Full-duplex, real-time dialogue with Kyutai

In this episode of the Practical AI podcast, Alexandre De Fussey, co-founder of the non-profit AI lab Kyutai, shares insights about their innovative open-source speech model, Moshi. The conversation highlights Kyutai's distinctive role in the French AI landscape, the challenges and breakthroughs in developing a full-duplex, low-latency speech model, and the data sets that powered Moshi's training. They also explore future research directions in real-time speech interaction and advancements beyond transformers. Additionally, the episode features sponsorship segments from Fly.io and Timescale.

Outlines

Sign in to continue reading, translating and more.

Open full episode in Podwise

Practical AI

Introduction and Sponsor Mentions

Introducing Kyutai and its Open-Source Approach to AI

Open Science and the Kyutai Approach

Introducing Moshi: A Real-Time, Full-Duplex Speech Foundation Model

The Evolution of Real-Time Speech Models

Moshi's Data and Training Methodology

Future Directions in AI and the Post-Transformer Era

Closing and Outro

Full-duplex, real-time dialogue with Kyutai

Practical AI

00:03Introduction and Sponsor Mentions

Introduction and Sponsor Mentions

03:15Introducing Kyutai and its Open-Source Approach to AI

Introducing Kyutai and its Open-Source Approach to AI

10:30Open Science and the Kyutai Approach

Open Science and the Kyutai Approach

19:04Introducing Moshi: A Real-Time, Full-Duplex Speech Foundation Model

Introducing Moshi: A Real-Time, Full-Duplex Speech Foundation Model

24:20The Evolution of Real-Time Speech Models

The Evolution of Real-Time Speech Models

37:12Moshi's Data and Training Methodology

Moshi's Data and Training Methodology

45:56Future Directions in AI and the Post-Transformer Era

Future Directions in AI and the Post-Transformer Era

49:15Closing and Outro

Closing and Outro