Heliosoph/LJ-Speech-Dataset

LJSpeech 1.1 — tar.gz re-encoding

A bit-for-bit re-encoding of Keith Ito's LJSpeech 1.1 from its upstream .tar.bz2 container into .tar.gz. The audio, transcripts, and directory layout are unchanged — only the outer compression wrapper differs.

Re-hosted under Heliosoph for ingestion-pipeline stability — Keith Ito's published archive is the authoritative source.

Credit: Keith Ito (LJSpeech 1.1, 2017) — recordings by LibriVox volunteers.

Why a tar.gz mirror?

bzip2 decode is single-threaded and roughly 2–5× slower than gzip on modern hardware. For tools that materialize the archive end-to-end before reading (most database / dataset ingestion pipelines do this), the bz2 wrapper dominates install time. The contained WAV bytes are already incompressible, so gzip is no worse than bzip2 on size — only on speed.

If your tooling unwraps bz2 natively at acceptable speed, prefer the upstream archive. If you're hitting a 10+ minute decode and want it under 2, this mirror exists for you.

What this repo contains

LJSpeech-1.1/
├── README                # upstream Keith Ito notes (preserved verbatim)
├── metadata.csv          # 13,100 rows: <wav-id>|<original>|<normalized>
└── wavs/
    ├── LJ001-0001.wav    # 22.05 kHz, mono, 16-bit
    ├── LJ001-0002.wav
    └── …                 # 13,100 clips total, ~24 hours

Archive: LJSpeech-1.1.tar.gz — roughly the same compressed size as the upstream .tar.bz2 (~2.6 GB).

How to use

Stream the archive directly with any tar reader:

import tarfile, soundfile as sf, io

with tarfile.open("LJSpeech-1.1.tar.gz", "r:gz") as tar:
    for member in tar:
        if not member.name.endswith(".wav"):
            continue
        with tar.extractfile(member) as f:
            audio, sr = sf.read(io.BytesIO(f.read()))
            # audio: float64 [-1, 1], sr: 22050
            ...

Or extract once and read flat:

tar -xzf LJSpeech-1.1.tar.gz
ls LJSpeech-1.1/wavs/ | head

Audio specs

	Spec
Format	RIFF WAV (PCM)
Sample rate	22050 Hz
Channels	1 (mono)
Bit depth	16
Total clips	13,100
Total duration	~24 hours
Per-clip duration	~1.1 – 10.1 sec (median ~6 sec)
Speaker	Single US English female

Transcripts

metadata.csv is a pipe-delimited file with three fields per row:

LJ001-0001|Printing, in the only sense with which we are at present concerned,…|Printing, in the only sense with which we are at present concerned,…
LJ001-0002|in being comparatively modern.|in being comparatively modern.

• Field 1: WAV identifier (matches the filename without extension).
• Field 2: original text as published.
• Field 3: normalized text (numbers expanded, abbreviations spelled out, punctuation simplified).

The same metadata.csv ships in the upstream archive — no edits.

When to pick LJSpeech

• Single-speaker TTS prototyping: the canonical small-scale training set.
• Vocoder development: clean studio-quality input, consistent recording chain.
• ASR sanity checks: one voice, no channel variation — easy to hear what your model is doing.

For multi-speaker variation, noisy conditions, or published WER benchmarks, reach for LibriSpeech instead.

License

Public domain. The source texts are 19th- and early-20th-century non-fiction passages whose copyrights have expired (United States); the audio was recorded by LibriVox volunteers and dedicated to the public domain by Keith Ito. No attribution required, but citing the upstream dataset is good practice — see keithito.com/LJ-Speech-Dataset/.

The HuggingFace metadata tags this cc0-1.0 as the closest formal SPDX identifier for HF's taxonomy; the upstream dedication is plain "public domain" rather than formal CC0, but the effective rights are equivalent.

• Upstream: keithito.com/LJ-Speech-Dataset
• Upstream archive (bz2): data.keithito.com/data/speech/LJSpeech-1.1.tar.bz2