PersonaPlex
A speech-to-speech translation model in the NVIDIA Nemotron speech model family. It supports full-duplex processing.
Timeline (last 90 days)
Moshi and its derivative model PersonaPlex are speech-to-speech LLMs capable of listening and speaking simultaneously, enabling concurrent generation.
In tests with five-minute segments of pure silence as input, Moshi exhibited inappropriate speech onset in 12 out of 40 tests, and PersonaPlex in 11 out of 40.
The suppression method successfully suppressed all 13/13 false starts in Moshi and all 9/9 false starts in PersonaPlex.
The method preserved all 40/40 genuine responses on each of the two models, with zero false suppressions.
When behavior needs to be inferred from a persona, F-Actor and PersonaPlex experience drops in instruction-following of 9.7% and 4.5% respectively
The research team conducted tests on six mainstream real-time voice systems, including F-Actor, PersonaPlex, GPT-Realtime, MiniCPM-o, and Fun-Audio-Chat
All Facts (4)
Moshi and its derivative model PersonaPlex are speech-to-speech LLMs capable of listening and speaking simultaneously, enabling concurrent generation.
50%UnverifiedIn tests with five-minute segments of pure silence as input, Moshi exhibited inappropriate speech onset in 12 out of 40 tests, and PersonaPlex in 11 out of 40.
50%UnverifiedThe suppression method successfully suppressed all 13/13 false starts in Moshi and all 9/9 false starts in PersonaPlex.
50%UnverifiedThe method preserved all 40/40 genuine responses on each of the two models, with zero false suppressions.
50%