HEFT-Net: A Hybrid Emotion-Feature Transformer Network for Emotion Detection in Farmer Speech

Abstract

The presented model HEFT-Net: A Hybrid Emotion-Feature Transformer Network is employed for Emotion Detection in Farmer Speech. HEFT‑Net. The model integrates two complementary components namely a dedicated emotion‑feature extractor that learns rich prosodic and spectral representations from raw or log‑Mel inputs and a Transformer‑based temporal encoder that models long‑range context and cross‑frame interactions. By using pretrained transformer and a shallow classifier, HEFT-Net achieves a balance between expressive power and computational efficiency, making it suitable for real-world local language farmer speech data. The Wav2Vec2 is adopted as a self-supervised speech learner. The model operates directly on raw audio waveforms and generates frame-level embeddings through multiple stacked transformer layers. The Wav2Vec2 is able to encode phonetic information, temporal dependencies and contextual patterns, making it suitable for emotion recognition in farmer speech. The classification accuracy of 93.44% is achieved.