ZYNAPTIQ PITCHMAP VERSION HISTORY 1.6 CODE

You imagine a sound you want to make, interpret it with your voice, and start singing. Your thoughts are the encoder, your voice is the embedding, and the resonance of the room is the WaveNet decoder. Imagine standing in a fun-house version of an echo chamber.

ThereвЂ™s a physical analogy that can be helpful in thinking about the system. Viewed from a different perspective, the WaveNet is actually a very nonlinear infinite impulse response (IIR) filter that is being driven by the embeddings.

ZYNAPTIQ PITCHMAP VERSION HISTORY 1.6 CODE

Since we literally add this latent code to the layers of the WaveNet decoder network at generation time, it works like an excitation/activation of a nonlinear feedback system, and the network learns to be silent unless it receives this excitation. Even some expressive information like vibrato is visible in the Trumpet figure above as quick fluctuations. ItвЂ™s interesting to note that the embeddings follow a similar dynamic envelope of the original sounds. You can find more details in the original post and paper. The whole system is trained so that the reproduced sound is as similar as possible to the real sound. A decoder network then converts it back into sound. In this simplified diagram, you can see that an encoder network transforms a sound into its z representation.

As a quick reminder, the NSynth algorithm works by finding a compressed representation of sound (letвЂ™s call it вЂњzвЂќ). We also had a great time developing these instruments and wanted to share some of our process in making a creative musical tool out of a generative algorithm. We hope that these instruments are one small step in that direction, and weвЂ™re really excited to see what you make with them. Instead of using AI in the place of human creativity, we strive to infuse our tools with deeper understanding so that they are more intuitive and inspiring. We want to empower creators with tools built with machine learning that also inspire future research directions. The goal of Magenta is not just to develop new generative algorithms, but to вЂњclose the creative loopвЂќ.

Zynaptiq pitchmap version history 1.6