Models
The app downloads these when you ask for them. It shows them under its own names, chosen by what each does. Standard is the one we trained. We didn’t make the others.
Express Speech in English
NVIDIA Parakeet TDT 0.6B v2, by NVIDIA. Converted to 8-bit ONNX by the sherpa-onnx project.
Universal Speech in any language
Whisper Large v3 Turbo, by OpenAI. Converted for whisper.cpp and made smaller (5-bit).
Lite Speech in any language
Whisper Base, by OpenAI. Converted for whisper.cpp.
Standard Cleanup
Trained by us for dictation from Qwen2.5 1.5B Instruct, by the Qwen team at Alibaba Cloud.
Advanced Cleanup, Command Mode and meeting notes
Qwen3 4B Instruct 2507, by the Qwen team at Alibaba Cloud. Made smaller (4-bit) by Unsloth.
English and Hindi Speech in meetings held in English and Hindi
Whisper-Hindi2Hinglish-Apex, by Oriserve. Converted for whisper.cpp and made smaller (5-bit) by us.
Voices Telling the people on a call apart
ERes2Net trained on VoxCeleb, by the 3D-Speaker project at Alibaba. As published by sherpa-onnx.
Engines inside the app
They run the models on your Mac. Their licence texts ship inside the app, and Settings › About › Show Licences opens them, along with those of the libraries sherpa-onnx is built on.
- whisper.cpp, by Georgi Gerganov and contributors. MIT.
- llama.cpp, by Georgi Gerganov and contributors. MIT.
- sherpa-onnx, by the k2-fsa project. Apache 2.0.
- ONNX Runtime, by Microsoft. MIT.
This website
- Astro. MIT.
- React. MIT.
- three.js. MIT.
- React Three Fiber and Drei. MIT.
- Bricolage Grotesque, Instrument Sans and JetBrains Mono. SIL Open Font License 1.1.
Names and logos
Names and logos of other companies and apps, here and on the rest of this site, belong to their owners. Using them to describe what Ramblebee works with doesn’t mean they endorse us.