test123124125/v5

LibriTTS research training subset

This repository contains unchanged LibriTTS waveform bytes, normalized transcripts, a versioned research split, and derived acoustic training features. It contains no private target-speaker recordings or application implementation source.

Source and attribution

The corpus is provided by Google LLC under CC BY 4.0. See the original OpenSLR release and licenses/LibriTTS-LICENSE.txt. Credit: Heiga Zen, Viet Dang, Rob Clark, Yu Zhang, Ron J. Weiss, Ye Jia, Zhifeng Chen and Yonghui Wu, LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech (2019). This is a repackaged research subset, not a release by the original authors.

The source is train-clean-100. Only recordings from 1 through 30 seconds in the project's existing index are included. Audio is unchanged. Normalized transcripts come from the source corpus. Repackaging adds per-sample SHA-256, TAR offsets, and a project-specific speaker-disjoint train/validation/test assignment. These are not the original publisher's development/test sets.

Indexed release: libritts-20261018-v1

Split	Samples	Hours	Speakers
test	1,979	3.5345	21
train	28,506	47.7166	217
validation	1,279	2.1312	9

The release manifests appear under releases/libritts-20261018-v1/manifests/ after all shards have passed full readback verification. foundation.json is the training index. all-training.json contains the same general-corpus rows. Integration-test releases under workflow-* contain small subsets of these recordings; they are not additional training speakers.

Reading the data

Shards under releases/<release>/shards/ are uncompressed TAR archives. Matching .wav and .json members share a stem. The JSON includes the normalized transcript, speaker/session/source identifiers, split, duration and sample SHA-256. The manifest's asset includes the archive key and SHA-256, byte length, member name, member length and exact waveform byte offset.

Pin a full repository commit and the manifest SHA-256. Request Range: bytes=<offset>-<offset+memberBytes-1> from the shard's resolve URL at that commit, require the expected content range, and verify the resulting sample SHA-256 before decoding. This permits bounded per-recording access without a full repository snapshot. Ordinary TAR readers can also read whole shards.

Derived arrays under derived/<feature-recipe-hash>/ are compressed NPZ files. They contain an identity record and numeric arrays, use allow_pickle=False, and bind the originating audio hash, feature settings and encoder/code identities. They are caches, not independent speech recordings.

Limitations

These are English audiobook recordings. The subset does not establish natural conversational delivery or target-speaker fidelity. The reviewed flag is inherited from corpus ingestion and does not claim a new human review of every waveform. Training execution, complete checkpoint recovery and voice-quality acceptance are distinct results.