Content ideas

[bmcfee]

• explaining resample-on-load
• introducing the block streaming interface
• how to do feature inversion
• how to make cqt's go fast
• how to use caching effectively
• why the name? / history of the project

What else should we post about?

[bmcfee]

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

[lostanlen]

• 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.

[bmcfee]

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

[lostanlen]

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.

[bmcfee]

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

[lostanlen]

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

[bmcfee]

Post ideas to accompany the 0.7.1 release:

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

[lostanlen]

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.

[bmcfee]

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