So happy to have revealed moshi, our new voice AI earlier today. If you miss it, you can see the keynote here: https://lnkd.in/d_tZWdNv And try out the model at https://lnkd.in/epAb-EeZ or https://lnkd.in/esRx5Gkw for US based users that want better latencies.
Unveiling of Moshi: the first voice-enabled AI openly accessible to all.
https://www.youtube.com/
Great start. I have tested #moshi in several user cases including Tutoring for students. I guess the focus has been on the vocal "wahou" effect which is quite well executed. But on the content production itself (the LLM DNA), progress need to be made as the outpiut to the prompt is not at the level of other models such as LLama3 , Mistral or GPT'4. Hope to see huge progress from this team in the next months 🚀
Xavier Lioneton, Marc de Vaugiraud , Raphaël Vienne , Thaïs Denoyelle , Rémy Gasmi
Looks interesting, why it starts answering the questions even before the question is complete (maybe we are too slow for it). It also seems to ignore the human commands. Has the model or its source been made available anywhere?
Super, ça marche en Français ?
I don't understand why, but this is completely inaccessible to me?
Awesome demo. I wonder will you open source your models? And if yes, when?
moshi moshi, brilliant name! and excellent keynote that we actually learnt something about the inner workings
Deeptech | Semiconductor Startup Executive | Artificial Intelligence | Corporate Development Strategy | Venture Capital
2moAmazing progress made by the Kyutai team in 6 months and with only 8 people since we discovered you at STATION F in November. Congrats ! Also very interesting to see how the demo quality shifted when trying to run it on device, without cloud… [~37:00 in the video]. The massive model compression, quantization, audio compression, so this could fit in the limited memory and gpu of the laptop, does have an impact on the experience. I see an opportunity for a better way to integrate compute and memory here.