Dictation that never leaves your Mac

What leaves your Mac

Speech recognition uses the SenseVoice model, which runs on your Mac after a one-time download of about 240 MB. Recent results are kept in memory only and cleared when you quit. Settings, including any API key, are stored in ~/.tidytalk.json, readable only by you.

The only other network request is the update check: at launch, at most once a day, Tidytalk reads the latest version number from GitHub. You can turn it off in Settings › General.

Set up Ollama

1. Install Ollama.
2. Download a model in Terminal, for example:

   ollama pull qwen2.5:7b

3. In Tidytalk, open Settings › Set Up AI, click "Presets", and choose "Ollama (local)". Check that the model name matches the one you pulled, then click "Save".
4. Click "Test Connection". When it says "Connected", you're done.

Set up LM Studio

1. Install LM Studio, download a model, and load it.
2. Start LM Studio's local server.
3. In Tidytalk, choose "LM Studio (local)" under Presets, enter the name of the loaded model, save, and click "Test Connection".

Local models don't need an API key, and Tidytalk gives them a 60-second timeout instead of 12 seconds.

Choosing a model

• Tidying is a light task, but the model has to follow instructions: tidy what you said instead of answering it. Small models are less reliable at this and at following custom rules.
• A model around 7B parameters or larger is a good starting point. qwen2.5:7b is about a 5 GB download and handles both Chinese and English well.
• If the local model is slow or returns something unexpected, Tidytalk pastes the raw transcript, so you never lose what you said.

No AI at all

Turn off "Tidy up text" in Settings › Set Up AI and Tidytalk becomes a fast, fully offline speech-to-text tool. Replacements, snippets, and Chinese–English spacing still work, because they run on your Mac without AI.

Verify it yourself

Tidytalk is open source under the MIT License. You can read exactly what it sends and when on GitHub, or build it from source.