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¶

Ten differences. Each one names the tool it is a difference from, and each ends with something you can check yourself instead of taking on trust — a file in the public repository, or a command you can run. Nothing here is a claim about the future.

1. Nothing you say leaves the machine — and the exceptions are enumerated, not asserted. Wispr Flow and Willow Voice send your audio to a server; that is their design, not a flaw. Handy, TalkType and nerd-dictation are genuinely offline — but "offline" is a claim you have to believe. YazSes publishes the list instead: five distinct mechanisms by which bytes could leave (importing a socket, spawning a program, handing a URL to your browser, asking a library to load a model by name, and a Unix socket that cannot reach a network at all), and exactly two code paths that can carry what you dictated. Both are constrained — the optional LLM cleanup is pinned to loopback, and yazses remote reaches only the host you typed on the command line. Check it: ADR-019 and tests/test_egress_inventory.py, which fails the build if any module under src/yazses/ grows a new outbound primitive without being registered.

2. A dependency list that says which parts you actually receive. Every project on this page has a large optional dependency tree; a vulnerability scanner reading one cannot tell what a normal install contains. YazSes ships a CycloneDX SBOM in which every component declares its scope — required for a plain install, optional for a feature extra you may never enable, excluded for build and test tooling that is never installed alongside the app. The scopes are derived from uv.lock by walking the real resolution graph, not hand-listed. Check it: sbom.cdx.json in the repository root, and the counts published in the privacy statement.

3. Three guards stand between the transcript and your keyboard. A dictation tool normally types whatever the model returned. YazSes stops in three cases: a dictated destructive command (rm -rf, mkfs, dd of=, a force-push) is held until you say a confirm word; a dictated number that fails its own check digit (Luhn, ISBN, Verhoeff) is held the same way, through the same gate, so you learn one release phrase rather than one per guard; and text with no editable target in focus goes to your clipboard with a notification instead of into the wrong window. Each is judged on how rarely it fires — a guard that interrupts a house number teaches you to dismiss it. Check it: src/yazses/cmdsafety/, src/yazses/checkdigit/, src/yazses/inject/target.py.

4. A broken config file cannot stop it starting. One bad line in a TOML file is enough to make most daemons refuse to run, which is the worst possible moment to lose your dictation. Config loading here is total: a repairable value is repaired ("0.004" → 0.004), anything else falls back to the documented default, unknown keys are dropped, and unparseable TOML still yields a working daemon. Every decision becomes a problem report that the daemon lists at start-up and yazses doctor shows. Check it: uv run yazses doctor and its Config validity row; the logic is in src/yazses/configcheck.py.

5. It can tell you which link in the chain is broken. yazses doctor proves the prerequisites (a mic exists, an injector is installed, the model is downloaded). yazses verify proves the function: it records, gates, transcribes and optionally injects for real, and names the first thing that broke with its fix rather than cascading. yazses report bundles versions, daemon state, a redacted config and a metadata-only log tail — locally; nothing is uploaded. Check it: uv run yazses verify and uv run yazses report on your own machine.

6. One install, three jobs. The same package and the same downloaded model do live dictation, offline transcription of recordings you already have (yazses transcribe), and whole-meeting capture with speaker labels (yazses meeting). Handy and TalkType are dictation tools, as their own sections below say; Otter-style meeting tools do meetings and are cloud services. Check it: uv run yazses transcribe --help and uv run yazses meeting --help.

7. Built for the people a hotkey excludes. Against every tool in the table: VAD calibration you run rather than guess at (yazses mic-level), a disfluency-friendly filter for stuttered and dysarthric speech, an on-device learning loop that adapts to your corrections, and a USB-serial or Bluetooth EMG muscle sensor as an alternative to pressing a key at all. Check it: uv run yazses mic-level, src/yazses/stt/filters/disfluency.py, src/yazses/platform/emg/.

8. Dictation and voice commands from one grammar. Handy has no voice commands; Voice In and Google voice typing have none; Windows, Apple and Wispr Flow dictation have a limited set (see the table above). Here, the same burst can type prose or fire "undo that", "save file", "go to line 42" as real key sequences, with per-editor grammar profiles and a personal dictionary for words Whisper keeps mis-hearing. Check it: src/yazses/commands/grammar.py, and uv run yazses vocab add <word>.

9. Self-improving, on your terms and on your disk. The learning corpus is off by default, encrypted with a machine-bound key, capped by size and retention, and never leaves the machine. Turned on, yazses tune re-transcribes your own audio with a larger model offline and proposes config changes — a vocabulary entry, a silence threshold — which you approve before anything is written. Check it: uv run yazses corpus status, uv run yazses tune (proposals only without --apply).

10. It builds and ships like infrastructure, not like a script. Cross-platform CI on Linux, macOS and Windows, on Python 3.11 through 3.14 and on arm64; a .deb, a .dmg, a Windows installer, a snap and PyPI from one tag; and build-provenance attestations on the release artifacts, so you can verify that the file you downloaded came from this repository's workflow. Check it: gh attestation verify <file> --repo MSKazemi/yazses on any release artifact, and the workflows in .github/workflows/.

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.

What the subscription actually costs over time¶

Dictation is a tool you use every working day for years, which makes a monthly fee an unusual shape of cost: you keep paying for speech recognition that already worked the first month. The comparison is worth making concretely.

Published list prices, checked 2026-08-11 (Willow and Voice In 2026-08-15; verify before quoting — these change):

For a team of five on Otter Business, the five-year figure is $6,000.

Three caveats, because a cost table that only argues one way is not an honest one:

• Free is not the same as costless. YazSes costs you setup time, about 1.25 GB of disk, and some CPU while it transcribes. If your time is worth more than the subscription, buy the subscription — that is a legitimate answer.
• The paid tools do things YazSes does not, and the sections below say where. Cloud models are larger than anything that fits comfortably on a laptop CPU.
• The prices above are list prices and both vendors run discounts.

Speaking is faster than typing — with an honest asterisk¶

Ordinary speech runs at roughly 120–150 words per minute, while most people type at 35–60 wpm. That is the reason dictation is attractive at all.

But raw throughput is the wrong number to plan around, and any tool that quotes it without qualification is overselling. What matters is throughput after corrections — and correcting a recognition error costs far more than the words it replaces, because you have to notice it, move the caret, and fix it.

So the useful framing is not "dictate everything, three times faster". It is:

• Dictation wins clearly for prose you would otherwise compose slowly — long-form writing, email, notes, documentation, chat.
• It wins overwhelmingly when typing is painful or impossible — RSI, injury, limited hand mobility. Here it is not a speed comparison at all.
• It wins least for dense code and identifiers, where accuracy per token matters most and correction cost is highest.

The measured accuracy behind those judgements is on the benchmarks page rather than asserted here. YazSes has not run a controlled throughput study against typing; that experiment is an open research question, deliberately listed as unanswered rather than guessed at.

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.

→ The detailed comparison, with a step-by-step migration guide

YazSes vs Vocalinux¶

Vocalinux is AGPL-3.0, 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 Willow Voice¶

Choose Willow if you want a polished commercial product on Mac, Windows or iPhone, with cloud-grade accuracy on technical vocabulary and a support contract behind it. It is a genuinely fast, well-made tool and it does things a laptop CPU cannot.

Choose YazSes if you are on Linux, or if the audio genuinely must not leave the machine.

One point is worth stating precisely, because it is easy to misread from comparison tables (including Willow's own, which lists Willow's offline support as an "optional mode"). Willow's Private Mode is a retention control, not local processing. Willow's privacy policy says plainly that "Willow uses cloud infrastructure to provide fast and accurate voice dictation, transcription, note-taking, and related features", and describes Private Mode as processing audio "transiently to return a transcription" without retaining it or training on it. That is a real and meaningful privacy commitment — no retention, no training, plus SOC 2 and HIPAA compliance — and for most users it is enough.

It is simply a different guarantee from the one this project makes. "We do not keep your audio" requires you to trust an operator, a policy, and a network path. "Your audio never left the machine" is a property of where the software runs, and you can verify it with tcpdump. Neither is automatically the right answer: if your threat model is a vendor mishandling data, Willow's compliance posture may be worth more than a self-hosted tool with no auditor. If your threat model includes the network itself, or you work under a rule that forbids third-party processing at all, no retention policy substitutes for the audio not being sent.

Checked against Willow's published privacy policy on 2026-08-15; verify before relying on it, since vendor policies change.

YazSes vs browser-extension dictation (Voice In, Google Docs voice typing)¶

Use a browser extension if everything you dictate lives in a browser tab. Voice In works across Chrome and Edge on thousands of sites in 50+ languages, installs in seconds, and has a free tier — for dictating into Gmail, Notion or a web CRM it is the lowest-friction option on this page. Google Docs' built-in voice typing is similar, and narrower still.

Choose YazSes when the text has to go somewhere a tab cannot reach. The limitation is structural, not a quality gap: a browser extension can only type into pages the browser renders. It cannot dictate into your terminal, your editor, an SSH session, a native desktop app, or an IDE. YazSes injects at the OS level, so the target window is not special to it.

Two things not to assume about browser dictation: it is usually not offline (the browser's speech engine typically streams audio to the platform vendor even when the extension itself stores nothing), and it does not do voice commands.

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.

YazSes vs Handy¶

Handy is the closest thing to a direct comparison: MIT-licensed, free, genuinely cross-platform (Linux, macOS, Windows), on-device, and built around the same hold-a-key-and-speak loop. It is popular for good reason.

Choose Handy if you want dictation and nothing else, with the smallest install and the least to configure. It does one job and does not ask you to care about anything else.

Choose YazSes if you also want the things that are not dictation: a documented voice-command grammar, whole-meeting capture with speaker labels, offline transcription of existing recordings, and a configuration surface deep enough to tune the silence gate, the disfluency filter and per-application tone.

The four adjectives people reach for — open source, local, cross-platform, free — describe both. They are not a differentiator, and this page would be dishonest if it implied otherwise.

YazSes vs OpenWhispr¶

OpenWhispr is MIT-licensed, runs on macOS, Windows and Linux, and transcribes locally with Whisper, Parakeet or Nemotron models. It also accepts spoken instructions such as "clean this up", and can optionally use a cloud model with your own API key.

Choose OpenWhispr if you want a polished desktop app and regard an optional cloud fallback as a feature rather than a dealbreaker.

Choose YazSes if "offline" needs to mean there is no cloud path at all — YazSes ships no remote inference of any kind, and the one place a local LLM endpoint can be configured refuses a non-loopback address unless you explicitly opt out. Also if you want speaker-labelled meeting transcripts, or published benchmark numbers with a method attached.

YazSes vs FluidVoice¶

FluidVoice is GPLv3, macOS-first, built on Parakeet, and has both a Write Mode and a Command Mode that can drive the OS itself — launching apps and running Shortcuts.

Choose FluidVoice if you are on macOS and want voice control of the machine, which is broader than what YazSes's editor-oriented command grammar does.

Choose YazSes if you are on Linux (FluidVoice announces it but does not ship it) or want the same experience across all three desktops.

Others in this space¶

Whispering, VoiceInk (macOS, paid) 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.