Seonglae ChoRVQ
‣
Residual Vector Quantization is a Data Compression technique found in SOTA Neural Audio Codec which in turn form the backbone of generative audio models.
RVQs
Vector quantization is a sparse MLP
Vector quantization is a sparse MLP
In this post, I argue that vector quantization, when combined with an embedding layer, is equivalent to a sparse multi-layer perceptron layer.
https://euclaise.xyz/vq-is-mlp/
blog - Residual Vector Quantization
An exploration of the heart of neural audio codecs
https://drscotthawley.github.io/blog/posts/2023-06-12-RVQ.html
