80% accuracy is potentially fine for a time series, assuming the errors are random rather than specific to a user. You can denoise this stuff, and the recognizer should get better as the sample size grows.