Emilia

Emilia is a speech-generation dataset assembled by the Amphion team from in-the-wild audio: podcasts, interviews, talk shows, and similar spontaneous recordings rather than clean studio audiobooks. The base release holds over 101,000 hours across Chinese, English, Japanese, Korean, German, and French, and the expanded Emilia-Large release grows this to more than 216,000 hours, with a large Emilia-YODAS portion drawn from YouTube speech. Each segment ships with a transcript and speaker, language, and quality annotations produced by Emilia-Pipe, the project's open preprocessing pipeline that standardises audio, separates sources, diarises speakers, segments by voice activity, transcribes, and filters.

Access is via Hugging Face. The dataset is gated: you agree to the terms and share contact details before downloading. Audio is distributed as segmented clips (MP3) paired with JSON metadata in WebDataset-style archives, so you can stream or bulk-download and load it with standard audio tooling.

For RAG and AI work this is primarily a training and fine-tuning corpus rather than a text retrieval source. It suits text-to-speech, voice cloning, speech synthesis, and speech-understanding models, and its in-the-wild variety tends to yield higher speaker similarity and naturalness than audiobook-only data. You could also index the transcripts for spoken-language retrieval experiments.

Watch the licence carefully. The original Emilia set is CC BY-NC 4.0, so commercial use is restricted, whereas the newer Emilia-YODAS subset is CC BY 4.0 and does permit commercial use. Terms differ across the compilation, and Amphion does not own the underlying audio copyright, which stays with the original creators, so due diligence is on you. If you need a fully open, commercially usable alternative, consider the People's Speech dataset or Mozilla Common Voice, both of which are likely catalogued here.