herimor/voxtream2-train

Model Card for VoXtream2 training dataset

This repository contains a training dataset for VoXtream2 TTS model.

The dataset contains 40k hours:

• 30k hours subset from Emilia dataset.
• 10k hours subset from HiFiTTS2 dataset (22 kHz).

All utterances are 55 seconds long. We concatenated multiple utterances within the same speaker and padded shorter clips with silence. Sampling rate: 24kHz.

Description

• mimi_codes - Tokens extracted by the Mimi audio codec (16 codebooks).
• phone_emb_indices - Alignment of phoneme tokens to Mimi audio frames extracted by ClapIPA forced aligner.
• phone_tokens - IPA Phoneme tokens extracted with espeak-ng phonemizer.
• sem_label_shifts - Monotonic phoneme alignment labels.
• punctuation - Punctuation tokens.
• spk_templates - Speaker templates for the first 3 seconds of audio extracted by ReDimNet model.

Sources

• Repository: repo
• Paper: paper
• Demo: demo

Usage

To download the dataset, use the following code:

from huggingface_hub import snapshot_download

local_dir = snapshot_download('herimor/voxtream2-train', repo_type='dataset')

Clone our repo and follow the instructions in the README file.

Citation

@inproceedings{torgashov2026voxtream,
  title={Vo{X}tream: Full-Stream Text-to-Speech with Extremely Low Latency},
  author={Torgashov, Nikita and Henter, Gustav Eje and Skantze, Gabriel},
  booktitle={Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2026},
  note={to appear},
  url={https://arxiv.org/abs/2509.15969}
}

@article{torgashov2026voxtream2,
  author    = {Torgashov, Nikita and Henter, Gustav Eje and Skantze, Gabriel},
  title     = {Vo{X}tream2: Full-stream TTS with dynamic speaking rate control},
  journal   = {arXiv:2603.13518},
  year      = {2026}
}