Attention-based Bi-LSTM model with hybrid wavelet–MFCC features for speech emotion recognition

Life Cycle Reliability and Safety Engineering · Springer Nature · 2026-08-27 · DOI: 10.1007/s10772-026-10287-4

Аннотация

It is still difficult to achieve the high accuracy of emotions recognition from the voice due to the variability of the speakers, conditions of recordings and non-stationarity of the speech signals. In this study, we have proposed the attention-based bidirectional Long Short-Term Memory (Bi-LSTM) network with hybrid Wavelet and Mel Frequency Cepstral Coefficient (MFCC) for robust speech emotion recognition. The proposed hybrid features extraction technique takes advantage of the time-frequency representations by combining the multi-resolution characteristics of the discrete wavelet transform with the perceptual spectral information provided by the MFCCs. The hybrid features are fed to the attention-based Bi-LSTM neural network, which can highlight emotionally important segments of speech, as well as model temporal dependencies between them. We evaluate our proposed method on the publicly available emotional speech datasets – EMO-DB, RAVDESS and CREMA-D, using the independent five-fold cross-validation methodology. The obtained experimental results prove the superiority of the proposed hybrid representation over the single Wavelet and MFCC features, achieving the overall recognition accuracy of 99% and average accuracy of 94%.

Читать у издателя →

Похожие статьи