A local voice dictation application like Voxtty stores only essential operational data on your Linux machine: primarily cached speech models, configuration files, and an optional local transcript history. Crucially, no audio data ever leaves the device. Only anonymized transcript text can leave, and only if the user explicitly opts into and configures an external AI cleanup service.
The distinction between local and cloud-based data handling is fundamental for privacy-conscious users. Cloud dictation services require sending your audio stream over the internet to a remote server for processing. This introduces multiple privacy vectors: your voice data resides on third-party infrastructure, subject to their retention policies, potential data breaches, or government subpoenas. An API outage can render the service unusable, and the opaque nature of cloud processing makes auditing impossible.
Storing data exclusively on-device mitigates these risks. It grants the user complete control over their sensitive voice data, ensuring that raw audio never traverses external networks. This local-first approach means you own and can audit every byte of data generated by the application. Your dictation capabilities remain robust even without an internet connection, removing reliance on external service availability.
Voxtty, running as a systemd user service, adheres strictly to a local-first data model. When you press Alt+D, speak, and release, the process is contained within your machine. The Voice Activity Detection (VAD) algorithm identifies speech segments, which are then transcribed by the faster-whisper engine using locally stored models. The resulting text is typed into the focused application via `ydotool`, a utility compatible with Wayland and X11.
Here's a breakdown of what Voxtty stores and where, typically following the XDG Base Directory Specification:
~/.cache/faster-whisper/. Depending on the model size selected (e.g., small, medium, large), this can range from several hundred megabytes to multiple gigabytes. These files are static and only updated if you choose a different model or if the faster-whisper library itself is updated.~/.config/voxtty/config.toml. This file is plain text and directly editable, providing transparent control over the application's behaviour.~/.local/share/voxtty/history.db. This database contains only the transcribed text, timestamps, and potentially the application it was dictated into—never the audio. Users can easily manage or clear this history through the Voxtty interface or by deleting the file directly.journalctl --user -u voxtty.service) or to a file such as ~/.local/state/voxtty/voxtty.log. These logs record application events, errors, and performance metrics, but do not contain sensitive audio or transcript data by default./tmp and are immediately discarded and securely deleted after transcription completes. Raw audio is never persistently stored on disk or transferred off-device.The only scenario where data leaves your machine is through Voxtty's optional AI cleanup feature. If enabled, the *transcript text* (after local transcription, never the audio) is sent to a third-party API, such as Claude. This requires you to explicitly provide your own API key. The cleaned text is then returned and typed into your application. This opt-in process is clearly delineated, and users retain full control over whether to engage external services for text refinement.
While on-device data storage offers significant privacy advantages, it comes with specific considerations. The Whisper models, particularly the larger ones, demand substantial disk space and RAM. A large-v3 model, for instance, can consume over 3GB of disk space and require several gigabytes of GPU VRAM or system RAM for efficient inference. Users with limited resources might need to opt for smaller, less accurate models, impacting transcription quality.
The privacy of the transcript history, if enabled, is contingent on the security of your local machine. If your system is compromised, the text stored in ~/.local/share/voxtty/history.db could be exposed. Users must manage this data like any other sensitive local file. Similarly, if you opt for AI cleanup, your API key is stored locally in config.toml. While Voxtty does not send this key off-device, its local security remains your responsibility. The external AI service will process the text you send, introducing a new privacy boundary for that specific text data.
Furthermore, while Voxtty ensures audio never leaves your device, the accuracy of local Whisper models can be influenced by factors like strong accents, background noise, or highly technical jargon. These are inherent limitations of the models themselves, which might perform differently compared to proprietary cloud services that often leverage larger, constantly evolving datasets and more powerful compute resources.
To understand Voxtty's local data footprint firsthand, install the application and try dictating a paragraph. Afterwards, inspect your ~/.cache/faster-whisper/ directory to observe the downloaded model files. Then, list the contents of ~/.config/voxtty/ and ~/.local/share/voxtty/ to see the configuration and any optional history files created. This direct inspection provides tangible evidence of what resides on your machine. Try Voxtty free by visiting https://voxtty.com.
Local-first voice dictation for Linux. Press Alt+D, speak, and your words land in whatever app has focus — nothing leaves your machine.
Try Voxtty free →