IMPROVED LONG-RANGE DEPENDENCY CAPTURE IN SPEECH SENTIMENT RECOGNITION: A COMPARATIVE STUDYOF MFCC-BILSTM AND MFCC-LSTM

The accurate analysis of sentiments conveyed through spoken language is of paramount importance in the domains of human-computer interaction.Speech sentiment recognition requires capturing long-range temporal dependencies within the speech signal.This study investigates the effectiveness of Bidirectional Long Short-Term Memory Networks (BiLSTMs) in capturing these dependencies compared to traditional unidirectional LSTMs. A hybrid approach combining Mel-Frequency Cepstral Coefficients (MFCCs) with BiLSTMs for speech sentiment analysis is proposed in this research work. The performance of the MFCC-BiLSTM model is evaluated and compared with that of an MFCC-LSTM model on TESS dataset. Experimental results demonstrate that the MFCC-BiLSTM model consistently outperforms the MFCC-LSTM model across various evaluation metrics, indicating improved accuracy of 98%in capturing long-range dependencies.The findings highlight the significance of bidirectional processing for enhancing speech sentiment recognition and provide valuable insights for developing more robust and accurate emotion recognition systems.