YazSes vs. other dictation tools¶

Short answer: YazSes is the tool to pick when you want hold-to-talk voice dictation that runs fully offline on Linux (and macOS/Windows), is free and open-source, and also does voice commands — without sending your voice to any cloud service. If you need cloud-grade AI reformatting, professional medical/legal accuracy, or deep voice-coding scripting, one of the alternatives below may fit you better. This page is deliberately honest about where each tool wins.

At a glance¶

Looking for meeting notes rather than dictation? YazSes also records whole meetings and transcribes existing recordings offline. That is a different competitor set (Otter.ai, Fireflies, Granola, Meetily) and has its own page: Offline meeting notes.

What makes YazSes different¶

• Fully offline & private by default. Audio is transcribed on-device with CPU faster-whisper (int8). No GPU, no network, no account — nothing you say leaves the machine.
• Dictation and voice commands. Speak to type, or use a fast regex command grammar (with an optional ~0.5B SLM router for low-confidence phrases) that maps "undo that", "save file", "go to line 42" to real key sequences.
• Hold-to-talk. Natural push-to-talk that types into whatever app has focus — editor, browser, terminal, chat.
• Linux-first, cross-platform. Works on X11 and Wayland, plus macOS and Windows.
• Built for accessibility. VAD calibration, a dysfluency-friendly mode for stuttered/dysarthric speech, and an optional EMG muscle-sensor trigger for hands-free use.
• Self-improving on your terms. An opt-in, encrypted, on-device learning corpus lets yazses tune propose accuracy fixes from your own corrections.
• One tool for three jobs. The same install and the same downloaded model do live dictation, offline transcription of existing recordings (yazses transcribe), and whole-meeting capture with speaker labels (yazses meeting). Every other tool on this page does one of the three.

The difference that is not a feature¶

Everything in the list above is copyable. Offline Whisper, a command grammar, Wayland support, a tray icon — any of the open-source projects on this page could ship all of it within a couple of releases, and some of them will. If you are choosing a tool for this month, compare the features. If you are choosing a project to depend on, or to contribute to, the more useful question is what it is structurally able to become.

Every dictation tool listed here answers "how does the user start talking?" the same way: a hotkey, hardcoded. That single assumption is load-bearing — it decides who the tool can serve. If you cannot reliably press and hold a key, a push-to-talk dictation tool is not accessible to you, no matter how good its transcription is.

YazSes treats the activation channel as a replaceable part. A hotkey is one implementation of a protocol, not a built-in assumption, and other implementations already exist in the tree: a USB-serial EMG muscle sensor, a BLE variant, and gaze-based window targeting on X11.

What that buys is a straight answer to a question the other tools cannot answer at all — "I can't press a key. Can I still use this?"

It is also why the research on the muscle and brain control page is not decoration. The 2025–26 silent-speech literature is converging on a result that maps directly onto this architecture: closed vocabularies of 10–30 words are decoded at 96–97% accuracy, while open-vocabulary silent speech still costs ~68% word error rate. So the defensible split is silent commands plus spoken prose — which needs a system where a command channel and a dictation channel can be owned by different hardware. That is a shape, not a feature, and it is not something a hotkey-shaped tool can add later without becoming a different program.

Will it still be here next year?¶

A fair question for any small open-source project, and the honest signals rather than a promise:

• Apache-2.0, no CLA, no open-core tier withheld from the repo. If this project stalls, the whole thing is forkable by anyone, with no permission required.
• 2,300+ tests across Linux, macOS and Windows on every push, so a fork or a new maintainer inherits something they can actually change safely.
• Published on four independent channels (PyPI, Snap, an APT repo, and .dmg/.exe release artifacts), so it does not disappear if one of them does.
• Ten people have contributed. That is a small number, said plainly — but the architecture, the ADRs and the design decisions are all in the open precisely so it does not depend on one person's continued attention.

The thing most likely to kill a project like this is not competition; it is a single maintainer losing interest with everything undocumented. That is the failure being designed against.

When another tool is the better choice¶

YazSes vs Dragon NaturallySpeaking¶

Choose Dragon if you need best-in-class accuracy for professional medical/legal dictation on Windows and a commercial license is acceptable. Dragon is a mature, paid, Windows-focused product with specialist vocabularies YazSes does not ship.

Choose YazSes if you are on Linux or macOS (Dragon is Windows-only), if a per-seat commercial licence is a blocker, or if you want the source to be auditable. On accuracy for general prose the gap is much smaller than it used to be; on specialist terminology it is not.

YazSes vs Talon Voice¶

Choose Talon if your priority is deep, scriptable voice coding. Talon has a powerful scripting ecosystem — Python configs, a large community grammar library, eye tracking — and for people who drive their whole desktop by voice it remains the most capable option.

Choose YazSes if you want dictation that works out of the box without learning a scripting system, want it fully open-source (Apache-2.0), or want file transcription and meeting capture from the same install. The two coexist happily; they are aimed at different points on the effort/power curve.

YazSes vs nerd-dictation¶

nerd-dictation is a single Python file using the VOSK API, GPLv3, with famously small models and no background process — dictation is started and stopped with explicit begin/end commands, and you customise output by writing Python string operations.

Choose nerd-dictation if you want maximum minimalism and hackability, the lowest possible resource footprint, or you like configuring behaviour in code.

Choose YazSes if you want a hold-to-talk key instead of begin/end commands, Whisper-class accuracy rather than VOSK, macOS/Windows support (nerd-dictation is Linux-only), or the packaged extras — voice commands, macros, personal vocabulary, file transcription, meeting capture.

YazSes vs Vocalinux¶

Vocalinux is GPLv3, supports whisper.cpp / Whisper / VOSK, runs on X11 and Wayland, has voice commands for text manipulation, and — notably — offers Vulkan GPU acceleration across AMD, Intel and NVIDIA.

Choose Vocalinux if you have a capable GPU and want to use it, or you want to pick between three recognition engines.

Choose YazSes if you need macOS or Windows too (Vocalinux is Linux-only), if you want the same install to also transcribe recordings and capture meetings with speaker labels, or if you need the accessibility-oriented pieces — dysfluency-friendly mode, EMG triggering, VAD calibration — and the opt-in on-device learning loop.

YazSes deliberately targets CPU int8 rather than GPU, so it runs on modest hardware; if you have the GPU, a whisper.cpp-based tool will transcribe faster.

YazSes vs TalkType¶

TalkType is the closest thing to YazSes on Linux: it is offline, Whisper-based, Wayland-first, and uses the same hold-to-talk gesture — press a key to talk, release to type. It ships as a zero-config AppImage with optional GPU acceleration.

Choose TalkType if you want a single-file AppImage with nothing to configure, and Linux is the only machine you dictate on.

Choose YazSes if you also work on macOS or Windows, if you want the same install to transcribe existing recordings and capture meetings with speaker labels, or if you need the accessibility and voice-command layers.

YazSes vs Wispr Flow¶

Choose Wispr Flow if you want polished, cloud-based AI formatting and rewriting and do not need offline operation or Linux support.

Choose YazSes if the audio must not leave the machine, if you are on Linux, or if you do not want a subscription. This is the clearest trade-off on the page: cloud polish versus local privacy.

YazSes vs Google / Apple / Windows built-in dictation¶

Use the built-in if it is already good enough and you are comfortable with cloud processing (Google), a walled ecosystem (Apple), or Windows-only (Windows Speech Recognition). They cost nothing and need no setup.

Choose YazSes if you want the same dictation behaviour across all three operating systems, need it to work with no network, or want voice commands that the built-ins largely do not offer. Note that Linux has no comparable built-in at all — that gap is the reason this project exists.

YazSes vs Whisper + your own scripts¶

Roll your own if you enjoy building and maintaining the glue.

Choose YazSes if you would rather not: it is that glue, productized and tested — hotkey capture across multiple keyboards, VAD calibration, pre-speech padding, command grammar, text injection that works on X11 and Wayland and in terminals, a no-text-target guard, mic-change auto-healing, and packaging for APT/Snap/PyPI.

YazSes vs the dictation built into your editor or AI coding tool¶

VS Code, Cursor and several AI coding assistants ship their own voice input, and for typing a prompt into that tool's own box they are the path of least resistance. The limitation is structural rather than a matter of quality: in-application dictation only reaches what the application owns.

YazSes types at the operating-system level — ydotool on Wayland, xdotool on X11 — into whichever window currently has focus. Nothing about the target window is special to it, so it does not care whether the shell behind that window is local, SSH'd, containerised or on the other side of the planet.

The practical consequence, and the reason developers working on remote machines tend to notice it first:

If most of your work happens over Remote-SSH — editing on a server, driving a build box, working in a container — this is the single biggest day-to-day difference, and it needs no configuration: install YazSes and dictate.

Details and the forwarding case (text typed on a remote host's own display): dictation over SSH.

Others in this space¶

Whispering, OpenWhispr, Handy and VOXD are also active open-source offline dictation projects, mostly built on whisper.cpp.

Speech Note is worth calling out separately because it is a different shape of tool: a notepad application you dictate into, which also does text-to-speech and offline translation. If you want a document to write in rather than dictation injected into whatever window has focus, it is the better fit — and it does more than YazSes on translation.

These are worth a look if the tools above do not fit; this page is a comparison, not a claim that YazSes wins every case.

Common questions¶

Is there a good open-source, offline alternative to Dragon or Wispr Flow on Linux? Yes — YazSes is an open-source (Apache-2.0), fully offline dictation tool that runs on Linux, macOS, and Windows and needs no cloud account.

What's the best free voice dictation for Linux that also does commands? YazSes combines on-device transcription with a voice-command grammar, so the same hold-to-talk key both types text and triggers editor/terminal actions.

Does YazSes send my audio anywhere? No. Transcription runs locally with faster-whisper; by default nothing leaves your machine.

Does voice dictation work over SSH, or in a VS Code Remote-SSH session? Yes, with no extra setup. YazSes injects keystrokes at the operating-system level into the focused window, not inside a particular application, so Remote-SSH editor panes, integrated terminals running a remote shell, tmux sessions and container shells all receive dictated text normally. This is where it differs most from the voice input built into editors and AI coding tools, which is bound to that application's own input handling.

Is there an offline, open-source alternative to Otter.ai for meeting notes? Yes. yazses meeting start / yazses meeting stop records a meeting hands-free and produces a speaker-labelled transcript on-device, with optional minutes from a local LLM — no account, no per-seat fee, and no bot joining the call. It records the room through your microphone rather than capturing a video call's system audio; see Offline meeting notes.

Can I transcribe an existing recording offline? Yes — yazses transcribe interview.m4a converts any audio or video file to txt, md, srt, vtt, or json, with --diarize for who-said-what speaker tags.

What is the best open-source dictation tool for Linux? It depends on what you weight. nerd-dictation is the most minimal, Vocalinux has GPU acceleration, Talon is the most powerful for voice control, and YazSes is the one that is cross-platform and covers dictation, file transcription and meeting capture from a single install. All four are free and run offline; the sections above lay out the trade-offs honestly.

Does YazSes work on Wayland? Yes. It probes the session at runtime and injects text through ydotool or wtype on Wayland and xdotool on X11. This is the part that most Linux dictation tools struggle with — see voice dictation on Linux.

Is there an alternative to Dragon NaturallySpeaking for Linux? Dragon does not run on Linux at all. The closest open-source options are YazSes, Talon Voice, Vocalinux and nerd-dictation. For general prose dictation the accuracy gap to Dragon is modest; for specialist medical or legal vocabularies it is not.

Honest limitations¶

• Accuracy is Whisper-class — 4.07 % WER on LibriSpeech test-clean with the default model, measured and reproducible. It is not tuned for specialized medical/legal vocabularies the way Dragon is, and real dictation in a room will be worse than a clean read-speech benchmark.
• On Wayland, global-hotkey and injection setup needs ydotool/ydotoold.
• The first run downloads the STT model.
• It is a dictation + command tool, not an LLM agent or a full voice-scripting platform like Talon.

Ready to try it? See Install on Linux — or pipx install yazses on any OS with Python ≥ 3.11.