Fast Discrete Orthogonal Stockwell Transform-driven Local-global Temporal Modelling with Attention-enhanced Dual-stage BiLSTM for Accurate Age and Gender Recognition from Speech
Year: 2026
Type: article
Abstract: Predicting age and gender from speech remains challenging because human speech contains attributes that are deeply intertwined, sensitive to noise, and marked by subtle, variable acoustic cues.This paper introduces a hierarchical feature-learning framework that combines fast discrete orthogonal Stockwell transform-based (FDOST) spectral representations with dual-stage temporal modeling to preserve and capture both local and global speech dynamics.We first extract FDOST features to obtain high-resolution time-frequency representations.These features are then passed through a local bidirectional long short-term memory (BiLSTM) module that captures fine-grained articulatory and harmonic detail, followed by a global BiLSTM designed to model long-range prosodic structure across entire utterances.The representations are further refined by a multi-head self-attention mechanism and a high-capacity, position-wise feed-forward network that flexibly highlights salient patterns.Extensive experiments show that the proposed architecture achieves reliable multi-class age classification and binary gender recognition.For age classification across eight groups, the model attains a mean absolute classification error (MACE) of 0.1396 and an F1score of 94.24%.For gender recognition, it reaches an accuracy of 97.16% and a macro-averaged F1-score of 97.05%.These results are supported by near-perfect ROC curves, well-balanced confusion matrices, and thorough ablation analyses, confirming the strength of the FDOST-driven hierarchical model for speech-based biometric prediction.
Institution: University of Tabriz
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7234963090
