Athena: fully offline voice assistant with emotion and memory
No cloud, no telemetry: a 100% local voice assistant that laughs and remembers
Igor Barshteyn has released Athena, a fully offline voice-to-voice assistant that runs entirely on local hardware. The system combines a large mixture-of-experts language model (Qwen3.5-397B), neural text-to-speech (Orpheus 3B), real-time speech recognition (Whisper-small.en), and a SNAC neural audio codec into a four-process pipeline — all implemented in C++ with zero Python at runtime (one optional Python script handles a one-time emotion2vec model conversion at setup). Athena is designed for maximum privacy: no cloud, no telemetry, no API keys, and it runs on a single consumer GPU plus system RAM.
Athena stands out for its advanced conversational abilities. It speaks with natural emotion — laughs, sighs, gasps — and can read the basic affect in your voice to respond accordingly. The assistant remembers across sessions with evolving long-term memory and a personality that persists between conversations. It maintains long conversational context and is interruptible mid-sentence: speak over Athena and it stops, keeping everything it already said in context. The current configuration is tuned to be friendly and connection-seeking, but these parameters can be adjusted via system prompts in the code. A demo video shows two sessions where Athena recalls planted memories, demonstrating its persistent memory capabilities.
- Uses Qwen3.5-397B mixture-of-experts LLM, Orpheus 3B TTS, Whisper-small.en speech recognition, and SNAC audio codec
- All C++ runtime with zero Python — one optional script for one-time model conversion
- Features emotion detection and response, long-term memory across sessions, and mid-sentence interruptibility
Why It Matters
Enables truly private, emotionally aware AI assistants that run locally, no cloud dependency needed.