A combined CNN-LSTM Network for Audio Emotion Recognition using Speech and Song attributs

Audio emotion recognition is a very active topic. It has many applications in various domains like health care, information technology and defence. This paper presents a hybrid CNN-LSTM Deep Network for audio emotion classification using speech and song description. The model combines the temporal modelling ability of a long shirt term memory (LSTM) with the CNN ability to learn invariant features. We evaluate our experements on RAVDESS Database that contains audio song and audio speech formats. We show the affectiveness of both CNN and LSTM neural networks for feature extraction and memorizing a long sentence in order to provide the stability and a good performance of the model by using MFCC which reinforces the idea of using frames instead of waves to improve the training process. We compare the obtained results for both modalities for the same amount number of emotions which are “happy”, “sad”, “angree”, “neutral”, “calm” and “fear”.