herimor/voxtream2.5-metadata

Dataset Card for VoXtream2.5 training metadata

This repository contains the metadata needed to rebuild the training dataset of VoXtream2.5 TTS model from the original audio. It holds no audio and no audio tokens.

The training dataset contains 54k hours of speech:

• 39k hours from the English part of Emilia dataset (46.8k hours, 18.1M utterances).
• 14.5k hours from a 15.6k-hour subset of HiFiTTS2 dataset (4.7M utterances).

All training entries are 80 seconds long: 1.91M from Emilia and 693k from HiFiTTS2. Each entry concatenates utterances of one speaker, cut at word boundaries and padded with silence at the end (74-75 s of speech on average). Emilia speakers are clustered with speaker embeddings, since its diarization labels are weak; HiFiTTS2 entries follow its speaker and session labels. Sampling rate: 24kHz.

Description

The emilia and hifitts2 directories hold:

• metadata.parquet - One row per utterance: path (relative to the root of the source dataset, without extension), duration (seconds) and text. All other files refer to an utterance by its row in this table.
• alignment.tar.gz - Alignments of every utterance, stored as {parent}/{name}.json and {parent}/{name}.seamless.json, where {parent} is the last directory of path (the Emilia batch, the HiFiTTS2 session):
  • {name}.seamless.json - Word alignment extracted by the Seamless aligner: duration and words (word, start, end, score).
  • {name}.json - Phoneme alignment of the espeak-ng IPA phonemes extracted by ClapIPA forced aligner: words (text, start, end, phonemes as [phoneme, start, end]) and punctuation ([symbol, word index]).
• spk_emb.npy - Speaker embeddings for the first 3 seconds of audio extracted by ReDimNet2 B3 (vb2+vox2+cnc2 lm) model: float16 [num_utterances, 192], L2-normalized, zero rows for unreadable files.
• alignment_filter.npy - Boolean [num_utterances] mask of the utterances that pass the transcript and alignment-quality filters (96.8% of Emilia, 99.0% of HiFiTTS2). Training entries are built only from these.
• crop_plan - The training entries:
  • crop_plan.parquet - One row per utterance segment: cluster_id (entry), order (position in the entry), row (utterance in metadata.parquet), src_start and src_end (sample range taken from the audio) and dst_start (where it is placed in the entry). Samples at 24kHz.
  • crop_clusters.parquet - One row per entry: cluster_id, plan_start and plan_count (its rows of crop_plan.parquet), speech_samples and num_speech_frames.
  • crop_plan_meta.json - Settings of the plan: sampling rate, entry length, pause limits.

valid_phoneme_vocab.json in the root holds the phoneme root counts that define the phoneme vocabulary of the model.

Sources

• Repository: repo
• Model: model

Usage

To download the dataset, use the following code:

from huggingface_hub import snapshot_download

local_dir = snapshot_download('herimor/voxtream2.5-metadata', repo_type='dataset')

Download the audio of the source datasets and place every utterance at {audio_dir}/{path}.mp3, one directory per dataset: the Emilia English mp3 files by batch (EN-B000000/EN_B00000_S00000_W000000.mp3), the HiFiTTS2 files converted to mp3 by speaker and session (100/2315/100_2315_sea_fairies_0812_librivox-18_baum_sea_fairies_17.mp3).

Clone our repo and rebuild the training dataset (Emilia; HiFiTTS2 is the same with its own files):

META=/path/to/voxtream2.5-metadata
mkdir -p /data/alignment && tar xzf $META/emilia/alignment.tar.gz -C /data/alignment

python -m voxtream.utils.dataset.emilia_train_arrays \
    -m $META/emilia/metadata.parquet -p $META/emilia/crop_plan \
    -e $META/emilia/spk_emb.npy -v $META/valid_phoneme_vocab.json \
    -d /data/alignment/emilia -o /data/stores/emilia -nj 32
python -m voxtream.utils.dataset.emilia_mimi_codes \
    -m $META/emilia/metadata.parquet -p $META/emilia/crop_plan -a /data/emilia \
    -o /data/stores/emilia -cb 32 -g 0,1,2,3
python -m voxtream.utils.dataset.silence_tokens -d /data/stores/emilia -g 0

See voxtream/utils/dataset/README.md in the repo for the details and for building training entries with other settings.

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}
}