Self-host microWakeWord-Trainer-Nvidia-Docker

microWakeWord Trainer is a self-hosted Docker application for training custom wake word models (microWakeWord, openWakeWord, NanoWakeWord CNN/GRU) using NVIDIA GPU acceleration. It provides a reactive Vue 3 + TypeScript web UI for managing positive and negative audio samples, reviewing captured audio from Tater Native or ESPHome satellites, and running automated training loops with local speech-to-text (Faster Whisper or Parakeet ONNX). The trainer generates text-to-speech samples via multiple providers (OmniVoice, Qwen3-TTS, MOSS-TTS-Nano, Piper), normalizes all audio to a standard sample rate and bit depth, and publishes trained wake word bundles (JSON plus .tflite/.onnx models) that Tater Native satellites can load live without firmware changes. It runs via a single Docker container with --gpus all, exposes the UI on port 8789 (configurable via REC_PORT), and persists all data in a mounted /data volume. The project includes a separate Docker image for RTX 50-series (Blackwell) GPUs using a community-built TensorFlow wheel (CC BY-NC 4.0). No license file is present in the repository.