Making Neural Decoders Robust to Variable SNR

Non-invasive speech decoding suffers from the problem of lower signal-to-noise ratio (SNR) than invasive methods. Trial averaging is a standard way to improve SNR, but each additional trial folded into an average reduces the number of training examples, creating a tradeoff between averaging level and training data. To investigate this tradeoff, we trained fixed-averaging models at various levels of averaging and tested them on those levels. Models tested on a certain level of averaging almost always performed best when the training level was the same or a similar level of averaging. This raises the problem of how to train models to be robust to different SNR. We tried different methods to improve robustness, and we found that exposure to higher averaging levels during training via stochastic and scheduled averaging improves robustness relative to a step-matched baseline, but not relative to a full-data baseline. How best to tackle the dilemma of controlling for data versus optimiser steps can be the subject of future work.