The LJSpeech Dataset

Datasheet: The LJ Speech Dataset

Overview

Dataset Description

The LJ Speech Dataset is a public domain, single-speaker speech dataset consisting of 13,100 short audio clips of a single female speaker reading passages from 7 non-fiction books. A text transcription is provided for each clip. Clips range from approximately 1 to 10 seconds in length, with a total duration of roughly 24 hours.

The dataset is widely used for training and benchmarking text-to-speech (TTS) synthesis systems. Its high audio quality, clean segmentation, and aligned transcripts make it a standard benchmark in neural speech synthesis research.

Language

All transcriptions and audio recordings are in English.

Creation

Audio Recordings

The audio was recorded in 2016–2017 by Linda Johnson as part of the LibriVox project. LibriVox is a volunteer-run initiative that produces free public domain audiobooks. The original LibriVox recordings were distributed as 128 kbps MP3 files, and as a result some clips may contain minor artifacts introduced by MP3 encoding.

Audio clips were segmented automatically based on silences in the recording. Clip boundaries generally align with sentence or clause boundaries, though not always.

Transcription & Annotation

Alignment and annotation were performed manually by Keith Ito. A quality assurance pass was done to ensure that text transcriptions accurately matched the spoken audio.

Source Texts

The dataset consists of excerpts from the following 7 public domain works (published 1884–1964):

Dataset Statistics

File Format

• Audio: Single-channel 16-bit PCM WAV files at a sample rate of 22,050 Hz

• Meta metadata.csv (pipe-delimited |), with three fields per record:

  1. ID — name of the corresponding .wav file

  2. Transcription — words spoken by the reader (UTF-8)

  3. Normalized Transcription — transcription with numbers, ordinals, and monetary units expanded into full words (UTF-8)

Intended Use

• Text-to-speech (TTS) synthesis model training and evaluation

• Neural speech synthesis benchmarking

• Voice modeling, prosody, and pronunciation research

• Educational and prototyping use cases for speech generation

Known Limitations & Notes

• Audio was originally encoded as 128 kbps MP3 by LibriVox; some clips may contain MP3 compression artifacts.

• Clip boundaries are based on silence detection and do not always align precisely with sentence boundaries.

• 19 transcriptions contain non-ASCII characters (e.g., raison d'être).

• The dataset contains a standard set of abbreviations (e.g., Mr., Dr., St.) with recommended expansions. Note that there is no standard expansion for Mrs.

License

This dataset is in the public domain in the United States, and most likely in other countries as well. There are no restrictions on its use.

"All text, audio, and annotations are in the public domain. We request that you use this dataset for good and not evil."

For more information, see: https://librivox.org/pages/public-domain

Creators

Citation

There is no requirement to cite this work (as it is in the public domain), but if you wish to cite it in a publication, the authors suggest the following:

@misc{ljspeech17,
  author       = {Keith Ito and Linda Johnson},
  title        = {The LJ Speech Dataset},
  howpublished = {\url{https://keithito.com/LJ-Speech-Dataset/}},
  year         = {2017}
}

Alternatively, you may link directly to: https://keithito.com/LJ-Speech-Dataset/

Changelog