VoxCeleb

VoxCeleb is a large speaker-recognition dataset from the Visual Geometry Group at the University of Oxford. It collects more than 100,000 short speech segments from over 7,000 speakers, cut automatically from celebrity interviews posted on YouTube. Because the clips are captured in real settings rather than a studio, they carry background noise, laughter, overlapping voices, and a broad spread of accents, ages, and recording quality, which makes them a realistic test of identifying who is speaking. A larger sequel, VoxCeleb2, extends this to over a million utterances from roughly 6,000 speakers.

You obtain the dataset by requesting access through the VGG project page and agreeing to the terms. Oxford supplies the annotations and speaker labels, while the audio itself is fetched from the original YouTube videos as WAV. Because videos get taken down over time, expect some link rot and missing clips, so budget for gaps and consider the mirrors that the research community maintains for reproducibility.

For a retrieval system, VoxCeleb is not a text corpus, so its place is upstream of RAG rather than inside it. It is built for the audio front end: training or evaluating speaker verification (deciding whether two clips are the same person), speaker diarisation (working out who spoke when), and voice embeddings that let a pipeline separate and label speakers before transcription. The transcripts you produce afterwards are what actually feed a RAG index over meetings, interviews, or podcasts, and better speaker labelling makes that index far more useful.

The licence needs care because it is split. Oxford's annotations are released under CC BY 4.0, which asks only for attribution, but the audio remains under the original YouTube terms, so the underlying content carries no blanket reuse right and commercial use is restricted. Treat VoxCeleb as a research and benchmarking resource, and check the source video terms before shipping anything built directly on the audio.