xszyou/fay is an agent framework that connects digital humans — virtual characters that talk and behave like people — to business systems. It handles 2.5D and 3D avatars, mobile devices, PCs and the web, OpenAI-compatible APIs, and large language models (LLMs) such as DeepSeek within a single framework. There is no publicly available latest release, so it cannot be confirmed that a new version has come out.
The core of this framework is combining automatic speech recognition (ASR), language models, and text-to-speech (TTS) to drive the avatar. ASR is recognition technology that turns a person's speech into text, and TTS is synthesis technology that turns text or audio signals into speech. The README describes Fay as an open-source framework that links digital human models with large language models and lets the TTS and ASR models be swapped. It supports offline use, streaming, and concurrent multi-user processing, and integrates with microcontrollers, apps, websites, large screens, and external business systems. Text and voice interaction, digital human driving, administrative control, automated broadcasting, and intent interfaces are listed as its features.
The implementation language is Python and the license is GPL-3.0. The repository was created on June 13, 2022, and has 13,581 stars.
However, only the README has been confirmed. The features summarized in the README are merely the scope the project presents, and the given record contains no measured performance figures, benchmarks, or actual deployment cases. The evaluation scores are low as well: only the Momentum entry is filled in, at 0.8 out of 10, while the novelty, technology, and quality entries are blank. Under the GPL-3.0 license terms, if you plan to combine it with other code for distribution, checking the conditions comes first. What can be confirmed right now is only that it is a digital human framework; actual response quality and concurrent processing scale need to be checked separately.