LJ Speech

LJ Speech is a small, clean corpus made for training and benchmarking text-to-speech (TTS, generating spoken audio from written text). It gathers 13,100 short clips of a single English-speaking woman reading passages from seven non-fiction books, with a transcript for every clip and durations of one to 10 seconds. The recordings were made under consistent conditions, and that steadiness is exactly why the dataset became a fixture of TTS research.

It ships as a single archive of about 24 hours of audio: WAV files at a fixed sample rate plus a metadata CSV mapping each file to its raw and normalised transcript. There is nothing to align or clean before you start, which is why most TTS toolkits include a recipe that reads LJ Speech directly. In a RAG setting its place is at the output end rather than in retrieval: it trains the voice that reads answers back to a user, so it pairs naturally with a text-to-speech layer sitting after your retriever.

It is the default choice for single-speaker TTS: prototyping a voice, reproducing a published model, teaching yourself the pipeline, or benchmarking vocoder and acoustic-model quality. Because so much research reports on it, results are easy to line up against one another, and its modest size means you can run experiments quickly on a single machine.

The limits are the flip side of that convenience. It is one speaker, one accent, and a fairly formal reading style, so a model trained on it alone will not give you multiple voices, expressive delivery, or robustness to conversational speech. The licence, happily, could not be simpler: the whole dataset is in the public domain, so you can use it for any purpose, commercial products included, with no restrictions and no attribution.

Among its neighbours, reach for VCTK or LibriTTS when you need many speakers, and LibriSpeech when your task is recognition rather than synthesis. LJ Speech stays the smallest, simplest starting point when you just want one clean voice.