If I was to carry on with it, I would look into using byteFrequencyData rather than timeDomainData. My initial attempts with the frequency data weren't great, but a few people have mentioned it so I'd take another look and see if I just wasn't doing something dumb.

I felt like the most fiddly work would be accurately sampling the sound. E.g. not recording when it was under (or maybe over) a particular volume, ignoring the initial thud of a key (particularly with the phone sitting on the piano), and making it so that the correct 'note' stayed blue as the sound trailed off and got less reliable.