EVP Labs
Credits
How the model was built
EVP Labs includes a small model that runs on your phone and measures voice-like acoustic structure. It was trained on speech with known labels (spoken and whispered) and on labeled everyday sounds such as knocks, rustling, hums and animal calls. It was not trained on recordings described as paranormal. None of the audio below is inside the app, except twelve test voices of about one second each that the app uses for its detection check: six cut from LibriSpeech (CC BY 4.0, shortened and level-adjusted) and six from the Thorsten-Voice whispered recordings (CC0 1.0).
Data sources
LibriSpeech (CC BY 4.0). V. Panayotov, G. Chen, D. Povey, S. Khudanpur, ICASSP 2015.
FSD50K (per clip CC0 1.0 or CC BY only). E. Fonseca, X. Favory, J. Pons, F. Font, X. Serra, TASLP 2022. The sounds were shared on Freesound by their authors. Full list of clips, authors and licenses.
Thorsten-Voice Dataset 2021.06 emotional, whispered recordings (CC0 1.0). Thorsten Müller and Dominik Kreutz. Used for testing only.
Hebrish Whispered-Sentences (CC BY 4.0). Daniel Rosehill.
Claris Whispered Hindi (MIT). Published by neilshahnms on Hugging Face. Used for testing only.
MUSAN, noise portion (CC BY 4.0 and public domain). D. Snyder, G. Chen, D. Povey, 2015.
Room Impulse Response and Noise Database (Apache 2.0). T. Ko, V. Peddinti, D. Povey, M. Seltzer, S. Khudanpur, ICASSP 2017.
Software
Built with Capacitor, React and onnxruntime-web, each under the MIT license.