LibriSpeech

LibriSpeech is the reference dataset for English automatic speech recognition (ASR, turning spoken audio into text). It is drawn from LibriVox, the volunteer-read public domain audiobook project, and pairs roughly 1,000 hours of read English speech with sentence-level transcripts, giving a model the sound-to-words mapping it needs to learn. The audiobooks were segmented, aligned to their text, and filtered to keep only recordings that track their transcripts closely.

The corpus lives on OpenSLR, split into named subsets: train-clean-100, train-clean-360, and train-other-500 for training, with matching dev and test sets, where "clean" holds clearer speakers and "other" holds harder, noisier audio. It downloads as a straightforward set of archives, and the audio is stored as FLAC, a lossless open format that common libraries decode without fuss. For RAG in particular, LibriSpeech is less a knowledge base than the front door to one: you use it to train or evaluate the speech-to-text stage that transcribes audio before that text is chunked, embedded, and retrieved.

It is at its best as a benchmark and a training set for English read speech: prototyping a transcription pipeline, fine-tuning or comparing ASR models, or producing the transcripts that feed a voice-driven RAG system. Because almost every ASR paper reports word error rate on test-clean and test-other, your own numbers land in a directly comparable frame.

The caveats follow from what it is. This is read audiobook speech, clearly enunciated and largely free of spontaneous talk, strong accents, crosstalk, or background noise, so a strong score here does not promise strong performance on phone calls or meetings. The licence, by contrast, is clean: CC BY 4.0 allows commercial use as long as you attribute.

Among its neighbours, LibriVox is the raw material one layer up if you want more audio to align yourself, LJ Speech is the single-speaker counterpart for text-to-speech, and larger modern corpora like Common Voice and GigaSpeech supply the spontaneous, accented speech LibriSpeech deliberately leaves out.