|
|
Emotion Recognition from Analysis of a Person’s Speech
Knutelský, Martin ; Shakil, Sadia (oponent) ; Malik, Aamir Saeed (vedoucí práce)
This thesis deals with the analysis of emotion recognition from human speech. It aims to design and implement a system that can automatically infer emotional states from speech recordings. The solution is based on the Audio Spectrogram Transformer (AST), a derivative of the Vision Transformer neural network, which accepts mel spectrogram as input. The implementation comprehends the pipeline with two stages. In the first stage, a mel spectrogram is obtained from the input speech recording and in the second stage, the pretrained AST model computes output in the form of probabilities of considered emotional classes. The AST implementation was trained and evaluated on three datasets: RAVDESS, Emo-DB and EMOVO. The obtained results in the form of unweighted accuracy are 84.5 % for RAVDESS, 91.6 % for Emo-DB and 73.8 % for EMOVO. During training, the consumed energy of the graphical processing unit was recorded for the calculation of the carbon footprint in terms of emitted CO2. The main contribution of this work is the utilization of neural network based on Transformer architecture, originally used for vision tasks, to classify emotions from speech. Another contribution is carbon footprint tracking of neural network training. The carbon footprint, expressed in emitted CO2 mass is 1058.37 grams.
|
host ::
RSS.