Methodology for Obtaining High-Quality Speech Corpora

Speech-based communication between users and machines is a very lively branch of research that covers speech recognition, synthesis, and, generally, natural language processing. Speech corpora are needed for training algorithms for human–machine communication, especially for automatic speech recognition and for speech synthesis. Generative artificial intelligence models also need corpora for training for every language implemented. Therefore, speech corpora are constantly being created. In this paper, we discuss how to create high-quality corpora. The technical parameters of the recordings and audio files are addressed, and a methodology is proposed for planning speech corpus creation with an emphasis on usability. The proposed methodology draws the attention of potential creators of speech corpora to often neglected aspects of the corpus creation process. The criteria for a quality assessment of particular components are also discussed. The author recommends not combining all quality metrics into one (or at least allowing users to adjust particular weights), as different users might be interested in different quality components. The presented guidelines lead to obtaining high-quality corpora that meet the needs of their end users and are easy to use.