UncommonVoice: A Crowdsourced Dataset of Dysphonic Speech

Abstract

To facilitate more accessible spoken language technologies and advance the study of dysphonic speech this paper presents UncommonVoice, a freely-available, crowd-sourced speech corpus consisting of 8.5 hours of speech from 57 individuals, 48 of whom have spasmodic dysphonia. The speech material consists of non-words (prolonged vowels, and the prompt for diadochokinetic rate), sentences (randomly selected from TIMIT prompts and the CAPE-V intelligibility analysis), and spontaneous image descriptions. The data was recorded in a crowdsourced manner using a web-based application. This dataset is a fundamental resource for the development of voice-assistive technologies for individuals with dysphonia as well as the enhancement of the accessibility of voice-based technologies (automatic speech recognition, virtual assistants, etc). Research on articulation differences as well as how best to model and represent dysphonic speech will greatly benefit from a free and publicly available dataset of dysphonic speech. The dataset will be made available at http://uncommonvoice. org. In the following sections, we detail the data collection process as well as provide an initial analysis of the speech corpus.