Autoencoder slashes SSL speech model training time by 8x for dysarthric ASR

A team of researchers led by Paban Sapkota (Aalto University, USC) introduced an autoencoder-based bottleneck layer for self-supervised learning (SSL) speech representations, targeting the specific challenge of dysarthric automatic speech recognition (ASR). SSL models like wav2vec 2.0 produce rich but high-dimensional features that are computationally expensive to process. By feeding these features through a carefully designed autoencoder, the team compresses them into a much smaller latent space while retaining the information critical for recognizing dysarthric speech patterns. This SSL-AE bottleneck reduces training time by a factor of 8 compared to the full SSL model, cutting GPU hours from days to hours without sacrificing accuracy.

The experiments, detailed in the arXiv preprint (2606.24088), show that the compressed representations actually improve Word Error Rates (WER) on standard dysarthric speech benchmarks. The autoencoder's reconstruction loss forces the bottleneck to capture only the most salient phonetic and prosodic cues, effectively denoising irrelevant variations. This makes the approach particularly valuable for deployment in clinics, mobile devices, and low-resource settings where computational power and labeled data are limited. The work demonstrates that careful feature compression can simultaneously reduce cost and boost performance, offering a practical path toward inclusive speech interfaces for people with speech impairments.