AI Lab Notes
RSS FeedAI experimentation, coding guides, and system administration notes. Written from a Pop!_OS workstation with an RTX 3090.
Featured
-
Local AI Video Generation with LTX-2 on an RTX 3090
Running Lightricks' LTX-2 19B video generation model on consumer hardware. Covers weight streaming, FP8 vs BF16 checkpoints, distilled model constraints, prompting lessons, and a practical img2vid pipeline.
-
Hello World: Welcome to AI Lab Notes
Introducing AI Lab Notes -- a technical blog covering local LLM setups, AI coding agents, browser automation, and Linux system administration.
Recent Posts
-
Speaker Verification for a Wake Word Daemon: Keeping 'Hey Jarvis' Personal
Add speaker verification to a wake word daemon so only your voice triggers it. Uses SpeechBrain ECAPA-TDNN, a rolling audio buffer, and cosine similarity -- 270ms latency, 86 MB VRAM.
-
Claude Code Hooks: Making Voice Mode Persistent Across Turns
Use Claude Code's Stop hook to inject a system reminder after every response, so the agent remembers voice mode without burning tokens on verbose skill instructions.
-
Adding Voice to Your AI Coding Agent: Text-to-Speech with Kokoro
Give your AI coding agent a voice with Kokoro TTS. Deploy a local, GPU-accelerated text-to-speech server and wire it into Claude Code for spoken responses.
-
Voice Dictation on Linux Wayland: Getting Push-to-Talk Working on COSMIC Desktop
How to set up push-to-talk voice dictation on Wayland Linux desktops using Voxtype and ydotool, with solutions for wtype failures and hotkey passthrough issues.