The speech models in TessFlo

TessFlo runs open speech models on your Mac. Pick one when you first set TessFlo up; each is a one-time download.

Jump to a section

Speech models

Scroll the table sideways to see every column.

In TessFloModelResultsDownload
Best QualityNVIDIA Parakeet TDT 0.6B v2Words appear as you speakAbout 550 MB
BalancedOpenAI Whisper large-v3-turboText after recordingAbout 1.7 GB
FastOpenAI Whisper smallText after recordingAbout 550 MB

Download sizes are approximate. All three models are for English dictation.

NVIDIA Parakeet v2

TessFlo uses the official FluidAudio package for on-device Parakeet v2 speech recognition, with the Core ML conversion by the FluidInference team. The model is licensed under CC BY 4.0. See the Parakeet v2 notice.

OpenAI Whisper

Balanced uses OpenAI Whisper large-v3-turbo and Fast uses Whisper small, with Core ML conversions by Argmax. The original code and weights are MIT-licensed. See the Whisper notices.

Optional Light Cleanup

Light Cleanup makes local grammar and punctuation edits and keeps your original transcript. It uses a quantized Gemma 2 2B instruction model that runs on your Mac. It’s a separate download with its own agreement; ordinary dictation doesn’t need it.

What stays on your Mac

Speech recognition and enabled Light Cleanup run locally on your Mac. TessFlo does not send audio or transcript content to its licensing service. An internet connection is used when you request model downloads, use licensing features, or check for updates.

Model downloads come from their hosts, which see normal connection details such as your IP address; a download may reveal which model you chose. See section 3, “Model downloads”, of the Privacy Policy.

Requirements

See how TessFlo compares with other apps: the best dictation apps for Mac.