Coming Soon: Chord Intelligence mkIV

I’m not one for pre-announcing features or releases, but 2020 has been a very unusual year so far, so let’s just go with it.

I first wrote about Capo’s Chord Intelligence technology over 6 years ago, where I described the third major revision of Capo’s chord detection technology. A few years later in March 2018, I spoke a bit about major updates to Capo’s MIR (Music Information Retrieval) engine that laid the groundwork for the subject of this post.

This fall we’re going to ship a huge update to Capo for macOS and iOS. Among other things, it includes an all-new chord detection engine that mops the floor with any of our previous iterations. It runs faster, it’s far more accurate, and detects a wider range of chords than before. There’s one more big trick up its sleeve, but I’m going to save that detail for the launch.

Because we’re shipping after the release of macOS 11 Big Sur, and because the development story is quite long, I’m going to split this up over a few posts. To whet your appetite for the upcoming blog entries, I’ll summarize some of the story below.

During WWDC 2018, Apple (formally) introduced new API that made it possible to train deep neural networks on the GPU. That, combined with the recent purchase of an iMac Pro, kicked off a journey that led me to build my own custom neural network training environment for my MIR research and development.

I didn’t quite build my own TensorFlow, PyTorch, or Turi create, but it sure felt like it. I used the low-level APIs that were intended for those packages to perform GPU-accelerated training on the Mac.

This work happened (sporadically) over the next 18 months, and it came to a head over the 2019/2020 holiday season when I felt compelled to acquire two beefy eGPUs to help me wrap up my research.

In the next installment, I'll share more about the challenges we faced while expanding our training data set.