Scholarlink

<p>Gender-aware emotion recognition from speech is critical in affective computing applications, including virtual assistants, telehealth, and intelligent surveillance, where interpretability and demographic fairness are paramount. However, real-world speech signals often suffer from compounded challenges such as noise, whispering, short utterance duration, multilingual variability, and age-related vocal changes all of which significantly degrade the accuracy and robustness of emotion recognition models, especially in preserving gender-specific prosodic cues like pitch and energy. To address these limitations, we propose HEDRNet: a Hybrid Diffusion-Transformer Network tailored for robust and equitable emotion recognition under naturalistic speech conditions. HEDRNet comprises four synergistic components engineered to resolve the multidimensional challenges of spontaneous speech. First, the DiffuLANS module leverages a diffusion-based denoising architecture with Latent Adaptive Noise Scheduling, which dynamically adapts the reverse diffusion trajectory based on input SNR and speech type. This ensures the preservation of emotionally salient prosody even in whispered, noisy, or aged speech. Second, the enhanced signal is processed by SwinSpeechX, a Speech Swin Transformer equipped with Prosodic-Guided Feature Enhancement and Temporal Dynamic Attention, enabling fine-grained modelling of emotion-specific temporal dynamics in short utterances. Third, Multilingual Feature Alignment is integrated via alignment loss to unify emotional feature representations across languages, ensuring robustness in code-switched and multilingual scenarios. Comprehensive evaluations on IEMOCAP, RAVDESS, CREMA-D, and EMO-DB datasets demonstrate that HEDRNet achieves a PSNR-equivalent accuracy of up to 94.54%, UA of 93.8%, F1-scores above 95%, and a PRI of 0.95, while maintaining 93.8% UA under 0&#xa0;dB SNR and 94.9% UA in cross-lingual tests. These results confirm HEDRNet’s effectiveness as a high-fidelity, demographically fair, and multilingual-ready solution for emotion recognition in complex speech environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

HEDRNet: A Hybrid Diffusion-Transformer Network Enhanced with Latent Adaptive Noise Scheduling and Prosodic Guided Attention for Gender Aware Emotion Recognition in Short Speech

  • 期刊论文

  • 发表日期：

    2026年3月19日

查看全文
  • Sanghamitra V. Arora

摘要

Gender-aware emotion recognition from speech is critical in affective computing applications, including virtual assistants, telehealth, and intelligent surveillance, where interpretability and demographic fairness are paramount. However, real-world speech signals often suffer from compounded challenges such as noise, whispering, short utterance duration, multilingual variability, and age-related vocal changes all of which significantly degrade the accuracy and robustness of emotion recognition models, especially in preserving gender-specific prosodic cues like pitch and energy. To address these limitations, we propose HEDRNet: a Hybrid Diffusion-Transformer Network tailored for robust and equitable emotion recognition under naturalistic speech conditions. HEDRNet comprises four synergistic components engineered to resolve the multidimensional challenges of spontaneous speech. First, the DiffuLANS module leverages a diffusion-based denoising architecture with Latent Adaptive Noise Scheduling, which dynamically adapts the reverse diffusion trajectory based on input SNR and speech type. This ensures the preservation of emotionally salient prosody even in whispered, noisy, or aged speech. Second, the enhanced signal is processed by SwinSpeechX, a Speech Swin Transformer equipped with Prosodic-Guided Feature Enhancement and Temporal Dynamic Attention, enabling fine-grained modelling of emotion-specific temporal dynamics in short utterances. Third, Multilingual Feature Alignment is integrated via alignment loss to unify emotional feature representations across languages, ensuring robustness in code-switched and multilingual scenarios. Comprehensive evaluations on IEMOCAP, RAVDESS, CREMA-D, and EMO-DB datasets demonstrate that HEDRNet achieves a PSNR-equivalent accuracy of up to 94.54%, UA of 93.8%, F1-scores above 95%, and a PRI of 0.95, while maintaining 93.8% UA under 0 dB SNR and 94.9% UA in cross-lingual tests. These results confirm HEDRNet’s effectiveness as a high-fidelity, demographically fair, and multilingual-ready solution for emotion recognition in complex speech environments.