[LAU] Audio to midi, finding notes from an mp3

[Sebastian Gutsfeld]

Chris Cannam <cannam at all-day-breakfast.com> writes:
> On Monday 24 September 2007 00:43, david wrote:
>> Not a single one produced anything usable! They were able to identify
>> notes, and output them as MIDIs, but the output was pretty useless:
>> every note of every instrument on the same staff.
>
> That's not a bad result for such a tough problem.  Even monophonic note 
> tracking is tricky, depending a great deal on the instrument. Note tracking 
> in a dense polyphonic mixture is much harder (if it's possible at all).

I've recently written a transcription system for polyphonic piano
music. It's based on the system described in "A Discriminative Model for
Polyphonic Piano Transcription" by Poliner and Ellis which uses support
vector machines for each piano note trained on spectral data to predict
sounding notes in small timeframes and hidden markov models to
temporally constrain those classification results. The results are
promising but not as good as I hoped (and sadly not as good as Poliner's
and Ellis'). I have a few ideas to improve the system but as I'm
currently searching for a job I have no time to implement those so I am
thinking about making my software open source (ATM it's a Python library
with command line tools).

If trained with other audio data than piano recordings it should also be
possible to transcribe other instruments but the restriction to one
instrument without slides, vibratos or anything like that remains.

   Sebastian