Open source
Saypane stands on open source software and openly licensed speech models. The Saypane app itself is closed source. Here is what it uses, and under which license.
Speech models
| Project | License | Used for |
|---|---|---|
| NVIDIA Parakeet TDT 0.6B v3 | CC-BY-4.0 | Speech recognition for 25 European languages. |
| OpenAI Whisper | MIT | Speech recognition for the other languages (small and large v3 turbo). |
| Silero VAD | MIT | Voice activity detection: finding speech and pauses in recordings. |
| pyannote segmentation 3.0 | MIT | Speaker segmentation for speaker labels. |
Software
| Project | License | Used for |
|---|---|---|
| whisper.cpp | MIT | Runs the Whisper and Parakeet models on the CPU and through Vulkan on the GPU. |
| ggml | MIT | The tensor library under whisper.cpp. |
| sherpa-onnx | Apache-2.0 | Runs the speaker label models. |
| NAudio | MIT | Microphone and PC sound capture. |
| CommunityToolkit | MIT | Building blocks of the app user interface. |
| Windows App SDK | MIT | The Windows app framework (WinUI). |
Attribution: NVIDIA Parakeet TDT 0.6B v3
Saypane uses NVIDIA Parakeet TDT 0.6B v3 by NVIDIA, licensed under the Creative Commons Attribution 4.0 International License (CC-BY-4.0). Model card: huggingface.co/nvidia/parakeet-tdt-0.6b-v3. Saypane ships a GGML conversion of the model quantized to 8 bits (q8_0), as published at huggingface.co/ggml-org/parakeet-GGUF; the model was converted and quantized, not retrained or otherwise changed.
Whisper model files are the GGML conversions published at huggingface.co/ggerganov/whisper.cpp. The speaker embedding model used together with pyannote segmentation will be listed here with its license when it is final.
Audio and video files are decoded with Windows Media Foundation, which is part of Windows. Saypane does not include FFmpeg.
Thank you to the authors and contributors of these projects. The full license texts are included with the app in its third party notices.