Open-source licenses

Effective October 31, 2026

AudioRinse is built on open-source software. Thank you to these projects.

DeepFilterNet

Used for the speech-enhancement model and its runtime (compiled to WebAssembly for the website, and built as a native program for AudioRinse in Claude, from the same pinned source, commit d375b2d).

License: MIT or Apache-2.0. © Hendrik Schröter and contributors.

Model trained on public speech and noise corpora, including the Microsoft DNS Challenge datasets; see the project's documentation for dataset credits.

https://github.com/Rikorose/DeepFilterNet

tract

Used for neural-network inference inside the DeepFilterNet runtime.

License: MIT or Apache-2.0. © Sonos, Inc. and contributors.

https://github.com/sonos/tract

Signalsmith Stretch

Used for changing pitch and speed separately in the pitch changer, key changer, speed changer and 432 Hz converter (its WebAssembly build, run in your browser).

License: MIT. © Signalsmith Audio Ltd..

https://signalsmith-audio.co.uk/code/stretch/

Mediabunny

Used for reading and writing audio and video files.

License: MPL-2.0. © Vanilagy and contributors.

Source code is available at the link.

https://github.com/Vanilagy/mediabunny

@mediabunny/aac-encoder

Used for AAC audio encoding in browsers without a built-in encoder.

License: MPL-2.0; includes the FFmpeg AAC encoder (libavcodec), LGPL-2.1-or-later. © Vanilagy; © the FFmpeg developers.

FFmpeg source code: https://ffmpeg.org/download.html. It is loaded as a separate module you can replace.

https://github.com/Vanilagy/mediabunny

@mediabunny/mp3-encoder

Used for MP3 encoding in the converters (browsers have no built-in MP3 encoder).

License: MPL-2.0; includes the LAME MP3 Encoder 3.100, LGPL-2.0-or-later. © Vanilagy; © the LAME project.

LAME source code: https://lame.sourceforge.io. It is loaded as a separate module you can replace.

https://lame.sourceforge.io

@mediabunny/ac3

Used for decoding Dolby Digital (AC-3, E-AC-3) sound, which browsers can't, for files that have it.

License: MPL-2.0; includes the FFmpeg AC-3 decoder (libavcodec), LGPL-2.1-or-later. © Vanilagy; © the FFmpeg developers.

FFmpeg source code: https://ffmpeg.org/download.html. It is loaded as a separate module you can replace.

https://github.com/Vanilagy/mediabunny

@mediabunny/dts

Used for decoding DTS sound, which browsers can't, for files that have it.

License: MPL-2.0; includes the FFmpeg DTS decoder (libavcodec), LGPL-2.1-or-later. © Vanilagy; © the FFmpeg developers.

FFmpeg source code: https://ffmpeg.org/download.html. It is loaded as a separate module you can replace.

https://github.com/Vanilagy/mediabunny

FFmpeg

Used for reading and writing audio and video files, the voice chain and loudness levelling in AudioRinse in Claude (the ffmpeg-static build, run as a separate program on our servers).

License: GPL-3.0. © the FFmpeg developers.

Source code: https://ffmpeg.org/download.html.

https://ffmpeg.org

WPE dereverberation

Used for room-echo reduction (our own implementation of Nakatani et al. 2010, the formulation of nara_wpe).

License: our own code. © Wikivec Inc..

https://github.com/fgnt/nara_wpe

Whisper large-v3-turbo and Whisper small

Used for speech recognition in the interview transcriber (the ONNX builds published by onnx-community, run in your browser).

License: MIT (large-v3-turbo); Apache-2.0 (small). © OpenAI.

https://github.com/openai/whisper

Transformers.js

Used for running Whisper in the browser.

License: Apache-2.0. © Hugging Face, Inc..

https://github.com/huggingface/transformers.js

ONNX Runtime Web

Used for neural-network inference for the transcriber, the vocal remover, the stem splitter and the sheet-music tool (WebGPU and WebAssembly).

License: MIT. © Microsoft Corporation.

https://github.com/microsoft/onnxruntime

HTDemucs (Hybrid Transformer Demucs)

Used for Standard quality in the vocal remover and stem splitter (the official htdemucs weights, as exported to ONNX by monteslu, fp16, and timcsy, fp32).

License: MIT. © Meta Platforms, Inc. and affiliates (Alexandre Défossez, Simon Rouard, Francisco Massa).

https://github.com/facebookresearch/demucs

Mel-Band RoFormer (vocals)

Used for Best quality in the vocal remover: Kimberley Jensen's vocals model, exported to ONNX by audiojs.

License: MIT. © Kimberley Jensen (weights); architecture by Ju-Chiang Wang, Wei-Tsung Lu, Minz Won, Kuan-Yi Chen and Xuchen Song (ByteDance); code by Phil Wang and Roman Solovyev.

https://huggingface.co/KimberleyJSN/melbandroformer

High-resolution piano transcription (piano_transcription_inference)

Used for hearing the notes in the sheet-music tool: the note model of the checkpoint CRNN_note_F1=0.9677_pedal_F1=0.9186 (Zenodo 4034264), run in your browser.

License: CC BY 4.0 (the checkpoint, https://creativecommons.org/licenses/by/4.0/); MIT (the code). © Qiuqiang Kong, Bochen Li, Xuchen Song, Yuan Wan and Yuxuan Wang.

Changed by us: converted to ONNX, split into its convolutional and recurrent halves, the pedal model left out, weights stored at half precision. The converted files are published with this notice.

https://github.com/qiuqiangkong/piano_transcription_inference

PM2S (performance MIDI to score)

Used for the beats, bar lines, hands and key signature in the sheet-music tool: its beat, hand-part, key-signature and time-signature models (Zenodo 10520196), run in your browser; the post-processing of their outputs is ported to TypeScript.

License: CC BY 4.0 (the checkpoints, https://creativecommons.org/licenses/by/4.0/); MIT (the code). © 2022 Lele Liu; models by Lele Liu, Qiuqiang Kong, Veronica Morfi and Emmanouil Benetos.

Changed by us: converted to ONNX, the beat model's unused tempo output removed, weights stored at half precision.

https://github.com/cheriell/PM2S

EfficientAT (mn10_as)

Used for the piano check in the sheet-music tool, an AudioSet sound tagger that confirms a recording has a piano in it.

License: MIT. © 2022 Florian Schmid (models by Florian Schmid, Khaled Koutini and Gerhard Widmer).

Converted by us to ONNX with half-precision weights.

https://github.com/fschmid56/EfficientAT

OpenSheetMusicDisplay

Used for drawing the score in the sheet-music tool, on the page and in the PDF.

License: BSD-3-Clause. © 2019 PhonicScore.

https://github.com/opensheetmusicdisplay/opensheetmusicdisplay

VexFlow

Used for music engraving inside OpenSheetMusicDisplay.

License: MIT. © 2010 Mohit Muthanna Cheppudira.

https://github.com/0xfe/vexflow

jsPDF and svg2pdf.js

Used for writing the sheet-music PDF.

License: MIT. © 2010–2025 James Hall; © 2015–2026 yWorks GmbH.

https://github.com/parallax/jsPDF

sherpa-onnx

Used for speaker detection in the transcriber (its WebAssembly build of offline speaker diarization, v1.13.7).

License: Apache-2.0. © Xiaomi Corporation and the k2-fsa contributors.

https://github.com/k2-fsa/sherpa-onnx

pyannote segmentation 3.0

Used for finding who speaks when, in the transcriber.

License: MIT. © CNRS.

https://huggingface.co/pyannote/segmentation-3.0

NVIDIA NeMo TitaNet-small

Used for telling speakers' voices apart, in the transcriber.

License: the NeMo Toolkit license (Apache-2.0). © NVIDIA Corporation.

https://catalog.ngc.nvidia.com/orgs/nvidia/teams/nemo/models/titanet_small

Kokoro-82M

Used for the synthetic voices in the before/after examples.

License: Apache-2.0. © hexgrad and contributors.

https://huggingface.co/hexgrad/Kokoro-82M

Archivo

Used for the main typeface (self-hosted with the site, so your browser never contacts Google Fonts; static cuts are used for link-preview images).

License: SIL Open Font License 1.1. © The Archivo Project Authors.

https://github.com/Omnibus-Type/Archivo

JetBrains Mono

Used for the monospaced typeface for labels and readouts (self-hosted with the site).

License: SIL Open Font License 1.1. © The JetBrains Mono Project Authors.

https://github.com/JetBrains/JetBrainsMono

Next.js and React

Used for the website.

License: MIT. © Vercel, Inc.; © Meta Platforms, Inc. and affiliates.

https://github.com/vercel/next.js

The room-echo reduction uses the weighted prediction error (WPE) method described by Nakatani et al. (2010) and Yoshioka & Nakatani (2012), implemented by us.