Speed of Sound Brings Offline Voice Typing to Linux

Speed of Sound, an open-source Linux desktop application, delivers fast, offline voice typing by running on-device automatic speech recognition and typing the result into any focused app. The project, authored by Antonio Zugaldia and distributed via Flathub and other packages, ships with a built-in `Whisper` model and now supports additional model families such as `Parakeet` and `Canary` to balance latency and accuracy. The app uses XDG Desktop Portals to simulate typing across GNOME, KDE, X11, and Wayland, and it includes optional LLM-based text polishing and self-hosted inference backends for more advanced workflows.

Speed of Sound performs transcription locally using on-device ASR with an ONNX runtime stack and JVM bindings. The app defaults to a lightweight `Whisper` model for broad multilingual coverage and low resource usage, while offering downloads for larger models to improve accuracy. Supported ASR families and integration points include:

• •`Whisper`, `Parakeet`, and `Canary` model families for offline transcription
• •Self-hosted LLM/text-polishing backends: vLLM, Ollama, llama.cpp
• •Cloud provider options when local compute is constrained
• •Desktop integration via XDG Desktop Portals for typed output across X11 and Wayland

Activation and UX are intentionally simple: users start recording with an in-app button, a global keyboard shortcut (the developer documentation cites Super+Z as an example), or system-tray controls, then stop recording to commit the transcription into the active text field. The app supports a primary and a secondary language and allows supplying contextual hints such as custom vocabulary and writing style to improve recognition. Recent Flathub release notes for version 0.12.0 added `NVIDIA Canary` and `Parakeet` support, fixed non-Latin script typing error propagation, and improved portal support detection.

Desktop voice typing has lagged behind mobile because of latency, privacy, and integration limits. The 2022 release of Whisper catalyzed a wave of on-device ASR projects; Speed of Sound packs that momentum into a polished Linux-native experience. For practitioners this matters because it makes private, offline speech-to-text accessible without cloud dependencies, and because it embraces self-hosted model stacks that teams can integrate into reproducible workflows. The combination of packaging on Flathub, Snap, and traditional Linux packages plus ONNX-based models means the app runs on both x86_64 and aarch64 hardware, with clear upgrade paths to GPU-accelerated or server-backed inference for heavier workloads.

The app is designed for discrete recordings rather than continuous streaming dictation; users must trigger and stop recordings which reduces complexity but limits hands-free workflows. Accuracy scales with model size and compute: the bundled multilingual Whisper is a sensible default, but professional transcription or noisy environments will require larger models or cloud/self-hosted inference. Optional LLM polishing introduces privacy and compute considerations when using external cloud LLMs; Speed of Sound explicitly supports self-hosted LLMs to retain local control.

Expect incremental improvements: streaming/real-time wake-word capture, tighter GPU acceleration on aarch64, more robust personal vocabulary adaptation, and enterprise packaging. Also watch for third-party extensions that automate post-processing pipelines and integrations with note-taking or accessibility tools.

Speed of Sound is a pragmatic, privacy-first voice typing tool for Linux that leverages modern on-device ASR ecosystems. It lowers the barrier for practitioners and power users to add offline voice input into everyday desktop workflows while offering clear upgrade paths to higher-accuracy or cloud-backed setups.