Generating long-form speech like audiobooks or podcasts requires keeping voice and style consistent. Existing public datasets only cover short clips or two speakers.
Emilia2 is a massive multilingual dataset from real-world recordings. It contains over two hundred thousand hours of speech spanning forty one languages with transcripts, speaker identities, and pauses. It includes natural-language descriptions of expressive style and tags for nonverbal sounds like laughter.
Models trained on Emilia2 achieved the lowest error rates on standard benchmarks: a character error rate of five point two seven percent and a word error rate of nine point nine three percent. They did so using far smaller models than competitors, proving that data quality is the key to practical long-form speech generation.