Descriptor: Enhancing conversations for the hearing impaired in the 9th computational hearing in multisource environments challenge (CHiME9 ECHI)

This article presents a dataset for enhancing conversations for the hearing impaired (ECHI). This is a publicly available dataset comprised of recordings of four-party conversations in a noisy, cafeteria-style environment, close-talk reference speech signals, clean voice samples from the participants, and motion tracking data. This dataset was created to be used as part of the Computational Hearing in Multisource Environments (CHiME) 9 challenge, task 2 (ECHI), which aims at developing systems to extract and enhance speech from the conversation partners, i.e. removing the background noise using audio from microphone signals of hearing aids or project aria smart glasses. This dataset can be used to both train and evaluate the performance of neural-network or signal-processing-based speech enhancement algorithms, and it was created due to the lack of datasets of conversational speech in noisy environments with clean reference audio. Groups of four people were seated around a table, surrounded by loudspeakers which played interfering speech, cafeteria-style sounds, and ambient background noise. Participants were instructed to have conversations guided by provided conversation prompts. Each participant also recorded a reading of the first paragraph of the rainbow passage as a clean sample of their voice, which may be used by a speech enhancement system to extract their speech from the noisy mixture. This dataset contains 29 h of audio from 190 participants. IEEE SOCIETY/COUNCIL Signal Processing Society (SPS) DATA TYPE/LOCATION Audio, Motion Tracking; University of Sheffield, Sheffield, U.K. DATA DOI/PID 10.57967/hf/5924