Article: Deep generative simulation modelling for time-variant processes of vocal emotion features Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.84 pp.49 - 77 Abstract: Vocal emotional expression is a highly complex, time-varying process. Current generative models often struggle with long-term dependencies and fine-grained dynamic transitions, causing over-smoothed acoustic features. To address this limitation, a temporal-conditioned diffusion model is proposed for the high-fidelity simulation of time-varying vocal emotional features. The model utilises a stochastic differential equation - based diffusion process operating on a multidimensional acoustic feature space. During the reverse generative phase, a hybrid LSTM-Attention network is employed to capture both macroscopic and microscopic emotional transitions. Comprehensive evaluations demonstrate that the proposed model significantly outperforms baselines. specifically, our model achieved a lower Mel-cepstral distortion (3.95 dB), reduced F0 RMSE (12.8 Hz), and a higher energy Pearson correlation (0.92). Furthermore, subjective mean opinion scores for acoustic naturalness (4.35) and emotional expressiveness (4.42) surpassed the best baseline. This study provides a robust generative paradigm for advanced affective computing. Inderscience Publishers - linking academia, business and industry through research

Title: Deep generative simulation modelling for time-variant processes of vocal emotion features

Authors: Yu Zhang

Addresses: Vocational and Technical College, Huanghe Science and Technology University, Zhengzhou, 450000, China

Abstract: Vocal emotional expression is a highly complex, time-varying process. Current generative models often struggle with long-term dependencies and fine-grained dynamic transitions, causing over-smoothed acoustic features. To address this limitation, a temporal-conditioned diffusion model is proposed for the high-fidelity simulation of time-varying vocal emotional features. The model utilises a stochastic differential equation - based diffusion process operating on a multidimensional acoustic feature space. During the reverse generative phase, a hybrid LSTM-Attention network is employed to capture both macroscopic and microscopic emotional transitions. Comprehensive evaluations demonstrate that the proposed model significantly outperforms baselines. specifically, our model achieved a lower Mel-cepstral distortion (3.95 dB), reduced F0 RMSE (12.8 Hz), and a higher energy Pearson correlation (0.92). Furthermore, subjective mean opinion scores for acoustic naturalness (4.35) and emotional expressiveness (4.42) surpassed the best baseline. This study provides a robust generative paradigm for advanced affective computing.

Keywords: vocal emotional characteristics; time-varying process; deep generative model; diffusion model; simulation modelling; emotional computing.

DOI: 10.1504/IJICT.2026.155480

International Journal of Information and Communication Technology, 2026 Vol.27 No.84, pp.49 - 77

Received: 27 Mar 2026
Accepted: 24 May 2026
Published online: 03 Aug 2026 *

Title: Deep generative simulation modelling for time-variant processes of vocal emotion features

Keep up-to-date