LibriHeavy TTS 3: A 50,000-Hour Speech Corpus for Text-to-Speech Training

LibriHeavy TTS 3 is an improved version of the LibriHeavy dataset, designed specifically for text-to-speech training quality. It is built on a 50,000-hour labeled ASR corpus derived from LibriLight, with audio encoded using the Opus 68kbps codec. The dataset, authored by brthor and last updated in April 2026, focuses on providing better audio and text supervision quality.

Use Cases

• Training text-to-speech models based on the high-fidelity audio and text supervision mentioned in the description.
• Fine-tuning speech synthesis systems based on the large-scale, 50,000-hour labeled corpus.
• Improving audio quality and naturalness in TTS outputs based on the use of the Opus codec for audio encoding.
• Researching speech representation learning based on the audio-text alignment and punctuation/casing context provided.

Strengths

• Contains 50,000 hours of audio data, providing a substantial resource for model training.
• Audio files are encoded with the Opus 68kbps codec, which is designed to retain quality while reducing file size.
• Includes punctuation, casing, and context labels for the text, which can improve supervision quality.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is marked as a work in progress (WIP) and not yet released, indicating it may be incomplete or unstable.

Provenance

Source

Built on top of mythicinfinity/libriheavy, which is a labeled version of LibriLight derived from LibriVox recordings.

Collection Method

Likely involves processing and labeling audio from the LibriVox public domain audiobook project.

Time Range

null

Freshness

Last updated 2026-04-27 16:43:06; freshness should be verified.

Geography

null