Beta

Voice dictation for Windows

Speak, and the words arrive in the application that has focus. mumblemuch turns what you said into clean text and pastes it where you were already working.

You do not dictate into a window of its own and copy the text out.

Press, speak, press again

The recording is a toggle. One press starts it, one press stops it. You are not holding a key down while you think.

The press does not have to be a key. Dictation ships on Ctrl + Win + Space, labelled "Type what I say", and it is bound to mouse button 4 and to the right bumper plus A on a gamepad out of the box as well. Each built-in mode has its own shortcut, and you can add your own.

The words land where you are working

When you stop the recording, mumblemuch transcribes what you said and pastes the result into the application that had focus. Mail client, editor, chat box, form field: the text arrives at the cursor.

The cleanup prompt is chosen by the application that has focus

Between the transcript and the paste there is a cleanup step. A prompt runs over the transcript, and which prompt runs is decided by the application in focus. A note to a colleague and a commit message do not want the same shape, so they do not get the same prompt.

Punctuation is settled at that step, read over the whole utterance, rather than taken from the pauses you happened to leave. Ramble, pause, restart a sentence: the cleanup prompt reads the whole transcript.

The step is one prompt over your words, so what it tidies is what that prompt asks for. The prompt is configuration. You can change the one an application uses in the workshop. mumblemuch can also draft a change to a prompt itself, and a drafted change waits for your approval instead of applying itself.

Names, jargon and the words only you use

A lexicon does two jobs. It maps a variant to the term you want written, so the name of your product comes out spelled the way you spell it. And it expands a spoken trigger into longer output, so a short phrase becomes the line you had been retyping.

Every match is gated by a six-level confidence rulebook, strongest first: exact, near exact, strong, partial, weak, disabled. The level a lexicon runs at decides how close a match has to be before it fires.

You do not have to build the list by hand. mumblemuch mines your own run history for repeatable patterns and surfaces each one as a suggestion. A lexicon entry becomes real when you approve it. Names and jargon are a list you keep, not a training session.

Where the speech and the cleanup run

Speech to text and the language model each run an ordered chain of providers, and mumblemuch fails over down the chain. The first entry can be the local template, your own endpoint, or mumblemuch cloud. That is the provider chain.

On the local template both stages run on your machine: the speech stage and the cleanup stage, on open speech and language models, on your GPU or CPU. You can keep the whole chain on your machine. mumblemuch runs locally by default, and nothing you dictate leaves this device unless a cloud provider is configured for that run.

Worth knowing if you are comparing this with Windows voice typing: Microsoft's page on voice typing documents it as online speech recognition that requires an internet connection.

• https://support.microsoft.com/en-us/accessibility/windows/use-voice-typing-to-talk-instead-of-type-on-your-pc - read 2026-09-21.

Language

English, with the built-in local model. That is the whole answer today, and the rest of the answers to the common questions are on the FAQ.

When you want an action instead of words

Dictation types. When you want a press to do something instead - translate the paragraph you just copied, explain the error message you copied - that is a different shortcut, and you can speak an instruction instead. mumblemuch classifies the utterance and runs one thing. You can see the three modes side by side.

Four things people do with it

Examples, not a feature list.

A mail client. Put the cursor in the message body, press the shortcut, say what you want to tell them, press again. The cleaned sentences arrive in the body, punctuated.

An editor, straight after. Same shortcut, same voice, different application in focus, so a different cleanup prompt runs over the transcript before the text lands.

A word that keeps coming out wrong. A lexicon entry maps the variant to the spelling you use, and the rulebook decides whether a given match is close enough to fire.

A line you retype every day. Say a short trigger and the lexicon expands it. If mumblemuch notices you making the same correction over and over, it proposes the entry itself and waits for you to approve it.

mumblemuch is in beta

mumblemuch is in closed beta. Leave your email and we will write to you when there is a build to try.