LJ2_Corpus

Harper Strickland, 2025

Description:

This repository contains a speech corpus in the format of the popular LJ Speech Dataset. It includes 26,200 audio files, with segmentation available in 12-, 24-, and 48-hours. It can be used in combination with LJ Speech to triple its size, compared with LJ Speech, or easily substituted into any model designed to use LJ Speech.

The single speaker's voice is the same as LJ Speech (Linda Johnson, an adult American woman). All texts are public domain nonfiction texts with audio recordings available from LibriVox.org. Selection decisions were made to maximize variety of subjects and authors, and terms over-represented in the texts were downsampled.

Audio file naming convention:

• /wavs1 subdirectory: LJ1##-####.wav
• /wavs2 subdirectory: LJ2##-####.wav
• /wavs3 subdirectory: LJ3##-####.wav
• /wavs4 subdirectory: LJ4##-####.wav

The first 3 digits designate all files partitioned from a single source text chapter. (Files in the LJ Speech Dataset are named LJ0##-####.wav.)

Segments:

• 12-hour: /wavs1
• 24-hour: /wavs1 + /wavs2
• 48-hour: /wavs1 + /wavs2 + /wavs3 + /wavs4

Summary Statistics Compared to LJ Speech:

segment	files	duration	mean (sec)	min (sec)	max (sec)	sources	chapters	authors	max source (hours)
LJ Speech	13,100	23:55	6.57	1.11	10.10	7	50	7	10:34
LJ2 12-hour	6,550	11:57	6.58	1.13	10.10	59	68	59	0:47
LJ2 24-hour	13,100	23:55	6.58	1.11	10.10	61	135	74	1:01
LJ2 48-hour	26,200	47:51	6.58	1.11	10.10	61	288	82	2:53

Files:

• metadata.csv

  Includes file id (same as audio filename without '.wav') and transcript for each selection, in the same format as LJ Speech.

• metadata.zip

  Compressed version of metadata.csv.

• LV_books_data_LJ2.json

  Original source metadata available via LibriVox's API: id number, title, description, language, copyright year, section/chapter count, urls (text source, rss, zip file, project, LibriVox), total time (h:m:s and seconds), authors (name and birth/death years).

• books_enrichment_LJ2.csv

  Additional source metadata: LibriVox id number (primary key), updated title (if incomplete in LibriVox API), LibriVox recordings upload date, count and duration of sections read by Linda Johnson ('LJ').

• sections_details_LJ2.csv

  Additional details about each source chapter represented in the corpus: index (first 3 digits in filename), LibriVox id number (foreign key), section number, authors (some texts are compilations with different authors for each chapter), count of files in corpus, total duration of corpus audio files (minutes), count of files included in '80 Excerpts' Corpus.

• Audio recordings:
  • Audio files can be downloaded from https://huggingface.co/datasets/speakingofdata/LJ2_Corpus/tree/main.