Orpheus is good. It's a very simple architecture, basically just llama3b trained to produce audio (SNAC) tokens instead of text tokens. This is nice because it benefits from the Llama open-source ecosystem which is very active. It being LLM based is also nice because it streams output and with some engineering effort supports continuation via prompt-filling (so streaming input can work).
The other one people have been excited about is Sesame. Both are very expressive and have much more room for optimization (both on the inference efficiency side quants, finetunes, distillation, etc) and on the quality side (throw more data at it). Exciting times!
The other one people have been excited about is Sesame. Both are very expressive and have much more room for optimization (both on the inference efficiency side quants, finetunes, distillation, etc) and on the quality side (throw more data at it). Exciting times!