In a short period of time, French AI startup Kyutai has developed an intriguing new conversational agent called Moshi. Unlike many other bots, Moshi is designed to understand the tone and emotion in a person’s voice during a dialogue. It can listen and respond at the same time without lag, a capability that could make conversations more natural.
Moshi replicates the nuances of human speech thanks to training on synthesized dialogs and collaboration with a professional voice actor. Though small compared to other language models, Moshi can carry out conversations using different accents and emotional styles. Its developers say the goal is for the technology to be open-source so people can use Moshi without privacy concerns.
One of Moshi’s standout features is its lightning-fast response time of just 200 milliseconds. For comparison, ChatGPT’s upcoming advanced voice mode is projected to take over 200 ms on average. Moshi was trained on 100,000 sample dialogs using text-to-speech to achieve this speed.
While Moshi is in early stages compared to behemoth projects like GPT-3 and ChatGPT, it shows promise as a prototype. Kyutai hopes to integrate Moshi with an audio identification system for more robust conversations. If open-sourced as planned, Moshi could offer a locally-run alternative to centralized AI models while continuing to get smarter through community contributions over time. Only continued development will show how Moshi may shape the future of human-AI interaction.
