TapeText

Open-source licences

TapeText would not exist without the work below. Everything listed here is downloaded to your browser when you use the tool, which makes publishing it distribution, which is why each entry names its licence and links to its source.

Whisper (small and base, ONNX)

The speech recognition model itself. Converted to ONNX by the onnx-community project, in the “timestamped” export that also emits cross-attention weights, which is what makes word-level timings possible.

Model weights © OpenAI, released under the MIT licence. The ONNX conversions are published by onnx-community on Hugging Face.

MITsource ↗licence text

Transformers.js

Runs the model: tokenisation, the audio feature extractor, the generation loop and the cross-attention alignment that produces word timings.

Apache-2.0source ↗licence text

ONNX Runtime Web

Executes the model graph, on WebGPU where the browser has it and on WebAssembly otherwise. The 21 MB binary served from /ort/.

MITsource ↗licence text

Mediabunny

Opens the media container and pulls out the audio track: MP4, QuickTime, Matroska, WebM, MP3, WAV, OGG, FLAC.

The Mozilla Public License requires that recipients of the compiled form be told how to obtain the source. It is at the address above, and the exact version in use is recorded in this site’s package-lock.json.

MPL-2.0source ↗licence text

Next.js, React and React DOM

The framework and view layer the pages are built with.

MITsource ↗licence text

@vercel/analytics

Counts page views and the small fixed set of events described on the privacy page.

MITsource ↗licence text

Syne, Schibsted Grotesk and DM Mono

The three typefaces. Self-hosted, so the browser fetches them from this site and never from a font CDN.

Copyright 2017 The Syne Project Authors; Copyright 2023 The Schibsted-Grotesk Project Authors; Copyright 2020 The DM Mono Project Authors. Serving a font file to a browser is distribution under the SIL Open Font License, which asks that the licence travel with it. None of the three carries a Reserved Font Name, which is what makes the subsetting next/font does unproblematic, since a subset is a Modified Version under OFL §3 and may not keep a reserved name.

OFL-1.1source ↗licence text

If something is used here and missing from this list, that is a mistake worth telling us about rather than a decision. The address is on the legal notice.