Speech to Text on Linux: 7 Tools Compared in 2026

What Linux Includes by Default

macOS and Windows expose system dictation in their default desktop settings. A typical GNOME or KDE installation does not provide the same system-wide voice-typing switch, though distributions and accessibility configurations differ. In practice, most Linux users install a dictation app, a script, or a transcription engine.

The missing built-in matters less than it used to. Whisper-class and Parakeet-class models run fast on ordinary Linux hardware, and the tooling around them grew from research scripts into real applications.

Recognition and text insertion are separate jobs. Recognition may succeed while the desktop blocks or misroutes the paste step. Compatibility therefore depends on both the app and the Wayland or X11 session.

7 Tools Compared

How to Choose a Linux Speech-to-Text Tool

Start with the destination for the text. Dictating into a browser, terminal, editor, and chat app needs reliable cross-application insertion. Transcribing a saved interview needs file support, timestamps, and export formats. Building speech recognition into another program needs an API or library. A tool can be excellent at one of those jobs and awkward at the others.

For voice typing across ordinary desktop apps, compare Spokenly, Handy, and nerd-dictation first. Spokenly favors a packaged setup, Handy favors an open-source desktop app, and nerd-dictation favors a small script that you can reshape. Speech Note is a better match when drafting inside one app is enough and manual copy-and-paste is acceptable.

For saved audio, start with whisper.cpp or a desktop file workflow. For an embedded project, Vosk exposes language bindings without a desktop interface. For full hands-free navigation and coding commands, Talon solves a broader control problem than dictation. The speech recognition software guide explains the same categories across operating systems.

System-Wide Dictation on Linux

Spokenly for Linux

Spokenly for Linux ships for x86_64 systems as amd64 deb, x86_64 rpm, and AppImage packages; ARM64 packages are not currently listed. A global hotkey starts system-wide dictation, and the app supports both X11 and Wayland text insertion. Local Parakeet and Whisper models work offline, while cloud providers are optional. It also transcribes audio files and exports TXT, SRT, VTT, and Markdown. The product page carries current packages and requirements.

Handy

Handy is a free, open-source desktop app that performs transcription locally and types into the focused field. Its Linux notes document extra setup: X11 typically uses xdotool, while Wayland needs wtype or dotool and may require a desktop-level shortcut. It is attractive when open source and offline processing matter more than a zero-configuration install.

nerd-dictation

nerd-dictation is a Python script around Vosk that you bind to a key and configure in code. Custom text processing is ordinary Python, which makes the workflow easy to adapt. In exchange, you select a Vosk model and configure the text-insertion path for your desktop instead of getting a packaged application.

Speech Note

Speech Note on Flathub bundles open speech recognition and text-to-speech models behind a desktop interface. Recognition runs offline. Text stays in Speech Note's own window, so cross-application Wayland insertion is not part of the dictation step; you copy the result where it needs to go.

Model choice changes the day-to-day feel of every local app. Smaller or quantized models download faster and use less memory, which helps older laptops. Larger models usually need more RAM and take longer to process, but may handle noisy speech, accents, and uncommon words better. Test a short sample from your real microphone before downloading several gigabytes of models.

Transcribing Audio Files on Linux

For recordings (meetings, interviews, voice notes), the terminal route is excellent on Linux:

1. 1whisper.cpp: a C and C++ implementation of Whisper with CPU and GPU backends. It is practical for batch jobs, scripts, and machines where a Python environment is undesirable.
2. 2OpenAI Whisper: the reference Python implementation. It is useful when you want the original model interface and the surrounding Python ecosystem.
3. 3Spokenly's Transcribe File: the same job without a terminal, with export to TXT, SRT, VTT, and Markdown for subtitles and notes.

The command-line setup, model download, and size trade-offs are covered step by step in running Whisper locally and the Whisper model size guide; both apply to Linux directly. whisper.cpp lives on GitHub.

Plan the output before a long transcription

Plain text is enough for search and notes. Subtitles need timed SRT or VTT output. Interviews may need speaker labels, but a speech model does not automatically provide reliable diarization just because it produces timestamps. Check that capability separately before processing hours of audio.

Long files also expose hardware limits. Run a five-minute sample, note the processing time and memory use, and verify names and language detection. Then choose the model and acceleration backend. A GPU can help, but CPU-only transcription remains useful when speed is not urgent. Keep the source recording until you have reviewed the transcript and subtitle timing.

Batch work benefits from stable file naming and explicit language settings. Store output beside the source or in a separate directory, and avoid overwriting the only copy. If a recording contains long silence, music, or repeated noise, voice activity detection or chunking can reduce wasted work and false text.

For Developers and Scripting

Two tools matter when speech to text is a component rather than an app:

See the Vosk project for language bindings and models. For dictating into AI coding agents, see voice dictation for developers; for hands-free control, compare Spokenly and Talon.

Linux Setup and Troubleshooting Checklist

Test the audio path and the text-insertion path separately. This prevents a microphone issue from looking like a Wayland issue, or a blocked paste from looking like a poor speech model.

1. 1Confirm the microphone first. Record ten seconds in the desktop sound recorder, play it back, and select the correct input device before changing speech-model settings.
2. 2Check the session with echo $XDG_SESSION_TYPE. Note the desktop and compositor too, because GNOME, KDE, Sway, Hyprland, and other environments expose different shortcut and input paths.
3. 3Download one small local model and test a short sentence. Confirm that recognition completes before adding a global hotkey, post-processing, or a larger model.
4. 4Test output in a plain text editor. If text appears there but not in a browser, terminal, or password field, the target app or field may handle synthetic input or paste differently.
5. 5On Wayland, follow the tool's documented helper setup and create a desktop-level shortcut when required. Check user groups and service permissions before running helpers as root.
6. 6For a packaged app, verify the architecture and package format. For source builds, record model paths, library versions, and launch commands so an update can be rolled back without rebuilding the whole setup from memory.

The Wayland vs X11 Catch

On Linux, recognition is the easy part. Typing the result into whatever window has focus is the hard part, because the two display protocols handle synthetic keystrokes in opposite ways:

If recognition succeeds but no text appears in the target app, session compatibility is a common cause. Check the session type with echo $XDG_SESSION_TYPE, then follow the app's documented setup for that desktop. Switching to X11 is a useful diagnostic step. Apps that keep text in their own window, such as Speech Note, avoid cross-application insertion.

FAQ

Read next

• Spokenly for Linux (deb, rpm, AppImage)
• How to Run Whisper Locally
• Whisper Model Sizes Explained
• Offline Speech-to-Text Software
• Voice Dictation for Developers
• Spokenly vs Talon Voice