Open-source credits

NueForm™ is built with open-source software and openly licensed models. Thank you to the people who make them.

These are the open-source projects and models behind features you use, with their licenses and the attribution their authors ask for.

Voice recording clean-up

When you clone a voice, NueForm can clean up your recording on our servers (background noise, hum and level) and check that the cleaned version still sounds like you. These run inside our audio engine:

  • DeepFilterNet

    Removes background noise from a voice recording.

    License: MIT or Apache-2.0 (code and model weights)

    DeepFilterNet3 by Hendrik Schröter et al. Code and model weights © Hendrik Schröter.

    Source: github.com/Rikorose/DeepFilterNet

  • tract

    Runs the noise-removal model inside our audio engine.

    License: MIT or Apache-2.0

    tract by Sonos.

    Source: github.com/sonos/tract

  • WeSpeaker CAM++ speaker model

    Checks that a cleaned recording still sounds like the same person.

    License: CC BY 4.0

    Speaker-similarity check: WeSpeaker CAM++ model (WeNet community, github.com/wenet-e2e/wespeaker), trained on VoxCeleb1 and VoxCeleb2 (A. Nagrani, J. S. Chung, W. Xie, A. Zisserman, University of Oxford), licensed under CC BY 4.0 (creativecommons.org/licenses/by/4.0). Model architecture: H. Wang et al., "CAM++: A Fast and Efficient Network for Speaker Verification Using Context-Aware Masking", Interspeech 2023. Converted to ONNX by WeSpeaker; run unmodified.

    Only the trained model is used, never VoxCeleb's audio, whose copyright stays with its owners.

    Source: github.com/wenet-e2e/wespeaker

Our audio engine also uses many smaller open-source Rust libraries, all under permissive licenses (MIT, Apache-2.0 and similar).

Questions about licensing? Contact us