# Shuush > Open-source local Superwhisper & Wispr Flow alternative for Linux and Wayland, built cleanly on top of MIT-licensed voxtype (https://github.com/peteonrails/voxtype). ## Canonical URLs - Site: https://shuush.kitsunekode.in/ - 1-Line Installer: https://shuush.kitsunekode.in/install.sh - Pre-Flight Doctor: https://shuush.kitsunekode.in/doctor.sh - Launch Blog & Benchmarks: https://shuush.kitsunekode.in/blog/why-we-built-shuush/ - RSS Feed: https://shuush.kitsunekode.in/rss.xml - GitHub Repository: https://github.com/KitsuneKode/shuush ## Verified Facts & Benchmarks - License: MIT (100% open-source, unlimited words, $0) - Operating Systems: Linux Wayland (Hyprland, Omarchy, Sway, Niri, Arch Linux) + first-party kuwai rice support - GPU Backends (via stock voxtype-bin): NVIDIA CUDA, AMD ROCm/Vulkan, Intel Vulkan, CPU - Speech Engines (via stock voxtype-bin): Parakeet Unified EN 0.6B fp16, Whisper large-v3-turbo, Cohere Transcribe GGUF, Moonshine - Measured daily-driver telemetry (Intel i5-13450HX, RTX 4050 Laptop 6 GB VRAM, 23 GiB RAM): - 333 real takes, 19,498 words dictated (~8.1 hours saved vs 40 WPM typing) - 516 ms P50 / 1,060 ms P90 end-to-end cleanup latency (24 ms when already clean) - 98.8% cleanup reliability; ~2.3 GB resident CUDA VRAM - Architecture: 1. Local ASR via stock voxtype (zero fork required) 2. 5-source bounded context: App + AT-SPI / + highlighted + fresh <=15s + 3. Deterministic anti-paraphrase span gate (voxtype_pp_gate.py) + 1-click Air-Gapped Ollama mode or multi-provider failover 4. Phonetic grounded self-learning (voxtype_learn.py) synced into voxtype config.toml [text] replacements + native GTK4 Studio