Julien Hauret Malo Olivier Thomas Joubaud Christophe Langrenne Sarah Poirée Véronique Zimpfer Éric Bavu

Abstract

Vibravox is a dataset compliant with the General Data Protection Regulation(GDPR) containing audio recordings using five different body-conduction audiosensors : two in-ear microphones, two bone conduction vibration pickups and alaryngophone. The data set also includes audio data from an airborne microphoneused as a reference. The Vibravox corpus contains 38 hours of speech samplesand physiological sounds recorded by 188 participants under different acousticconditions imposed by an high order ambisonics 3D spatializer. Annotationsabout the recording conditions and linguistic transcriptions are also includedin the corpus. We conducted a series of experiments on various speech-relatedtasks, including speech recognition, speech enhancement and speakerverification. These experiments were carried out using state-of-the-art modelsto evaluate and compare their performances on signals captured by the differentaudio sensors offered by the Vibravox dataset, with the aim of gaining a bettergrasp of their individual characteristics.