Local desktop AI girlfriend demo fits Whisper, llama.cpp and Qwen3-TTS into 15GB VRAM

xiaosun86 · x · 2026-07-26

A local-only AI girlfriend/desktop companion demo combines speech-to-speech components without using the cloud.

The cited setup uses Silero VAD v5 for voice activity detection, Whisper for STT, llama.cpp for the LLM, and Qwen3-TTS for speech synthesis. The creator says all four models fit into 15 GB of VRAM and can be hot-swapped freely on a fully local stack built on an open-source speech-to-speech project.

Original post →

More from Fun

Fun channel →