from Hacker News

Llasa: Llama-Based Speech Synthesis

by CalmStorm on 5/1/25, 4:43 PM with 22 comments

• by ks2048 on 5/1/25, 8:03 PM

  Odd that the page doesn't seem to link to either,

  paper: https://arxiv.org/abs/2502.04128

  github: https://github.com/zhenye234/LLaSA_training

• by CalmStorm on 5/1/25, 4:43 PM

  LLaSA is a simple framework for speech synthesis that employs a single-layer vector quantizer (VQ) codec and a single Transformer architecture to fully align with standard LLMs such as LLaMA.

• by dheera on 5/1/25, 8:02 PM

  > employs a single-layer vector quantizer (VQ) codec and a single Transformer architecture to fully align

  I really wish when new models were released that they would draw a diagram of all the layers and the tensor input and output sizes at each layer, with zoom in/out capabilities if needed using D3.js or whatever visualization framework if needed. Every single layer should be on there with its input and output sizes.

  These one-sentence descriptions, and approximate block diagrams with arrows pointing at each other are never enough to understand how something is actually implemented.

• by StevenNunez on 5/1/25, 6:00 PM

  I can't wait see this integrated into Open WebUI! These sound amazing.

• by mring33621 on 5/1/25, 7:08 PM

  the long 'uuuuhhhhhhh' from some of the lesser models is killing me.