Talk in English - UK at Rust Nation UK 2025
Track Name:
The Fastly Track - Talk Track
Short URL: https://joind.in/talk/934e2
(QR-Code (opens in new window))
This session will take you through the development of Moshi, the first open-source speech-to-speech model, highlighting the crucial role Rust played in its success. While the primary training loop was built in Python, Rust was instrumental in optimizing the entire pipeline—handling data pre-processing, data loading, and powering the entire inference stack. We'll explore key decisions, including the shift from traditional frameworks like PyTorch to Rust/Candle for inference, and how we benefited from this choice, allowing us to serve over 500'000 Moshi sessions last summer on a low-cost GPU cluster.
Comments
Please login to leave a comment