Request access to tokenized OmniVoice training shards

Access is gated because these shards are derived from speech data. Only request access if your intended use follows the original dataset licenses and speaker-rights requirements.

By requesting access, you agree not to redistribute the files, attempt speaker re-identification, or use the data for unauthorized voice impersonation.

Log in or Sign Up to review the conditions and access this dataset content.

OmniVoice Higgs Tokenized V2

This repository contains versioned, tokenized OmniVoice training shards from quality-audited speech sources. It is intended for OmniVoice-style training and evaluation, not for direct audio playback.

Access

This dataset repository is intended to be public but gated with manual approval mode. Gating controls access to files, but it does not replace the source licenses or consent requirements.

Source Dataset

  • darus/CNVVE; langswap/dialogs-ru-emotional-conversations; yfish/WESR-Bench

Contents

  • data/darus__CNVVE/train/data.lst: 2 examples, 0.000 hours.
  • data/langswap__dialogs-ru-emotional-conversations/dev/data.lst: 92 examples, 0.155 hours.
  • data/langswap__dialogs-ru-emotional-conversations/train/data.lst: 8,735 examples, 14.937 hours.
  • data/yfish__WESR-Bench__en/dev/data.lst: 1 examples, 0.003 hours.
  • data/yfish__WESR-Bench__en/train/data.lst: 143 examples, 0.290 hours.
  • data/yfish__WESR-Bench__zh/dev/data.lst: 1 examples, 0.003 hours.
  • data/yfish__WESR-Bench__zh/train/data.lst: 212 examples, 0.584 hours.
  • */audios/shard-*.tar: tokenized audio code shards.
  • */txts/shard-*.jsonl: text and metadata shards.
  • tokenized_dataset_summary.json: generated upload summary.

The uploaded data.lst files are rewritten to use repo-relative paths. The original pod paths are not required after download.

Summary

  • Approximate examples in train + dev: 9,186
  • Approximate hours in train + dev: 15.97
manifest shard rows examples hours missing refs
data/darus__CNVVE/train/data.lst 1 2 0.00 0
data/langswap__dialogs-ru-emotional-conversations/dev/data.lst 1 92 0.15 0
data/langswap__dialogs-ru-emotional-conversations/train/data.lst 9 8,735 14.94 0
data/yfish__WESR-Bench__en/dev/data.lst 1 1 0.00 0
data/yfish__WESR-Bench__en/train/data.lst 1 143 0.29 0
data/yfish__WESR-Bench__zh/dev/data.lst 1 1 0.00 0
data/yfish__WESR-Bench__zh/train/data.lst 1 212 0.58 0

Training Streams

Training stream metadata was not found in the run config.

Known Tokenization Skips

Raw errors.jsonl files are not uploaded by default. Counts are included here for traceability.

errors file rows
data/darus__CNVVE/train/errors.jsonl 0
data/langswap__dialogs-ru-emotional-conversations/dev/errors.jsonl 0
data/langswap__dialogs-ru-emotional-conversations/train/errors.jsonl 0
data/yfish__WESR-Bench__en/dev/errors.jsonl 0
data/yfish__WESR-Bench__en/train/errors.jsonl 0
data/yfish__WESR-Bench__zh/dev/errors.jsonl 0
data/yfish__WESR-Bench__zh/train/errors.jsonl 0

Usage

After downloading or streaming the repository, point the OmniVoice training data config at the relative manifests:

train/data.lst
dev/data.lst

These shards are tied to the OmniVoice/Higgs audio token format used during preparation.

Safety and License Notes

  • Do not redistribute the files outside the gated repository.
  • Do not use the dataset for unauthorized voice cloning, speaker impersonation, or speaker re-identification.
  • Make sure each downstream use complies with the original speech dataset licenses and any consent constraints.
  • Gating is access control, not legal clearance.

Generation

This README was generated by omnivoice/push_tokenized_to_hf.py. Target repository: instinct1912/omnivoice-higgs-tokenized-v2.

TTS-AGI Emilia Yodas ScribeEvents (JA/EN/DE)

Append-only verified tranche: 14,278 examples across 6 train/dev manifests. Source revision: e45dc6334d5396fc2e62b7210369df0d98e4d2e1. All rows passed the staged eight-codebook, holdout, provenance, and tag-contract audits before publication.

Bookbot NonverbalTTS filtered cleaned

Append-only verified tranche: 4,451 examples across 2 train/dev manifests. Source revision: 4bbed9d0abece3bd7e64fb9583429c14e59cae52. All rows passed the staged eight-codebook, holdout, provenance, and tag-contract audits before publication.

High quality English EARS and Expresso

Append-only verified tranche: 8,004 examples across 6 train/dev manifests. Source revision: EARS@613935413a2eac115d191fffc5c74b34ca7cc884+Expresso@9fb79a189698de3255eff48edd2bc0d9e487adc0+qa43@5d60517ffc5a7db3ff35a415558f37d0d5446ab383be178bdafe3c33dbc7e79e. All rows passed the staged eight-codebook, holdout, provenance, and tag-contract audits before publication.

Downloads last month
7