Content ideas about blog HOT 9 OPEN

• i would love to see a post about waveplot and how it keeps a low memory usage. The email to John Davis from a few days ago (copy-pasted with permission on this StackOverflow board) could be used as a starting point: https://stackoverflow.com/questions/57260065/why-does-librosa-plot-differ-from-matplotlib-and-audacity

• as i wrote earlier on the Slack channel, librosa.util.valid_audio is one of my most cherished yet under-appreciated librosa features. I would be in favor of giving it more publicity :) we could write a short post about that.

from blog.

We should probably do a comprehensive post about all the various efficiency improvements in the 0.10 release.

from blog.

@lostanlen I also drafted up a post on block streaming, care to give it a looksy before I publish?

from blog.

Maybe also a plot on all the bells and whistles of specshow, compared to (say) imshow.

from blog.

Yes, specshow is a major feature in librosa, and some uses are not trivial. For example:

• Explaining how the color map gets assigned (magma vs diverging) + a cautionary word against jet?
• How to make sure that the axis ticks show physical units, not frame indices (one of my biggest pet peeve as a reviewer!).
• How to make sure that the contrast is right.
• How to have several spectrograms share the same colormap and the same axes
• How to make the y-axis in CQT specshow be scientific pitch
• How to annotate a spectrogram with piptrack
• How to annotate a spectrogram with beat_track

This covers probably more than one post, but it's all very relevant IMO.

from blog.

SGTM in general! I'm not sure what exactly you have in mind for piptrack, can you elaborate?

from blog.

What i mean is to have a pitch curve (or spectral centroid, or whatever frequency-valued signal) overlayed on top of the CQT

from blog.

Post ideas to accompany the 0.7.1 release:

• framing
• feature inversion (incl cqt griffin-lim)
• replicating HTK features
• pitch tracking and plotting?

from blog.

replicating HTK features

Yes. I think that the big take-home message here is that there is no such thing as "the MFCC" of a signal. It all depends on what well-established implementation we're trying to replicate, which ultimately falls back on which psychometric study we're relying on for the mel scale.
This (under-cited) paper about "fitting the mel scale" was a big eye-opener for me in this regard:
https://ieeexplore.ieee.org/document/758101

Pitch tracking and plotting: yes! will submit a WIP PR tonight.

from blog.