You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This repository contains reconstructable speech latents derived from the Emilia English dataset. Access is restricted to non-commercial research and education. Applicants must already be eligible to use the original Emilia dataset, comply with its Terms of Access and CC BY-NC 4.0, and must not redistribute this derived dataset or use it commercially. Copyright in the underlying recordings remains with the original rights holders. Requests are reviewed manually by the repository owner.

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

Emilia EN Pocket Mimi continuous latents

This gated repository contains the English Emilia training data representation used by the LatentTTS experiments in this project. Audio was encoded offline with the continuous Gaussian Mimi speech VAE used by Pocket TTS. The files are intended to let an authorized researcher reproduce latent-domain training without encoding the source audio again.

Access and licensing

This is a derived representation of the official Emilia Dataset. It is provided only for non-commercial research and educational use under CC BY-NC 4.0.

The Emilia authors state that they do not own the copyright in the underlying audio. Those rights remain with their respective owners. Continuous latents can retain enough information to reconstruct recognizable speech, so these files must not be treated as anonymized data.

Access is manually reviewed. Applicants must:

  1. be eligible to access and use the official Emilia Dataset;
  2. accept the official Emilia Terms of Access and CC BY-NC 4.0;
  3. use the files only for non-commercial research or education;
  4. not redistribute the files or grant access to unapproved third parties;
  5. honor deletion or takedown requests concerning source material.

Approval by this repository owner does not replace the applicant's obligation to check whether their specific use is lawful.

Contents

Emilia_EN/
  latents_mimi_fp32/
    metadata.json
    READY.json
    equivalence_report.json
    train/
    valid/
  manifests/
    PREPARED.json
    ALIGNED.json
    train_aligned.jsonl
    valid_aligned.jsonl
models/
  normalize/
    pocket_mimi/
      emilia_latent_stats.json

The training split contains 18,088,966 aligned utterances (approximately 46,626.18 hours after preprocessing); the validation split contains 3,916 utterances (approximately 10.39 hours).

Latent format

  • codec: Pocket TTS continuous Gaussian Mimi VAE
  • sampling rate: 24,000 Hz
  • latent frame rate: 12.5 Hz
  • waveform samples per latent frame: 1,920
  • latent dimension: 32
  • storage dtype: little-endian float32

The canonical layout, shard sizes, offsets, extraction configuration, weight hashes, normalization information, and validation results are recorded in Emilia_EN/latents_mimi_fp32/metadata.json and the split index files. Training uses train_aligned.jsonl, valid_aligned.jsonl, and emilia_latent_stats.json together with the latent cache.

Manifest path fields preserve paths from the source machine for provenance. Latent-only training identifies samples through the manifest and latent index; those absolute audio paths are not expected to exist on a different server.

Download

After the owner approves an access request:

hf download Gong1212/emilia-token \
  --repo-type dataset \
  --local-dir /path/to/emilia-token

Keep the repository layout unchanged when using it with the accompanying LatentTTS configuration. Verify downloaded files against the hashes recorded in the included metadata before deleting any existing local copy.

Citation

Please cite the official Emilia dataset and Pocket TTS work when this derived data is used. Refer to their official repositories and papers for the current citation text.

Downloads last month
-