Scholarlink

Class imbalance challenges dysarthric speech recognition systems, compromising performance for minority-class speakers, particularly those with severe articulation problems. We propose FFA-GAN, a fresh architecture combining Generative Adversarial Networks (GANs) with the Firefly Algorithm (FFA) for data augmentation, to handle this. FFA dynamically optimizes GAN hyperparameters to produce high-quality, varied dysarthric speech samples, improving minority-class representation. Experiments on the UASpeech dataset show FFA-GAN lowers Word Error Rate (WER) by 14.3% compared to baseline GANs and increases minority-class F1 scores by 22% relative to conventional oversampling techniques. The adaptive optimization of the framework reduces mode collapse and stabilizes training, producing more discriminative synthetic data. These outcomes show how well bio-inspired optimization ad-dresses data imbalance and increases dysarthric speaker recognition accuracy. While preserving linguistic diversity in synthesized speech, the method offers a scalable solution for clinical applications limited in resources.

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

FFA-GAN: A Novel Firefly Algorithm-Optimized GAN Framework for Enhanced Dysarthric Speech Recognition

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kapil Bhaiyalal Kotangale,
  • Yuvraj Vijay Parkale

摘要

Class imbalance challenges dysarthric speech recognition systems, compromising performance for minority-class speakers, particularly those with severe articulation problems. We propose FFA-GAN, a fresh architecture combining Generative Adversarial Networks (GANs) with the Firefly Algorithm (FFA) for data augmentation, to handle this. FFA dynamically optimizes GAN hyperparameters to produce high-quality, varied dysarthric speech samples, improving minority-class representation. Experiments on the UASpeech dataset show FFA-GAN lowers Word Error Rate (WER) by 14.3% compared to baseline GANs and increases minority-class F1 scores by 22% relative to conventional oversampling techniques. The adaptive optimization of the framework reduces mode collapse and stabilizes training, producing more discriminative synthetic data. These outcomes show how well bio-inspired optimization ad-dresses data imbalance and increases dysarthric speaker recognition accuracy. While preserving linguistic diversity in synthesized speech, the method offers a scalable solution for clinical applications limited in resources.