A talking anime companion in under 500 lines of Python

What it does Riko is a push-to-talk voice companion: press Enter, speak, press Enter, and she replies out loud while a 3D VRM avatar in your browser lip-syncs along. The loop is Faster-Whisper (local or Groq) for ears, OpenAI GPT for the brain, and GPT-SoVITS or Fish Audio for the voice. Chat history persists in chat_history.json, so conversations pick up where they left off.

The interesting bit The whole thing is under 500 lines, with each pipeline stage isolated in one file — riko/asr.py, riko/llm.py, riko/tts.py, riko/viewer.py. It’s less a product and more a readable reference implementation of the voice-companion pattern, which is exactly what makes it forkable. Personality, avatar, and voice all live in a single character_config.yaml.

Key highlights

• Full loop from your voice to hers: ASR → LLM → TTS → lip-synced VRM avatar
• Local speech recognition via Faster-Whisper; local voice cloning via GPT-SoVITS (NVIDIA GPU recommended)
• Swap personality, avatar (any .vrm), voice, and model in one config file
• Pre-flight setup checker catches missing keys, servers, or files before launch
• MIT licensed, Windows and Linux

Caveats

• This repo is a frozen “50K snapshot” — no updates, and GitHub issues aren’t monitored, so there’s no support here
• The brain always requires an OpenAI API key; there’s no fully offline mode
• Newer features (hands-free listening, expressions, long-term memory, VR) are gated behind the paid Patreon tiers

Verdict Great if you want a clean, hackable base for your own voice companion and don’t mind wiring up API keys. Skip it if you expected a maintained product or a fully local stack — the interesting parts beyond this snapshot live behind a paywall.