SayIt
SayIt is an open-source voice typing app for Windows under AGPL-3.0. Press a shortcut, speak, and cleaned-up text is typed at the cursor. It runs in three modes: local, with seven GGUF speech models on your PC; cloud, with your own provider keys and no SayIt server involved; or server, against a backend you host in Docker. The cleanup step uses Ollama or any OpenAI-compatible endpoint, and every prompt is editable.
What does SayIt do?
SayIt's README opens with "Open-source voice typing for Windows." Press a shortcut, speak, and the cleaned-up text is typed where the cursor is, press-to-toggle by default with hold-to-talk available. It runs in three modes. Local mode does recognition on the PC with seven GGUF models the app ships: Parakeet Unified EN, SenseVoice Small, Fun-ASR Nano, Nemotron 3.5 ASR (32 languages) and three sizes of Qwen3-ASR, with GPU acceleration. Cloud API mode uses your own keys with Doubao, Qwen, Xiaomi MiMo or Groq Whisper, and the README says "No SayIt server is involved." Server mode talks to a backend you host or to a free public trial server.
The cleanup step can use DeepSeek, Qwen, Groq, MiMo, Ollama or any OpenAI-compatible endpoint, and the README says "Every prompt is yours to change." Context-aware writing, off by default, reads the text around the cursor, and selected text becomes an editing instruction; password fields are skipped. It has an English and Chinese interface, hotwords, per-app rules, a waveform overlay with live captions, and a local history with re-transcribe.
Key facts
- Licence: AGPL-3.0 on the code. The seven bundled GGUF models' licences are not stated in the README.
- Local models: Parakeet Unified EN, SenseVoice Small, Fun-ASR Nano, Nemotron 3.5 ASR (32 languages), and Qwen3-ASR in three sizes, with GPU acceleration.
- Three modes: local recognition on the PC, cloud API with your own keys, or a self-hosted server; a free public trial server exists.
- Cleanup providers: DeepSeek, Qwen, Groq, MiMo, Ollama, or any OpenAI-compatible endpoint; with cleanup off, the README says nothing leaves the device.
- Self-hosted server: Docker Compose with a FastAPI backend, WebSocket streaming and Qwen3-ASR; an NVIDIA GPU is required and 16 GB or more of VRAM is recommended for the default server model.
- Server performance reference in the README: Qwen3-ASR-1.7B with vLLM on an A10G 24 GB, 30 seconds of audio in about 0.8 seconds, 5 minutes in about 3.0 seconds.
- Client build needs Node.js 18+, Rust 1.75+, CMake 3.20+ and the Vulkan SDK; the first native build "may take around 20 minutes".
- Latest tagged release when read: v0.2.0 on 2026-09-09; 407 stars on 2026-09-17.
What does it replace, and where does it fall short?
SayIt replaces Wispr Flow and Superwhisper on Windows, and its own README says so. It goes further than Handy on the cleanup side, with an editable prompt and a choice of local or hosted models for it, and further on the model side, with Nemotron and Qwen3-ASR alongside Parakeet. It sits on the open-source alternatives to Wispr Flow page beside Handy, VoiceInk, OpenWhispr and the others.
Where it falls short: Windows only for the client. The self-hosted server needs an NVIDIA GPU with 16 GB or more of VRAM recommended. The community channels are Chinese-language, by the README's own note. Building the client from source compiles a C++ speech engine and takes about 20 minutes the first time.
How does SayIt run?
Install the Windows client from GitHub Releases. For local mode nothing else is needed; the models ship with the app and use the GPU where available. For server mode, run the backend with Docker Compose on a Linux machine with an NVIDIA GPU, or point the client at the public trial server. Cloud mode needs your own API keys for the providers you choose. There is no paid tier stated; the trial server and the web demo at sayitapp.site are free.
Who is SayIt for?
A Windows user who wants dictation with a strong cleanup step and control over every prompt, or a team that wants a shared recognition server on its own GPU. Mac users should look at VoiceInk or FluidVoice; Linux users at Handy, OpenWhispr or SpeakoFlow.
What limits does the README state?
From the README: the community channels are Chinese-language; GPU speech recognition on the server requires an NVIDIA GPU; on non-English Windows installations the build needs CL=/utf-8 set first.
Questions people ask
Is SayIt open source?
Yes. The client and the server are under AGPL-3.0, an OSI approved copyleft licence, and the README states the obligation that comes with it. The seven speech models bundled for local mode carry their own licences, which the README does not list.
Does SayIt keep my voice on my PC?
In local mode, yes: recognition runs on the PC with the bundled models, and the README says that with AI cleanup off, nothing leaves the device. Cloud mode sends audio to the provider you chose with your own key. Server mode sends it to the backend you host, or to the public trial server if you use that.
Does SayIt run on a Mac or Linux?
The client is Windows only. The self-hosted server runs in Docker on Linux with an NVIDIA GPU. For macOS, see VoiceInk or FluidVoice; for Linux, Handy, OpenWhispr or SpeakoFlow, all on the same alternatives page.
How does SayIt compare with Wispr Flow?
SayIt does the same hotkey-to-text job with cleanup, on Windows, with the recognition model and the cleanup model both under your control and no subscription. Wispr Flow adds macOS and phone apps and a polished default. The open-source alternatives to Wispr Flow page compares the options by need.
Which speech models does SayIt ship?
Seven GGUF models for local mode: Parakeet Unified EN, SenseVoice Small, Fun-ASR Nano, Nemotron 3.5 ASR with 32 languages, and Qwen3-ASR in three sizes. The self-hosted server runs Qwen3-ASR through vLLM, with the README quoting 30 seconds of audio transcribed in about 0.8 seconds on an A10G.
Sources
- SayIt README and AGPL-3.0 LICENSE: github.com/crosswk/SayIt, read 2026-09-17.
Compared with the others
On the open-source alternatives to Wispr Flow page, SayIt is the pick for windows, with an editable cleanup step. AGPL-3.0, seven bundled local models including Nemotron 3.5 and Qwen3-ASR, cleanup through Ollama or any OpenAI-compatible endpoint, and a self-hosted server option.
Also on that page: Handy for all three desktops, nothing else running, VoiceInk for a mac, with modes per application, FluidVoice for a mac with the newest local models.
On the open-source alternatives to Superwhisper page, SayIt is the pick for windows, with an editable cleanup step. AGPL-3.0, seven bundled local models including Nemotron 3.5 and Qwen3-ASR, cleanup through Ollama or any OpenAI-compatible endpoint, and a self-hosted server option.
Also on that page: VoiceInk for a mac, with modes per application, Handy for all three desktops, nothing else running, FluidVoice for a mac with the newest local models.
More voice dictation apps
A free, offline push-to-talk dictation app for macOS, Windows and Linux that runs Whisper or NVIDIA Parakeet on your own machine and pastes the text at your cursor, under MIT.
A GPL-3.0 macOS dictation app that runs Nemotron, Parakeet, Cohere Transcribe, Apple Speech or Whisper on the device, with command and write modes, per-app prompts, and an optional closed local enhancement model.
A dictation and meeting-notes app under MIT that runs Whisper or Parakeet locally or cloud models with your own key, detects Zoom, Teams and FaceTime calls, and labels speakers on device.
A macOS dictation app under GPL-3.0 whose README names Superwhisper and Wispr Flow as the tools it replaces, with local models, per-app modes, a personal dictionary and an AI assistant mode.
A speech-to-text app inside the Epicenter monorepo under AGPL-3.0 that records, transcribes with a provider you choose, cloud or local, optionally polishes the text, and pastes it at the cursor.
An MIT desktop voice assistant for Windows, macOS and Linux that extends Handy's dictation core with a trigger phrase, a local cleanup model, screen-aware answers and Kokoro speech, with no account and no telemetry.
Added September 17, 2026. Every claim above comes from the project's README, LICENSE or model card, read on September 17, 2026, or from the GitHub API on the date shown in the panel. Found an error? Write to reveneau@licheo.com and it is fixed in the next weekly pass. Repository: github.com/crosswk/SayIt.
Running one of these inside your own environment, with your own data and your own security rules, is the kind of work Reveneau does. Read how a forward deployed engagement works.