artifacts

Local speech-to-text (provider stt_enhance) with faster-whisper or opt-in sherpa-onnx Parakeet V3 and long-audio chunking, plus an optional bounded LLM pass that enhances the transcript using a custom skill, an inline prompt or the default, and keeps the raw transcript on failure. Disclosure — audio is decoded locally. Post-processing is on by default and sends the transcript text to the LLM Hermes routes the stt_enhance auxiliary task to (your main, usually remote, provider unless you point it at a local endpoint or set post_processing.enabled: false). It reads the body of the SKILL.md you configure (never executed) and runs ffmpeg/ffprobe. The opt-in Parakeet backend reads a local model export and, on first use and only when your config.yaml explicitly sets security.allow_lazy_installs: true, installs pinned sherpa-onnx and numpy from PyPI into ~/.hermes/plugin-runtimes/hermes-stt-enhance/ (shared by all profiles). No tools, hooks, background processes, telemetry or stored credentials.