Case Study: University of Texas at Austin achieves 96% imagined-speech decoding accuracy with MathWorks (MATLAB) wavelets and deep learning

UT Austin Researchers Convert Brain Signals to Words and Phrases Using Wavelets and Deep Learning

The University of Texas at Austin faced the challenge of decoding imagined speech from magnetoencephalography (MEG) signals to give ALS patients a faster, noninvasive communication channel. Researchers needed to denoise and analyze hundreds of trials, preserve signal characteristics across neural frequency bands, and rapidly iterate on machine learning and deep learning approaches. To meet these needs they used MathWorks’ MATLAB platform and associated toolboxes (Wavelet Toolbox, Deep Learning Toolbox, Statistics and Machine Learning Toolbox, and Parallel Computing Toolbox).

Using MathWorks’ tools, the team applied wavelet multiresolution analysis and scalograms to extract time‑frequency features, trained SVM and shallow ANN baselines, and then fine‑tuned pretrained convolutional networks (AlexNet, ResNet, Inception‑ResNet) on the scalograms, training on a seven‑GPU server. The MathWorks workflow produced a measurable leap in performance—classification accuracy rose from ~80% with classical methods to 96% with wavelets plus deep learning—and training ran about 10× faster by switching to multi‑GPU training, enabling rapid prototyping and iteration.

University of Texas at Austin

Jun Wang

Associate Professor Of Communication Sciences & Disorders And Neurology And Director