Emotion Recognition with Intensity Level from Bangla Speech using Feature Transformation and Cascaded Deep Learning Model
Masum Billah, Md. Likhon Sarker, M. A. H. Akhand, Md Abdus Samad Kamal
Abstract
Speech Emotion Recognition (SER) identifies and categorizes emotional states by analyzing speech signals. The intensity of specific emotional expressions (e.g., anger) conveys critical directives and plays a crucial role in social behavior. SER is intrinsically language-specific; this study investigated a novel cascaded deep learning (DL) model to Bangla SER with intensity level. The proposed method employs the Mel-Frequency Cepstral Coefficient, Short-Time Fourier Transform (STFT), and Chroma STFT signal transformation techniques; the respective trans-formed features are blended into a 3D form and used as the input of the DL model. The cascaded model performs the task in two stages: classify emotion in Stage 1 and then measure the intensity in Stage 2. DL architecture used in both stages is the same, which consists of a 3D Convolutional Neural Network (CNN), a Time Distribution Flatten (TDF) layer, a Long Short-term Memory (LSTM), and a Bidirectional LSTM (Bi-LSTM). CNN first extracts features from 3D formed input; the features are passed through the TDF layer, Bi-LSTM, and LSTM; finally, the model classifies emotion along with its intensity level. The proposed model has been evaluated rigorously using developed KBES and other datasets. The proposed model revealed as the best-suited SER method compared to existing prominent methods achieving accuracy of 88.30% and 71.67% for RAVDESS and KBES datasets, respectively.