HiFiTTS-2 source audio archive

This hysi-lab mirror preserves NVIDIA's pinned HiFiTTS-2 metadata and the original LibriVox chapter audio referenced by its 22 kHz inventory. The 44 kHz metadata is also retained; it describes a subset and must not be counted as additional hours. Sources: https://huggingface.co/datasets/nvidia/hifitts-2 and https://librivox.org/. Please credit Ryan Langman et al., HiFiTTS-2: A Large-Scale High Bandwidth Speech Dataset (Interspeech 2025), NVIDIA, and the original LibriVox readers/books.

Audio acquisition is in progress. Tar shards under archives/ (including hash subdirectories) contain original compressed chapter bytes at audio/<SHA256-of-source-URL>.media and MANIFEST.jsonl, which maps each member to its original URL, byte digest and measured duration. Each sidecar manifest identifies the exact tar checksum. This is source audio, not pre-cut utterance FLACs. Use the original chapter offsets and utterance manifests to reconstruct the desired 22.05 kHz or 44.1 kHz split. No ASR, speaker separation, resampling or transcript changes are applied during acquisition.

acquisition_status.json reports verified uploads and failures. Upstream nominal hours and metadata completion are not delivered-audio counts. A chapter may be unavailable or may fail a duration/decode check; failures remain recorded. The NVIDIA metadata is CC BY 4.0. LibriVox source recordings are public domain in the USA; retain source attribution and consult original book/recording details.