Speech Emotion Recognition (SER) is an important component of effective computing, and can be used in Human-computer interaction, healthcare, and intelligent systems applications. This paper gives a comparative analysis of Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) in recognizing emotions of Kannada speech. The dataset consists of 2,500 utterances in five categories of feelings, which were recorded under controlled conditions by 20 speakers, and it was additionally enriched to 11,250 utterances through the use of pitch shifting, time-stretching, noises, and SpecAugment. Both models used Mel-frequency cepstral coefficients (MFCCs) as their input features. The RNN was experimentally found to be more accurate with a 98.84% accuracy rate over the CNN, which had a 93.75% accuracy rate. These findings show the utility of the temporal modelling in low-resource language SER and demonstrate the efficiency of the proposed data augmentation strategy. The further development will involve an investigation of the more sophisticated architectures and testing the performance of systems in real-life.