Colin Raffel

Loading seems to be 10-50x faster than deepdish/h5, size is comparable to compressed h5, and types are preserved completely: ``` Python import msgpack import msgpack_numpy msgpack_numpy.patch() import numpy as np...

Better docstrings for separation

2

Currently, the docstrings for the metrics in `mir_eval.separation` are not super helpful - they basically say "Computes xyz function from bss_eval". Specifically, this makes it unclear what the difference between...

Implement more separation metrics

29

Although `bss_eval` is pretty standard, there are more: http://bass-db.gforge.inria.fr/peass/ http://labrosa.ee.columbia.edu/projects/snreval/

Chroma transcription metrics

7

To cover all evaluation done in MIREX, we also need to add the ability to evaluate transcription annotations after mapping the pitch values to a chroma (single octave) scale, as...

We should check that metric parameters are provided with sane values, and throw helpful errors when they are not.

Writing and reading back in should be roughly lossless

11

I.e. if you write a MIDI file out and read it back in, you should get at least the same collection of notes, time signature changes, tempo changes, pitch bends,...

Feature extraction

3

Some enthusiastic individual might be interested in implementing all of the features described in section 4.5 here: http://jmir.sourceforge.net/publications/PhD_Dissertation_2010.pdf

The estimate_* functions (`estimate_tempi`, `estimate_tempo`, and `estimate_beat_start`) are now out of the scope of `pretty_midi`. They should be deprecated, and moved to separate examples in the example folder.

We need tests.

Instead of djitw. This simplifies dependencies. Plus, librosa's DTW has actual tests.