Learnable Deep Wavelet Packet Transform for Speech Emotion Recognition in High-Risk Suicide Calls

In human-computer interaction and psychological evaluation, speech emotion recognition (SER) is crucial for interpreting emotional states from spoken language. Although there have been advancements, challenges such as system complexity, issues with feature distinctiveness, and noise interference continue to persist. This paper presents a novel end-to-end (E2E) deep learning multi-resolution framework for SER, which tackles these limitations by deriving significant representations directly from raw speech waveform signals. By leveraging the properties of wavelet packet transform (WPT), our approach introduces a learnable model for both wavelet bases and denoising through deep learning techniques. Unlike discrete wavelet transform (DWT), WPT offers a more detailed analysis by decomposing both approximation and detail coefficients, providing a finer resolution in the time-frequency domain. This capability enhances feature extraction by capturing more nuanced signal characteristics across different frequency bands. The framework incorporates a learnable activation function for asymmetric hard thresholding of wavelet packet coefficients. Our approach exploits the capabilities of wavelet packets for effective localization in both time and frequency domains. We then combine one-dimensional dilated convolutional neural networks (1D dilated CNN) with a spatial attention layer and bidirectional gated recurrent units (Bi-GRU) with a temporal attention layer to efficiently capture emotional features' nuanced spatial and temporal characteristics. By handling variable-length speech without segmentation and eliminating the need for pre or post-processing, the proposed model outperforms state-of-the-art methods on our NSPL-CRISE dataset and on the public IEMOCAP dataset. The source code of this paper is shared on this repository: https://github.com/alaaNfissi/WPT-Deep-Learning-SER-for-Suicide-Monitoring.