Open-source credits
PaperEdit transcribes, tells speakers apart and translates on your own computer, with these open models and libraries. Thank you to their authors. If you choose, it translates with Anthropic's Claude Haiku or Claude Sonnet instead.
Models
- Whisper large-v3-turbo · Transcription
By OpenAI. Licence: MIT.
- Whisper base · Transcription without a graphics chip
By OpenAI. Licence: MIT.
- WeSpeaker ResNet34-LM (VoxCeleb) · Telling speakers apart
By the WeSpeaker project (Hongji Wang et al.). Licence: CC BY 4.0.
- M2M100 418M · Translation between other languages
By Meta AI (Angela Fan et al.). Licence: MIT.
- Opus-MT · Translation
By Jörg Tiedemann, Santhosh Thottingal and the Language Technology Research Group at the University of Helsinki. Licence: Apache 2.0; CC BY 4.0 for English→German, Russian→English and Chinese→English.
Libraries
- Transformers.js · Running the models
By Hugging Face. Licence: Apache 2.0.
- ONNX Runtime Web · Running the models
By Microsoft. Licence: MIT.
- Mediabunny · Reading and making media files
By Vanilagy. Licence: MPL 2.0.
- FFmpeg's ProRes and DNxHD decoders, built with libav.js · Small picture copies of ProRes and DNx interviews
By the FFmpeg developers; libav.js by Yahweasel. Licence: LGPL 2.1 (libav.js unmodified; FFmpeg built for speed, its ProRes decoder with our half-size patch, public/paperedit/libav-3/prores-lowres.patch; the build is described in public/paperedit/libav-3/README.txt).
Translation with Claude
- Claude Haiku and Claude Sonnet · Translation, when you choose it
By Anthropic. Not open source, and they don't run on your computer: the transcript's text and speaker names go to Anthropic to be translated, never the sound or the picture. Claude Haiku is Standard, included with your plan's hours. Claude Sonnet is Best quality, bought by the hour.
Accuracy was measured with Rev's Earnings-21 recordings and Meta's FLORES-200 sentences (both CC BY-SA 4.0). Neither is shipped with PaperEdit.