<p>Speech Emotion Recognition (SER) aims to detect and interpret human emotions from speech, that has vast applications across different fields. This research aims at developing SER system for Punjabi language. The SER process consists of mainly five phases, including the dataset creation, signal pre-processing, feature extraction, feature selection and classification. The scope of this paper is to explain all these phases for Punjabi SER system. The authors have designed and recorded two types of Punjabi Emotional Speech databases with professional, non-professional speakers, Punjabi language sentences, six emotions (happy, sad, neutral, surprise, fear and anger). A set of 16 features is extracted, including Mel spectrogram, Mel Frequency Cepstral Coefficients (MFCC), contrast, pitch, chroma, zero crossing rate (ZCR), Linear Prediction Cepstral Coefficients (LPCC), tonnetz, formant, jitter, shimmer, entropy, duration, harmonic, Perceptual Linear Predictive (PLP), and energy, with a total of 523 feature attributes. Feature selection is performed using Least Absolute Shrinkage and Selection Operator (LASSO) algorithm. Then a hybrid 1-D Convolutional Neural Network (CNN) with Long Short-Term Memory (LSTM) is used for classification. A number of experiments are carried out, firstly to select the best set of features for the feature extraction phase, then experimenting with four feature selection methods (LASSO, ANOVA, Recursive Feature Elimination (RFE), t-statistics) and selecting the best one. Then seven classifiers, namely Support Vector Machine (SVM), Decision Tree, Random Forest, Multi-Layer Perceptron (MLP), 1-D CNN, 2-D CNN and hybrid 1-D CNN with LSTM are experimented. The hybrid model outperformed other classifiers with an average accuracy of 81.6%. It has shown good performance metrics with precision, recall, f1-score, mean absolute error, mean square error and root mean square error as 83.5, 81.6, 81.7, 0.43, 1.25 and 1.11 respectively.</p>