Skip to content
NewPhonon-2 is coming

Talk. It types.
Nothing leaves your device.

Talkies is open-source dictation for macOS, Windows, Linux and Android. Hold a hotkey, speak, and your words land wherever the cursor is. Recognition, cleanup and typing run locally once you have downloaded a model.

Download TalkiesView source on GitHub (opens in a new tab)
  • macOS
  • Windows
  • Linux
  • Android
Listening0:04
On this device

Send Maya the draft by Friday and flag the budget numbers.

Illustration of the Talkies dictation panel listening and typing a sentence, with a badge showing processing happens on this device.

How it works

Your voice never takes a round trip.

Cloud dictation sends your audio to someone else's computer. There is no Talkies server to send it to: the whole job happens on your machine.

  1. Download a model once

    You choose the model. Talkies fetches it over HTTPS and checks it against a pinned SHA-256 hash before using it.

  2. Hold your hotkey and talk

    Audio is captured on your device and handed straight to the local engine. Nothing is streamed anywhere.

  3. Recognised on your hardware

    A local speech model turns audio into text. Optional S1-mini cleanup tidies punctuation and phrasing, also locally.

  4. Text lands at your cursor

    Talkies types the result into the app you were using. Switch off Wi-Fi and it keeps working.

Data flow

Model host · one-time download

Your device
  1. Microphone
  2. Speech engine
  3. Cleanup
  4. Your cursor
The network is used for the model download you start. After that, recording, recognition, cleanup and typing stay inside the dashed line.

Left out on purpose

  • No account or sign-in
  • No audio or transcript uploads
  • No analytics or tracking
  • No hosted speech service

No subscription, no billing, no trial timer. Your operating system's own sync or backup settings still apply.

Speech engines

Two engines. One rule: it runs here.

Phonon-2 will handle English. Whisper covers every other language. Both run entirely on your hardware after a one-time download.

Coming soonNew default for English

Phonon-2

Fermion Research's compressed build of NVIDIA Parakeet TDT. A compact English speech model that will become the default engine for English dictation.

Size · Apple Silicon
~164 MB
Language
English
Runs
On-device

Speed and accuracy figures for Phonon-2 are Fermion Research's own reported numbers, not Talkies measurements. We will publish ours when the engine ships.

Available todayFallback for other languages

Whisper

The open multilingual speech model, run through a native runtime on each platform. It powers every Talkies build today and stays on for languages beyond English.

Languages
Multilingual
Desktop
WhisperKit · Whisper.net · whisper.cpp
Android
Whisper tiny

Optional cleanup with S1-mini

A small local language model that tidies punctuation, filler words and tone. Off until you download it.

~462 MiB · CPU · Apache-2.0

Platforms

Native everywhere you type.

Each client is built with its platform's own toolkit, not a web wrapper. Pick yours.

Current builds are prereleases from GitHub Releases. Windows packages are unsigned for now. Verify any download against the SHA256SUMS file in the same release.

Open source

MIT licensed.
Read every line.

You shouldn't have to trust a privacy promise. Audit how audio is handled, build it yourself, or fork it. Model weights and third-party libraries keep their own licenses.

git clone https://github.com/4cecoder/talkies
Read the MIT License (opens GitHub in a new tab)

FAQ

Questions, answered plainly.

Which platform should I use?
Use the one you type on. macOS, Windows and Linux have native builds on GitHub Releases; Linux is newer and still maturing. Android is in early access with a signed APK on the download page. See the roadmap for current status.
Is it really free?
Yes. Talkies is MIT-licensed open source with no subscription, paywall, in-app purchase or account.
Does my audio leave my device?
No. Recognition runs locally with WhisperKit on macOS, Whisper.net on Windows and whisper.cpp on Linux and Android. Optional cleanup runs locally too. Audio and transcripts are not uploaded to any server.
Does it work offline?
Yes, after the one-time model download. The first-run guide walks through model setup, microphone permissions and text insertion.
What is Phonon-2, and when does it arrive?
Phonon-2 is Fermion Research’s compressed build of NVIDIA Parakeet TDT, about 164 MB on Apple Silicon, for English. It is in progress as the new default English engine. Whisper remains available for other languages. There is no release date yet.
How do I build it from source?
Each platform has its own toolchain. Start with AGENTS.md for commands, then the platform guide for mac/, windows/, linux/ or mobile/android/.
Where do I report bugs?
Open an issue on GitHub. Please leave recordings and private transcript text out of public reports.

Your words, on your machine.

Free, open source and private by construction. Download a build or read every line first.

Download TalkiesAll platforms
4cecoder/talkies on GitHub (opens in a new tab)