YazSes — Subsystem Architecture

Local, offline hold-to-talk voice dictation. Same diagram as yazses-architecture.mmd (Mermaid) and .txt (ASCII).

yazses CLI44 commands · 6 help panels

yazses-traymenu-bar UI (macOS / Windows)

JSON-RPC 2.0 · newline-delimited · Unix socket / Windows named pipe

▼

IPC serveripc/

▼

yazses-daemon — orchestrator (core/daemon.py)

1 Hotkey backendkeyboard hook OR EMG muscle-sensor (USB serial / BLE)

▼

2 Audiorecorder → VAD gate (calibrated) → pre-speech padding

▼

3 STTfaster-whisper (CPU int8) · streaming · disfluency filter · initial_prompt bias

▼

4 Post-processcleaner · voice-punctuation · continuation spacing · optional LLM cleanup

▼

5 CommandsTier-1 regex grammar → Tier-2 SLM router (low confidence) · LSP context

▼

6 · dispatch()

dictate

Injectorauto → ydotool / xdotool / wtype / clipboard

command

Key sequencectrl+z · ctrl+s · arrows · …

▼

Focused application (local)

— or — SSH reverse tunnel (remote/), text only, audio stays local

yazses-agentremote host

Configurationconfig.toml / config.py — every section has defaults; v2 features default off

Feature registrysystem/features.py — 135 capabilities across 9 categories

Platform abstractionplatform/base.py Protocols → linux · macos · windows · emg (Hotkey / Injector / Lifecycle / IPC / Permissions / Tray)

Opt-in subsystems (OFF by default)Learning loop (encrypted corpus · yazses tune) · v2 cognitive layer (voiceprint · gaze · personalize · polyglot · cocktail) · Recording import (yazses transcribe · diarization)

User-facing Orchestrator / core Output / injection Config / registry Opt-in (off by default)

Everything runs on-device. Nothing leaves the machine except the optional SSH remote path (final text only).