Model Card for VoXtream2.5

VoXtream2.5 is a zero-shot full-stream TTS model with dynamic speaking-rate control that can be updated mid-utterance on the fly. It builds on VoXtream2, introduced in the paper VoXtream2: Full-stream TTS with dynamic speaking rate control, with:

• Hybrid depth transformer head: 32 Mimi codebooks are generated in 4 steps (3 AR and a single Mixture of Gaussians step), which makes the model run ~10x faster than real-time on an RTX 3090.

• CPU inference: GGUF (int8) model runs in real-time on a single CPU core on llama.cpp.

• Long-form generation: Generation runs for up to 5 minutes with a WER below 2% measured on the LibriSpeech-long benchmark.

Developed by: Nikita Torgashov, Gustav Eje Henter, Gabriel Skantze

Model Sources

• Repository: https://github.com/herimor/voxtream
• Blogpost: https://herimor.github.io/voxtream2.5
• Live Demo: https://huggingface.co/spaces/herimor/voxtream2.5

Files

• model.safetensors, config.json: the model weights and config.
• phoneme_vocab.json: the phoneme vocabulary the training data was built with.
• *-q8_0.gguf, voxtream_gguf.json: the temporal transformer, depth transformer, MoG head and Mimi decoder transformer for llama.cpp.

Get started

Installation

eSpeak NG phonemizer

# For Debian-like distribution (e.g. Ubuntu, Mint, etc.)
apt-get install espeak-ng
# For RedHat-like distribution (e.g. CentOS, Fedora, etc.) 
yum install espeak-ng
# For MacOS
brew install espeak-ng

Pip package

pip install "voxtream>=0.2.6"

Usage

• Prompt audio: a file containing 3-10 seconds of the target voice. The maximum supported length is 20 seconds (longer audio will be trimmed).
• Text: What you want the model to say.
• Speaking rate (optional): target speaking rate in syllables per second.

Output streaming

voxtream \
    --prompt-audio assets/audio/english_male.wav \
    --text "In general, however, some method is then needed to evaluate each approximation." \
    --output "output_stream.wav"

Full streaming (slow speech, 2 syllables per second)

voxtream \
    --prompt-audio assets/audio/english_female.wav \
    --text "Staff do not always do enough to prevent violence." \
    --output "full_stream_2sps.wav" \
    --full-stream \
    --spk-rate 2.0

CPU inference with llama.cpp (int8)

voxtream \
    --prompt-audio assets/audio/english_male.wav \
    --text "In general, however, some method is then needed to evaluate each approximation." \
    --output "output_cpu.wav" \
    --threads 2

• Note: Initial run may take some time to download model weights and warmup model graph.

Training Data

The model was trained on Emilia and HiFiTTS2 datasets. For more details, please check our paper.

Out-of-Scope Use

Any organization or individual is prohibited from using any technology mentioned in this paper to generate someone's speech without his/her consent, including but not limited to government leaders, political figures, and celebrities. If you do not comply with this item, you could be in violation of copyright laws.

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}
}