My Journey to a reliable and enjoyable locally hosted voice assistant
9 hours ago
- #HomeAssistant
- #LLMIntegration
- #LocalVoiceAssistant
- Switched from Google Home to local HomeAssistant with assist, backed by local first + llama.cpp.
- Optimized performance by fixing unwanted behaviors, improving unclear requests, and reducing token usage.
- Tested various GPUs (RTX 3090, RX 7900XTX, etc.) for local assist, noting response times and model compatibility.
- Used models like GGML GPT-OSS:20B and Unsloth Qwen3.5-35B-A3B for tool calling and context understanding.
- Implemented speech-to-text (Wyoming ONNX ASR) and text-to-speech (Kokoro TTS, Piper) solutions.
- Integrated LLM Conversation and LLM Intents for improved Assist functionality (weather, search, music).
- Customized LLM prompt to refine responses, remove emojis, and ensure concise outputs.
- Created automations for music playback, mapping satellites to media players for seamless control.
- Trained a custom wakeword ('Hey Robot') for better user experience and reduced false positives.
- Achieved a locally run, privacy-focused voice assistant with reliable core task handling.