LJ Speech: 13,100 Single-Speaker Audio Clips with Transcriptions

LJ Speech contains 13,100 short audio clips of a single speaker reading from seven non-fiction books, totaling approximately 24 hours of English speech. Released by Keith Ito, the dataset provides expert-generated transcriptions for every recording to support speech synthesis and recognition tasks.

Use Cases

• Training Text-to-Speech (TTS) models using the transcription field and audio clips
• Fine-tuning Automatic Speech Recognition (ASR) systems on the 13,100 audio-text pairs
• Acoustic modeling and prosody analysis of a consistent single speaker

Strengths

• 13,100 expert-annotated records
• 24 hours of high-quality single-speaker audio
• Public domain license for unrestricted commercial and research use

Limitations

• Single-speaker bias prevents models from generalizing to multiple voices or accents
• Vocabulary is limited to the domain of 7 non-fiction books
• Audio requires manual conversion to float32 for most modern ML pipelines

Provenance

Source

Keith Ito

Collection Method

Recorded readings of seven non-fiction books with expert-generated transcriptions

Freshness

Last updated August 2024; the source audio and text from non-fiction books are static.

Geography

United States