Winter Week 5 - Template Matching Model

This week, my partner and I were able to finish implementing the chroma feature generation and template matching model for chord recognition on the LCDK. Then, we began testing the model with our dataset, which consists of several audio files for each of the 24 guitar chords (major and minor).

Our code was adapted from the template matching model described in the paper "Real-Time Chord Recognition for Live Performance". The paper uses a complementary template whose dot product with the chroma vector is minimized, rather than maximized. This achieves the same effect when calculating similarity scores, but is more robust to noise, due to greater sensitivity in the dot product.

I used my laptop to output the audio using the analog line out, and input to the LCDK using the line in. This way, we were able to use the same audio codec for sampling as in the MFCC project. We can extend this to live recordings through the microphone easily as well. We found that the template method was pretty accurate for our pre-recorded audio, as it matched to most of the audio files in our preliminary testing. We believe that we won't have to use a more complex model for chord recognition.

Our next step is to test with live audio from the microphone, which is the last part of our week 7 milestones. After this is complete, we hope to generate sheet music from these live recordings.