honwee/mynah is a real-time digital human project that runs on your own server. The latest release recorded in the repository is v0.1.0, and the provided information does not list what changed in this version. According to the repository description and README, a core written in Go delivers the avatar worker's video and audio to the browser over WebRTC. It was released under the Apache-2.0 license.
The project focuses on connecting avatars with speech recognition and synthesis and language models to provide voice-to-voice conversation in the browser. The core opens a bidirectional gRPC session with the avatar engine and sends TTS audio along with encoded video and audio frames to the browser. An ASR worker and a turn-detector worker that determines the end of an utterance also take part in the conversation flow. Avatar, TTS, ASR, LLM and turn detection form a replaceable plugin architecture. When the database option is enabled, you get an admin API, configuration changes applied immediately, a knowledge base and channel publishing. Workers use gRPC streaming, and the TTS and LLM connections support OpenAI-compatible HTTP.
The operating scope depends on the hardware and on how sessions are handled. The README states that the default FlashHead engine requires an NVIDIA GPU in the 4090 class, while the lighter wav2lipLS profile targets low-spec GPUs. One avatar worker handles one live session, and even without a GPU the components can be configured as remote endpoints. The README contains no performance figures or benchmarks, so actual latency and image quality cannot be confirmed. In other words, the current material reveals the feature composition and the operating conditions, but makes it hard to judge real-time quality.