Enhancing Dysarthric Speech for Improved Clinical Communication: A Deep Learning Approach

Dysarthric speech poses significant challenges to modern speech processing systems due to its inherently low intelligibility, irregular prosody, and atypical articulation patterns. Traditional methods that rely on intermediate automatic speech recognition (ASR) stages often perform poorly under such conditions, especially when speech is severely degraded. In this work, we propose a fully end-to-end enhancement pipeline that directly improves dysarthric speech quality and intelligibility using GAN-based models, bypassing the limitations of transcription-based systems. We employ a MelSEGAN architecture coupled with a SepFormer to address spectral and temporal distortions in the speech signal. Through a comparative analysis of preprocessing strategies, we find that dynamic time warping (DTW) in conjunction with variational mode decomposition (VMD) offers more stable and intelligible outputs than conventional voice activity detection (VAD), particularly in cases of temporally misaligned or fragmented speech. DTW not only enables better convergence during training but also results in clearer formant structures and reduced background artifacts in the enhanced speech. Further, we extend our pipeline with Model-Agnostic Meta-Learning (MAML) to improve speaker-specific adaptation. The MAML-augmented models demonstrate superior generalization and refinement of harmonic features, especially when paired with DTW-based preprocessing. Additionally, we are investigating an alternative enhancement path that combines a UNet-based encoder-decoder with a HiFi-GAN vocoder. Early qualitative assessments suggest that this hybrid model produces higher naturalness and improved intelligibility, offering a promising direction for future development. Overall, our findings highlight the importance of robust temporal preprocessing and adaptive learning strategies in building effective enhancement systems for disordered speech scenarios.