Yoruba speech corpus expands language technology resources

Researchers have released the \`I}r\`o}y\`i}nSpeech text corpus, a collection of 24,905 hand-verified Yoruba sentences, to support speech and language technologies.

The corpus contains 275,897 words and builds on a 42-hour audio recording distributed since 2024 by ELRA, a language resource distribution center. Approximately 11,000 sentences came from news sources while the remaining were newly written to move beyond religious texts common in existing Yoruba language datasets.

Each sentence underwent manual review for tonal accuracy and was adjusted for clear pronunciation and a neutral style. The creators also adapted names and places to Yoruba forms. Preparing the corpus for release revealed widespread inconsistencies in how Yoruba characters are encoded in computers, affecting over 60% of the lines, which the team has now corrected. The \`I}r\`o}y\`i}nSpeech corpus provides a valuable resource for developing tools like text-to-speech systems, improving speech recognition, and furthering research into Yoruba orthography.

It also offers a tested set of prompts for future audio recordings.