Request: Calibration function


Hey, I’ve had a little time to play around with my muse and it’s pretty evident that the calibration routine that runs in your own app makes a world of difference for reading data from the user’s brain.

A question to the Muse-team, to which I haven’t stumbled on the answer, is if you have any plans on releasing the calibration function so we all can use it in our apps? I really don’t see why it’s worth keeping it constrained to apps included with the Muse when it could be a much greater platform if we all could use it? Would be a great incentive to develop apps. Currently, there’s a huge hurdle to overcome, namely getting any comprehensible and usable signals from the Muse. That will probably keep most potential developers from even attempting to build anything.

Or have i missed some relevant piece of information?


Hey alinder,

Some type of calibration is definitely an important component of most BCI applications, agreed! It’s definitely something we’re seriously considering including in future updates to the SDK. Of course the algorithm in the official Muse application is proprietary, though. It will not be included. Can’t say much more than that right now, though.

Building your own calibration function is not necessarily that difficult. Not much more than building an algorithm, which is something many developers will have to do (BCI algorithms usually have to be task-specific). Once you have a classifier/algorithm, including a brief recording session for calibration ahead of the main activity is not, relatively speaking, a major technical obstacle.


Hi Alinder,

I will eventually publish a blog article on this topic (,, but I’ll quickly go over the way I do it.

I begin by recording 15 seconds of my subject’s awake brain EEG. I then compute basic statistics on the Fourier transform of the signal to know the mean and standard deviation of each frequency bins I’m interested in profiling. I then use these parameters to transform every subsequent samples into z-scores (or standard-score, this helps to adapt the reference frame for a new subject, on the fly.

Alternatively, if you look at the shape of the signal (I consider only the alpha wave, you can use the adjacent frequencies, which remain unaffected by mediation, as the reference frame and extrapolate the awake brain state from them and then quantify the strength of the alpha wave. This method should work without a training sequence, but I haven’t implemented it yet and you should expect it to have slightly worst results than the trained approach. That’s the necessary trade-off when looking for a shortcut.

These are the most basic approach I can think of. the particularities of the eeg signal signature you want to capture will influence the calibration procedure. This one work very well for the alpha wave in occipital lobe and other obvious signals. Search in the blog I linked, I published an offline data analysis method that is more powerful and based on the FDA.