I tune speech models to run offline on small hardware: a Raspberry Pi 400, an Android phone and a Pico.
- oido: offline speech-to-text with a tuned Pi 400 build. It also runs on the Pi's GPU, level with the CPU on a 10 s clip. How it got there.
- oido-streaming: the live streaming dictation code for oido on the Pi 400 GPU (chunked encoder on the GPU, VAD with a 5 s pre-roll).
- gpu-port-kit: tools, layout tests and measured results for porting models to the Pi 4/400 GPU (Vulkan, V3D), including a hand-written QPU matmul.
- pi-gpu-asr4all: asr4all-m speech-to-text on the Pi GPU with hand-written shaders.
- pi400-ai-notes: measured results for speech-to-text, text-to-speech and small LLMs on a Pi 400.
- pi-wake-word: train a tiny wake-word model on your own voice on a Pi.
- pi-code: a small dependency-free agent CLI for a local llama.cpp server.
- pico-experiments and universal_build_pico: RP2040 experiments, including llama2.c stories260K on the Pico.

