-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 On 10.05.2014 00:18, Lucas Takejame wrote: I would be interested in trying a different approach to this problem. Namely using machine learning. Would the guitar players on this list be willing to provide training data? I.e. audio files of you playing single notes with an additional text file describing the played pitches. A format like frame-number start 440 frame-number end e.g. you have an audio clip at 1000hz sampling rate where an 440hz A starts on frame 500 (at 0.5 secs) that lasts to frame 2500 (so, 2 secs duration): 500 start 440 2500 end And so on.. Also for a usable guitar synthesizer one would maybe have to go a step further. The closer one is to the onset of the note the more uncertain the pitch is (loosely and much simplified speaking: for a low note it takes quite a while until even half a cycle has passed - during that time the pitch is very uncertain), so one approach would be to use a percussive sound for the start of the note (onset detection is easier than pitch detection I would figure)

and then once the pitch is established fade to a second synthesis model that produces a clearly defined pitch. Mapping this to a midi stream introduces some loss of information, but could also be possible (e.g. two midi outs, one for the percussive onset detection and one for the (slower) pitch detection). I wanted to try this for a long time.. Maybe using decent machine learning and lots of training data one could even identify the pitch based on the characteristics of the very early signal from a note, but I'm not really certain about that. Sorry for hijacking your post and talking about several separate things at once, but maybe someone is interested in a project like this.. Flo

I would be interested in trying a different approach to this problem. Namely using machine learning. Would the guitar players on this list be willing to provide training data? I.e. audio files of you playing single notes with an additional text file describing the played pitches. A format like frame-number start 440 frame-number end e.g. you have an audio clip at 1000hz sampling rate where an 440hz A starts on frame 500 (at 0.5 secs) that lasts to frame 2500 (so, 2 secs duration): 500 start 440 2500 end

-----BEGIN PGP SIGNED MESSAGE----- Hash: SHA1 On 10.05.2014 00:18, Lucas Takejame wrote: I would be interested in trying a different approach to this problem. Namely using machine learning. Would the guitar players on this list be willing to provide training data? I.e. audio files of you playing single notes with an additional text file describing the played pitches. A format like frame-number start 440 frame-number end e.g. you have an audio clip at 1000hz sampling rate where an 440hz A starts on frame 500 (at 0.5 secs) that lasts to frame 2500 (so, 2 secs duration): 500 start 440 2500 end And so on.. Also for a usable guitar synthesizer one would maybe have to go a step further. The closer one is to the onset of the note the more uncertain the pitch is (loosely and much simplified speaking: for a low note it takes quite a while until even half a cycle has passed - during that time the pitch is very uncertain), so one approach would be to use a percussive sound for the start of the note (onset detection is easier than pitch detection I would figure) and then once the pitch is established fade to a second synthesis model that produces a clearly defined pitch. Mapping this to a midi stream introduces some loss of information, but could also be possible (e.g. two midi outs, one for the percussive onset detection and one for the (slower) pitch detection). I wanted to try this for a long time.. Maybe using decent machine learning and lots of training data one could even identify the pitch based on the characteristics of the very early signal from a note, but I'm not really certain about that. Sorry for hijacking your post and talking about several separate things at once, but maybe someone is interested in a project like this.. Flo