Hearsay is built on open work by the people and teams below. Their models run on your own device, in your browser; Hearsay uses them without modification, in the ONNX versions published by the onnx-community group on Hugging Face so that they can run in a web page.
OpenAI
Turns speech into words, with the time of each word. Hearsay offers three sizes: whisper-base (Fast), whisper-small (Balanced) and whisper-large-v3-turbo (Most accurate).
Hervé Bredin, Alexis Plaquet and the pyannote.audio project
Finds where people are speaking and when the speaker changes, which Hearsay uses to cut the audio at pauses and to tell voices apart.
The WeSpeaker team (WeNet community), packaged by pyannote
Makes a “voiceprint” from a few seconds of someone's voice, so that Hearsay can group the parts of a recording spoken by the same person. Trained on the VoxCeleb dataset.
Hugging Face
Runs the models in the browser. github.com/huggingface/transformers.js · Apache 2.0
Microsoft
The engine underneath, using your graphics card (WebGPU) or processor (WebAssembly). github.com/microsoft/onnxruntime · MIT
Anthropic
Bundled with this site and loaded only when you use the optional Claude features with your own API key. github.com/anthropics/anthropic-sdk-typescript · Apache 2.0
Inter by Rasmus Andersson and Literata by TypeTogether, both under the SIL Open Font License 1.1, served by Google Fonts.
The “Try a sample” interview is synthetic: a script read by the Microsoft Zira and David speech voices in Windows. The people in it are fictional.
Hearsay's own code and design are © 2026 Applied Arts. The models, software and typefaces above belong to their authors and are used under the licenses shown. Transcripts you make are yours.