transfer underscores the importance of selecting suitable source datasets and models to improve speech emotion recognition performance across diverse linguistic contexts.</p>