Credits

Audio Magic is a browser wrapper around other people’s research. The speech recognition, noise suppression and translation models below were built and published by the researchers and engineers named here, and released under licences that let anyone use them. This page is where they get the credit, in the citation format each project asks for.

Artwork

The mage — the mark in every masthead here, and the face of Audio Magic — was drawn by Gobby, an artist and musician from the Bronx.

Speech recognition

Transcription runs OpenAI’s Whisper, a transformer trained on 680,000 hours of multilingual audio.

  1. [1]

    Whisper

    MIT

    Transcribing speech to text and subtitles

    Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. 2022. Robust Speech Recognition via Large-Scale Weak Supervision. arXiv:2212.04356.

    Read the paperProject source

    +BibTeX
    @article{radford2022whisper,
      title   = {Robust Speech Recognition via Large-Scale Weak Supervision},
      author  = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg
                 and McLeavey, Christine and Sutskever, Ilya},
      journal = {arXiv preprint arXiv:2212.04356},
      year    = {2022}
    }

Speech synthesis

Generate Speech runs Kokoro, an 82-million-parameter voice model small enough to send to your browser instead of sending your text to a server. Clone a Voice runs Chatterbox, which copies a voice from a few seconds of it.

  1. [2]

    Chatterbox

    MIT

    Cloning a voice from a short recording and speaking text in it

    Resemble AI. 2025. Chatterbox-TTS. Model weights and code, released under the MIT licence.

    Project source

    +BibTeX
    @misc{chatterboxtts2025,
      author       = {{Resemble AI}},
      title        = {{Chatterbox-TTS}},
      year         = {2025},
      howpublished = {\url{https://github.com/resemble-ai/chatterbox}},
      note         = {GitHub repository}
    }
  2. [3]

    Kokoro

    Apache-2.0

    Generating spoken audio from text

    hexgrad. 2025. Kokoro-82M. Model weights, released under the Apache 2.0 licence.

    Project source

  3. [4]

    StyleTTS 2

    MIT

    The synthesis architecture Kokoro is built on

    Yinghao Aaron Li, Cong Han, Vinay S. Raghavan, Gavin Mischler, and Nima Mesgarani. 2023. StyleTTS 2: Towards Human-Level Text-to-Speech through Style Diffusion and Adversarial Training with Large Speech Language Models. In Advances in Neural Information Processing Systems (NeurIPS). arXiv:2306.07691.

    Read the paperProject source

    +BibTeX
    @inproceedings{li2023styletts2,
      title     = {{StyleTTS} 2: Towards Human-Level Text-to-Speech through Style
                   Diffusion and Adversarial Training with Large Speech Language
                   Models},
      author    = {Li, Yinghao Aaron and Han, Cong and Raghavan, Vinay S. and
                   Mischler, Gavin and Mesgarani, Nima},
      booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
      year      = {2023}
    }
  4. [5]

    iSTFTNet

    Apache-2.0

    The vocoder that turns Kokoro’s output into a waveform

    Takuhiro Kaneko, Kou Tanaka, Hirokazu Kameoka, and Shogo Seki. 2022. iSTFTNet: Fast and Lightweight Mel-Spectrogram Vocoder Incorporating Inverse Short-Time Fourier Transform. In Proceedings of ICASSP. arXiv:2203.02395.

    Read the paperProject source

    +BibTeX
    @inproceedings{kaneko2022istftnet,
      title     = {{iSTFTNet}: Fast and Lightweight Mel-Spectrogram Vocoder
                   Incorporating Inverse Short-Time Fourier Transform},
      author    = {Kaneko, Takuhiro and Tanaka, Kou and Kameoka, Hirokazu and
                   Seki, Shogo},
      booktitle = {Proc. IEEE ICASSP 2022},
      year      = {2022}
    }
  5. [6]

    phonemize

    MIT

    Converting text into the phonemes Kokoro reads, without a GPL dependency

    hans00. 2025. phonemize: a pure-JavaScript rule-based grapheme-to-phoneme converter.

    Project source

Speech enhancement

Both noise-removal engines pair a neural network with classical signal processing — which is what makes them fast enough to run on your own machine.

  1. [7]

    RNNoise

    BSD-3-Clause

    Fast background noise removal

    Jean-Marc Valin. 2018. A Hybrid DSP/Deep Learning Approach to Real-Time Full-Band Speech Enhancement. In Proceedings of the IEEE International Workshop on Multimedia Signal Processing (MMSP). arXiv:1709.08243.

    Read the paperProject source

    +BibTeX
    @inproceedings{valin2018rnnoise,
      title     = {A Hybrid {DSP}/Deep Learning Approach to Real-Time Full-Band
                   Speech Enhancement},
      author    = {Valin, Jean-Marc},
      booktitle = {Proceedings of the IEEE International Workshop on Multimedia
                   Signal Processing (MMSP)},
      year      = {2018}
    }
  2. [8]

    DeepFilterNet3

    MIT or Apache-2.0

    Higher-quality background noise removal

    Hendrik Schröter, Tobias Rosenkranz, Alberto N. Escalante-B., and Andreas Maier. 2023. DeepFilterNet: Perceptually Motivated Real-Time Speech Enhancement. In Proceedings of INTERSPEECH. arXiv:2305.08227.

    Read the paperProject source

    +BibTeX
    @inproceedings{schroeter2023deepfilternet3,
      title     = {{DeepFilterNet}: Perceptually Motivated Real-Time Speech
                   Enhancement},
      author    = {Schr{\"o}ter, Hendrik and Rosenkranz, Tobias and
                   Escalante-B., Alberto N. and Maier, Andreas},
      booktitle = {Proceedings of INTERSPEECH},
      year      = {2023}
    }
  3. [9]

    DeepFilterNet

    MIT or Apache-2.0

    The deep-filtering framework DeepFilterNet3 is built on

    Hendrik Schröter, Alberto N. Escalante-B., Tobias Rosenkranz, and Andreas Maier. 2022. DeepFilterNet: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering. In Proceedings of ICASSP.

    Read the paperProject source

    +BibTeX
    @inproceedings{schroeter2022deepfilternet,
      title     = {{DeepFilterNet}: A Low Complexity Speech Enhancement Framework
                   for Full-Band Audio based on Deep Filtering},
      author    = {Schr{\"o}ter, Hendrik and Escalante-B., Alberto N. and
                   Rosenkranz, Tobias and Maier, Andreas},
      booktitle = {ICASSP 2022 - IEEE International Conference on Acoustics,
                   Speech and Signal Processing},
      year      = {2022}
    }

Music source separation

Splitting a song back into vocals, drums, bass and everything else is done by Demucs — a hybrid model that works in both the spectrogram and the waveform domain.

  1. [10]

    HT-Demucs (Demucs v4)

    MIT

    Separating a song into stems, and removing vocals

    Simon Rouard, Francisco Massa, and Alexandre Défossez. 2023. Hybrid Transformers for Music Source Separation. In Proceedings of ICASSP. arXiv:2211.08553.

    Read the paperProject source

    +BibTeX
    @inproceedings{rouard2023hybrid,
      title     = {Hybrid Transformers for Music Source Separation},
      author    = {Rouard, Simon and Massa, Francisco and D{\'e}fossez, Alexandre},
      booktitle = {ICASSP 23},
      year      = {2023}
    }
  2. [11]

    Hybrid Demucs

    MIT

    The hybrid spectrogram/waveform design HT-Demucs builds on

    Alexandre Défossez. 2021. Hybrid Spectrogram and Waveform Source Separation. In Proceedings of the ISMIR 2021 Workshop on Music Source Separation.

    Read the paperProject source

    +BibTeX
    @inproceedings{defossez2021hybrid,
      title     = {Hybrid Spectrogram and Waveform Source Separation},
      author    = {D{\'e}fossez, Alexandre},
      booktitle = {Proceedings of the ISMIR 2021 Workshop on Music Source Separation},
      year      = {2021}
    }

Sound event detection

Detect Sounds recognises a sound you teach it by standing on Google’s YAMNet — a network trained on AudioSet, two million human-labelled clips from YouTube — and learning only the last step from your own examples.

  1. [12]

    YAMNet

    Apache-2.0

    Turning audio into the embeddings a custom sound detector is trained on

    Shawn Hershey, Sourish Chaudhuri, Daniel P. W. Ellis, Jort F. Gemmeke, Aren Jansen, R. Channing Moore, Manoj Plakal, Devin Platt, Rif A. Saurous, Bryan Seybold, Malcolm Slaney, Ron J. Weiss, and Kevin Wilson. 2017. CNN Architectures for Large-Scale Audio Classification. In Proceedings of ICASSP. arXiv:1609.09430.

    Read the paperProject source

    +BibTeX
    @inproceedings{hershey2017cnn,
      title     = {{CNN} Architectures for Large-Scale Audio Classification},
      author    = {Hershey, Shawn and Chaudhuri, Sourish and Ellis, Daniel P. W. and
                   Gemmeke, Jort F. and Jansen, Aren and Moore, R. Channing and
                   Plakal, Manoj and Platt, Devin and Saurous, Rif A. and
                   Seybold, Bryan and Slaney, Malcolm and Weiss, Ron J. and
                   Wilson, Kevin},
      booktitle = {2017 IEEE International Conference on Acoustics, Speech and
                   Signal Processing (ICASSP)},
      year      = {2017}
    }
  2. [13]

    AudioSet

    CC BY 4.0 (ontology and labels)

    The labelled corpus YAMNet learned what sounds are made of from

    Jort F. Gemmeke, Daniel P. W. Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R. Channing Moore, Manoj Plakal, and Marvin Ritter. 2017. Audio Set: An ontology and human-labeled dataset for audio events. In Proceedings of ICASSP, New Orleans, LA.

    Read the paperProject source

    +BibTeX
    @inproceedings{gemmeke2017audioset,
      title     = {Audio Set: An ontology and human-labeled dataset for audio events},
      author    = {Gemmeke, Jort F. and Ellis, Daniel P. W. and Freedman, Dylan and
                   Jansen, Aren and Lawrence, Wade and Moore, R. Channing and
                   Plakal, Manoj and Ritter, Marvin},
      booktitle = {Proc. IEEE ICASSP 2017},
      address   = {New Orleans, LA},
      year      = {2017}
    }
  3. [14]

    MobileNets

    Apache-2.0

    The depthwise-separable convolution architecture YAMNet is built on

    Andrew G. Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. 2017. MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. arXiv:1704.04861.

    Read the paperProject source

    +BibTeX
    @article{howard2017mobilenets,
      title   = {{MobileNets}: Efficient Convolutional Neural Networks for Mobile
                 Vision Applications},
      author  = {Howard, Andrew G. and Zhu, Menglong and Chen, Bo and
                 Kalenichenko, Dmitry and Wang, Weijun and Weyand, Tobias and
                 Andreetto, Marco and Adam, Hartwig},
      journal = {arXiv preprint arXiv:1704.04861},
      year    = {2017}
    }

Speaker diarization

Label Speakers is two networks and a clustering step: pyannote’s segmentation model finds when each voice is active in a ten-second window, and a speaker-embedding network turns those excerpts into vectors close enough to group. Neither the segmentation nor the identity comes from one model alone.

  1. [15]

    pyannote powerset segmentation

    MIT

    Finding when each person is speaking, overlap included

    Alexis Plaquet and Hervé Bredin. 2023. Powerset multi-class cross entropy loss for neural speaker diarization. In Proceedings of INTERSPEECH 2023.

    Read the paperProject source

    +BibTeX
    @inproceedings{Plaquet23,
      author    = {Plaquet, Alexis and Bredin, Herv{\'e}},
      title     = {{Powerset multi-class cross entropy loss for neural speaker
                   diarization}},
      booktitle = {Proc. INTERSPEECH 2023},
      year      = {2023}
    }
  2. [16]

    pyannote.audio

    MIT

    The diarization pipeline this tool reimplements in the browser

    Hervé Bredin. 2023. pyannote.audio 2.1 speaker diarization pipeline: principle, benchmark, and recipe. In Proceedings of INTERSPEECH 2023.

    Project source

    +BibTeX
    @inproceedings{Bredin23,
      author    = {Bredin, Herv{\'e}},
      title     = {{pyannote.audio 2.1 speaker diarization pipeline: principle,
                   benchmark, and recipe}},
      booktitle = {Proc. INTERSPEECH 2023},
      year      = {2023}
    }
  3. [17]

    WeSpeaker

    Apache-2.0

    The ResNet34 speaker embedding model, one of the two voice comparisons offered

    Hongji Wang, Chengdong Liang, Shuai Wang, Zhengyang Chen, Binbin Zhang, Xu Xiang, Yanlei Deng, and Yanmin Qian. 2023. Wespeaker: A Research and Production oriented Speaker Embedding Learning Toolkit. In Proceedings of ICASSP. arXiv:2210.17016.

    Read the paperProject source

    +BibTeX
    @inproceedings{wang2023wespeaker,
      title     = {Wespeaker: A Research and Production oriented Speaker Embedding
                   Learning Toolkit},
      author    = {Wang, Hongji and Liang, Chengdong and Wang, Shuai and
                   Chen, Zhengyang and Zhang, Binbin and Xiang, Xu and
                   Deng, Yanlei and Qian, Yanmin},
      booktitle = {ICASSP 2023 - 2023 IEEE International Conference on Acoustics,
                   Speech and Signal Processing (ICASSP)},
      year      = {2023}
    }
  4. [18]

    ERes2Net

    Apache-2.0

    The other voice comparison model, the faster of the two

    Yafeng Chen, Siqi Zheng, Hui Wang, Luyao Cheng, Qian Chen, and Jiajun Qi. 2023. An Enhanced Res2Net with Local and Global Feature Fusion for Speaker Verification. In Proceedings of INTERSPEECH 2023. arXiv:2305.12838.

    Read the paperProject source

    +BibTeX
    @inproceedings{chen2023eres2net,
      title     = {An Enhanced {Res2Net} with Local and Global Feature Fusion for
                   Speaker Verification},
      author    = {Chen, Yafeng and Zheng, Siqi and Wang, Hui and Cheng, Luyao and
                   Chen, Qian and Qi, Jiajun},
      booktitle = {Proc. INTERSPEECH 2023},
      year      = {2023}
    }
  5. [19]

    3D-Speaker

    Apache-2.0

    The toolkit ERes2Net’s released checkpoint was trained and published with

    Yafeng Chen, Siqi Zheng, Hui Wang, Luyao Cheng, Tinglong Zhu, Rongjie Huang, Chong Deng, Qian Chen, Shiliang Zhang, Wen Wang, and Xihao Li. 2024. 3D-Speaker-Toolkit: An Open-Source Toolkit for Multimodal Speaker Verification and Diarization. arXiv:2403.19971.

    Read the paperProject source

    +BibTeX
    @article{chen20243dspeaker,
      title   = {{3D-Speaker-Toolkit}: An Open-Source Toolkit for Multimodal
                 Speaker Verification and Diarization},
      author  = {Chen, Yafeng and Zheng, Siqi and Wang, Hui and Cheng, Luyao and
                 Zhu, Tinglong and Huang, Rongjie and Deng, Chong and Chen, Qian and
                 Zhang, Shiliang and Wang, Wen and Li, Xihao},
      journal = {arXiv preprint arXiv:2403.19971},
      year    = {2024}
    }
  6. [20]

    VoxCeleb

    CC BY 4.0

    The corpus both speaker embedding models learned to tell voices apart from

    Arsha Nagrani, Joon Son Chung, and Andrew Zisserman. 2017. VoxCeleb: a large-scale speaker identification dataset. In Proceedings of INTERSPEECH 2017. arXiv:1706.08612.

    Read the paperProject source

    +BibTeX
    @inproceedings{nagrani2017voxceleb,
      title     = {{VoxCeleb}: a large-scale speaker identification dataset},
      author    = {Nagrani, Arsha and Chung, Joon Son and Zisserman, Andrew},
      booktitle = {Proc. INTERSPEECH 2017},
      year      = {2017}
    }

Machine translation

Translated transcripts come from Helsinki-NLP’s OPUS-MT models, trained with the Marian toolkit.

  1. [21]

    OPUS-MT

    Apache-2.0

    Translating transcripts into other languages

    Jörg Tiedemann and Santhosh Thottingal. 2020. OPUS-MT – Building open translation services for the World. In Proceedings of the 22nd Annual Conference of the European Association for Machine Translation (EAMT), pages 479–480.

    Read the paperProject source

    +BibTeX
    @inproceedings{tiedemann-thottingal-2020-opus,
      title     = {{OPUS-MT} {--} Building open translation services for the World},
      author    = {Tiedemann, J{\"o}rg and Thottingal, Santhosh},
      booktitle = {Proceedings of the 22nd Annual Conference of the European
                   Association for Machine Translation},
      pages     = {479--480},
      year      = {2020}
    }
  2. [22]

    Marian NMT

    MIT

    The toolkit the OPUS-MT models were trained with

    Marcin Junczys-Dowmunt, Roman Grundkiewicz, Tomasz Dwojak, Hieu Hoang, Kenneth Heafield, Tom Neckermann, Frank Seide, Ulrich Germann, Alham Fikri Aji, Nikolay Bogoychev, André F. T. Martins, and Alexandra Birch. 2018. Marian: Fast Neural Machine Translation in C++. In Proceedings of ACL 2018, System Demonstrations, pages 116–121.

    Read the paperProject source

    +BibTeX
    @inproceedings{junczys-dowmunt-etal-2018-marian,
      title     = {{M}arian: Fast Neural Machine Translation in {C}++},
      author    = {Junczys-Dowmunt, Marcin and Grundkiewicz, Roman and
                   Dwojak, Tomasz and Hoang, Hieu and Heafield, Kenneth and
                   Neckermann, Tom and Seide, Frank and Germann, Ulrich and
                   Aji, Alham Fikri and Bogoychev, Nikolay and
                   Martins, Andr{\'e} F. T. and Birch, Alexandra},
      booktitle = {Proceedings of ACL 2018, System Demonstrations},
      pages     = {116--121},
      year      = {2018}
    }

Audio codecs

Browsers decode MP3 but will not encode it, so writing an MP3 takes someone else’s encoder.

  1. [23]

    LAME (via lamejs)

    LGPL-3.0

    Encoding MP3 files in Convert Audio

    The LAME project, https://lame.sourceforge.net. JavaScript port: lamejs by Alex Zhukov (@zhuker), a rewrite of jump3r-code, itself a rewrite of libmp3lame. Used unmodified as @breezystack/lamejs 1.2.7.

    Project sourceLicence notice

Colour

Make a Spectrogram is our own signal processing, but the colour ramps it maps loudness onto are somebody else’s research. Heat, Ice, Grayscale and anything built in the custom editor are not — they are ours, and are absent here for that reason.

  1. [24]

    Viridis, Magma, Inferno and Plasma

    CC0-1.0 (public domain dedication)

    The perceptually uniform palettes in Make a Spectrogram

    Nathaniel J. Smith and Stéfan van der Walt. 2015. A Better Default Colormap for Matplotlib. SciPy 2015. Viridis additionally credited to Eric Firing.

    Read the paperProject source

  2. [25]

    Cividis

    CC0-1.0 (public domain dedication)

    The colour-vision-deficiency palette in Make a Spectrogram

    Jamie R. Nuñez, Christopher R. Anderton, and Ryan S. Renslow. 2018. Optimizing colormaps with consideration for color vision deficiency to enable accurate interpretation of scientific data. PLOS ONE 13(7): e0199239.

    Read the paperProject source

  3. [26]

    Turbo

    Apache-2.0

    The rainbow palette in Make a Spectrogram

    Anton Mikhailov, Google LLC. 2019. Turbo, An Improved Rainbow Colormap for Visualization. Published as a 256-entry sRGB lookup table.

    Project sourceLicence notice

  4. [27]

    Cubehelix

    Free to use, citation requested

    The greyscale-safe palette in Make a Spectrogram

    D. A. Green. 2011. A colour scheme for the display of astronomical intensity images. Bulletin of the Astronomical Society of India, 39, 289.

    Read the paperProject source

Running models in the browser

None of the above would run client-side without the work that compiled these runtimes to WebAssembly.

  1. [28]

    Transformers.js

    Apache-2.0

    Running the Whisper and OPUS-MT models in a Web Worker

    Joshua Lochner and the Hugging Face team. Transformers.js: State-of-the-art machine learning for the web.

    Project source

  2. [29]

    ONNX Runtime

    MIT

    Executing the models via WebAssembly

    ONNX Runtime developers. 2021. ONNX Runtime: Cross-platform accelerated machine learning.

    Project source

  3. [30]

    rnnoise-wasm

    Apache-2.0

    The WebAssembly build of RNNoise we vendor

    Tim Guggenmos (@timephy). rnnoise-wasm.

    Project source

  4. [31]

    demucs.cpp

    MIT

    The C++/WebAssembly engine that runs HT-Demucs in Separate Stems

    Sevag H (@sevagh). demucs.cpp — a C++17 inference library for the Demucs v3/v4 hybrid transformer music source separation models, built on Eigen and compiled to WebAssembly.

    Project source

Typefaces

The site is set in Inter, Rasmus Andersson’s grotesque drawn for screen text at small sizes. Measured values — the figure’s axes and captions — stay in Computer Modern’s typewriter cut, which Donald Knuth designed with TeX and which holds a column of numbers in line because every character is the same width.

  1. [32]

    Inter

    SIL OFL-1.1

    Every letterform on the site except the figure annotations

    Rasmus Andersson. Inter — a typeface carefully crafted and designed for computer screens.

    Project sourceLicence notice

  2. [33]

    Computer Modern Unicode

    SIL OFL-1.1

    The monospaced figure annotations — axis chrome, captions, and measured values

    Andrey V. Panov. Computer Modern Unicode fonts — Type 1 and OpenType conversions of Donald E. Knuth’s Computer Modern family, extended to Unicode coverage.

    Project sourceLicence notice

Corrections

Every citation above was checked against the primary source, but if your work is listed incorrectly — or is used here and missing entirely — that is a bug worth reporting, and it will be fixed.