Speech recognition (SR) is an crucial job in human–computer interface (HCI) with applications in healthcare, virtual assistants, and affective computing. Conventional SR models tend to fail in capturing emotional subtleties as a result of having limited training data and domain fluctuation. In this paper, an Updated Generative Adversarial Network (Updated GAN) is presented for emotional speech generation and recognition. Our model uses a GAN-based method to produce high-quality emotional speech samples to augment the training dataset for better generalization. This work use the RAVDESS dataset to train a speaker recognition model based on VGG16, using spectrogram representations for strong feature extraction. The Updated GAN produces emotional speech in eight categories: neutral, cool, joyful, unhappy, mad, awful, hatred, and amazed. The synthesized AUDIO is subsequently used to enrich the training data and enhance the speech emotion recognition (SER) framework’s performance. Experimental evidence shows that the proposed method markedly improves emotion classification accuracy (97.84%), which is measured in terms of confusion matrix analysis. The results further show that embedding GAN-generated speech into the conventional SER model results in enhanced and more consistent speaker recognition accuracy. This research helps to push the boundaries of emotion recognition technology by alleviating data sparsity and enhancing classification resilience.