Moshi Chat is a new native speech AI model developed by French startup Kyutai, offering a similar experience to GPT-4o but with the ability to understand tone of voice and be interrupted. Unlike GPT-4o, Moshi is a smaller model that can be installed locally and run offline, making it suitable for smart home appliances in the future. However, in online demos, conversations with Moshi tend to lose cohesion after five minutes, with some instances of argumentative behavior.
The Kyutai research lab built Moshi from scratch six months ago, aiming to create an open and expandable model for generative voice AI. The core functionality of Moshi is comparable to GPT-4o but from a smaller model, available for immediate use. The team envisions applications for Moshi in roleplay scenarios or as a motivational coach during training sessions.
Moshi, a 7B parameter multimodal model named Helium, is trained on text and audio codecs, specializing in speech in and speech out. The team plans to collaborate with the community to enhance Moshi’s knowledge base and factuality, enabling more sophisticated and extended conversations. The next steps involve refining the model and scaling it up for improved performance.
While Moshi is not a direct competitor to OpenAI’s GPT-4o advanced voice, it represents a significant advancement in open-source AI development by offering a locally running model with similar capabilities. The potential for Moshi to evolve into a powerful assistant through community support and further development is promising.
Source link
Source link: https://www.tomsguide.com/ai/moshi-chats-gpt-4o-advanced-voice-competitor-tried-to-argue-with-me-openai-doesnt-need-to-worry-just-yet