Athena
A fully offline, privacy-first voice assistant that combines a local mixture-of-experts language model, neural text-to-speech, real-time speech recognition and an audio codec in a C++ pipeline with no Python at runtime.
Athena is an offline voice-to-voice assistant that runs entirely on local hardware, combining a mixture-of-experts language model, neural text-to-speech, real-time speech recognition and a neural audio codec in a four-process C++ pipeline. It speaks with emotional inflection, reads and responds to emotion in the user's voice, remembers across sessions through an evolving long-term memory, and can be interrupted mid-sentence.
Its text-to-speech path streams tokens over SSE with incremental decoding, so audio for a chunk begins around 200 to 400 milliseconds after submission, and playback stays gapless through a persistent raw-PCM audio sink. A health watchdog restarts the text-to-speech backend after a crash, and the installer hardens the NVIDIA driver configuration.
It needs a single consumer GPU with system RAM, a recent NVIDIA driver, CUDA and cuDNN; models are not bundled and are fetched by an idempotent, resumable install script that downloads about 179 GB of data.