The SWARA speech corpus: A large parallel Romanian read speech dataset

Abstract

This paper introduces one of the largest Romanian speech datasets freely available for both academic and commercial use. The dataset comprises speech data recorded over the last year from 12 speakers, along with 5 other speakers previously recorded in a separate environment. The data was manually segmented at utterance-level and semi-automatically labelled at phone-level. The resulting corpus amounts to approximately 21 hours of high-quality read speech data, split into over 19,000 utterances. The speakers read between 921 and 1493 utterances each. 880 utterances are common to all speakers and add up to over 16 hours of parallel data. We present the steps of performing the recordings and data segmentation, as well as a first use of this corpus in the context of synthetic voice development.