Speech + LLMs @nvidia | Previously: @Meta MSL, @jhuclsp, @IITGuwahati
Full-duplex/realtime/interaction models have suddenly become popular due to TML's demos. I thought it would be a good idea to create a (visual) primer on what these models are.
1/ It's a 276B-A12B MoE, and they "plan to release larger models" later this year. The post already talks about the complexities of frequent prefills and decodes required for 200ms chunks. As these mo…
Google released several variants of Gemma-4 under an Apache 2.0 license yesterday. Of these, the smaller "edge" models (E2B, E4B, and their instruction-tuned variants) come with audio encoders. I was …
since ASR is suddenly cool again due to @cohere Transcribe, let's talk about what lessons are illustrated from its model design.
