Open-source licences and attribution

StemSing builds on the following models and code. Only components whose code and weights permit commercial use are shipped; the full machine-readable list is in THIRD_PARTY_LICENSES.

Pocket TTS (on-device voice engine)

Model weights CC-BY-4.0 (Kyutai); browser implementation MIT (vlapky); ONNX export Apache-2.0 (vlapky)

Attribution: Pocket TTS by Kyutai, ONNX conversion and browser implementation by vlapky.

Used for on-device voice cloning and computer speech. Commercial use permitted with attribution. Cloning requires explicit lawful consent; AI output is labelled and watermarked.

https://huggingface.co/kyutai/pocket-tts

WeSpeaker ResNet34 (independent speaker verification)

Apache-2.0

Attribution: WeSpeaker by WeNet Community.

Server-side WASM speaker embedding used to verify that consent and enrolment recordings are the same voice.

https://github.com/wenet-e2e/wespeaker

Whisper base (consent phrase transcription)

MIT (OpenAI code and weights); Apache-2.0 (ONNX conversion tooling)

Attribution: Whisper by OpenAI; ONNX export by onnx-community.

Independent, unconstrained transcription of the stored consent recording; the expected phrase is never injected.

https://github.com/openai/whisper

ONNX Runtime Web

MIT

Attribution: Microsoft ONNX Runtime.

Runs the voice and verification models in the browser (WebGPU/WASM) and on the server (WASM).

https://github.com/microsoft/onnxruntime

Demucs / MDX separation models

MIT (code); model weights as listed in THIRD_PARTY_LICENSES

Attribution: Meta Research and MDX challenge contributors.

Browser stem separation for karaoke making.

https://github.com/facebookresearch/demucs