roest-v3-chatterbox-350m

Røst-v3-chatterbox-350m

This is a Danish open sourced state-of-the-art Text to Speech (TTS) model. It is trained as part of the CoRal project by the Alexandra Institute.

Looking to improve this model or adapt it to your domain? The finetuning framework used to train this model is fully open-sourced at coral_chatterbox.

Examples

<table> <tr> <th>Settings</th> <th>Mic</th> <th>Nic</th> </tr> <tr> <td colspan="3"><em>Dette er en tekst-til-tale model, trænet til at oplæse dansk. Lyt med og bedøm selv, hvordan den håndterer sprogets mange forskellige lyde og rytmer.</em></td> </tr> <tr> <td><b>temp=0.7, top_p=0.95, top_k=600</b></td> <td><audio controls src="https://huggingface.co/CoRal-project/roest-v3-chatterbox-350m/resolve/main/audio_samples/01_mic_00_t0.7_p0.95_k600_r1.2.wav"></audio></td> <td><audio controls src="https://huggingface.co/CoRal-project/roest-v3-chatterbox-350m/resolve/main/audio_samples/01_nic_00_t0.7_p0.95_k600_r1.2.wav"></audio></td> </tr> <tr> <td><b>temp=0.8, top_p=0.95, top_k=1000</b></td> <td><audio controls src="https://huggingface.co/CoRal-project/roest-v3-chatterbox-350m/resolve/main/audio_samples/00_mic_00_t0.8_p0.95_k1000_r1.2.wav"></audio></td> <td><audio controls src="https://huggingface.co/CoRal-project/roest-v3-chatterbox-350m/resolve/main/audio_samples/00_nic_00_t0.8_p0.95_k1000_r1.2.wav"></audio></td> </tr> <tr> <td colspan="3"><em>København er Danmarks hovedstad og ligger på øerne Sjælland og Amager, hvor mange turister besøger de smukke kanaler og historiske bygninger.</em></td> </tr> <tr> <td><b>temp=0.7, top_p=0.95, top_k=600</b></td> <td><audio controls src="https://huggingface.co/CoRal-project/roest-v3-chatterbox-350m/resolve/main/audio_samples/01_mic_01_t0.7_p0.95_k600_r1.2.wav"></audio></td> <td><audio controls src="https://huggingface.co/CoRal-project/roest-v3-chatterbox-350m/resolve/main/audio_samples/01_nic_01_t0.7_p0.95_k600_r1.2.wav"></audio></td> </tr> <tr> <td><b>temp=0.8, top_p=0.95, top_k=1000</b></td> <td><audio controls src="https://huggingface.co/CoRal-project/roest-v3-chatterbox-350m/resolve/main/audio_samples/00_mic_01_t0.8_p0.95_k1000_r1.2.wav"></audio></td> <td><audio controls src="https://huggingface.co/CoRal-project/roest-v3-chatterbox-350m/resolve/main/audio_samples/00_nic_01_t0.8_p0.95_k1000_r1.2.wav"></audio></td> </tr> </table>

Inference Quickstart

Start by installing the required libraries:

$ pip install chatterbox-tts huggingface_hub torchaudio

To just run a simple inference

import torchaudio as ta
from chatterbox.tts_turbo import ChatterboxTurboTTS
from huggingface_hub import snapshot_download
import os

REPO_ID = "CoRal-project/roest-v3-chatterbox-350m"
device = "cuda" # Change to cpu if no GPU is available

# Load the model
model_dir = snapshot_download(
        repo_id=REPO_ID,
        token=os.getenv("HF_TOKEN") or True,
        # Optional: Filter to download only what you need
        allow_patterns=["*.safetensors", "*.json", "*.txt", "*.pt", "*.model"],
    )

model = ChatterboxTurboTTS.from_local(model_dir, device)


text = "Hej, hvordan går det? Jeg er dansk Chatterbox, og taler naturligt dansk."
wav = model.generate(text)
ta.save("test.wav", wav, model.sr)

# If you want to synthesize with a different voice, specify the audio prompt
AUDIO_PROMPT_PATH = "YOUR_FILE.wav"
wav = model.generate(text, audio_prompt_path=AUDIO_PROMPT_PATH)
ta.save("test-2.wav", wav, model.sr)

Model Details

The model is a finetuned variant of Chatterbox-Turbo, one of the leading open-source text-to-speech models. The base Chatterbox-Turbo model uses the GPT-2 Medium architecture and tokenizer, and is trained on more than 500,000 hours of high-quality English speech.

We finetuned the model on over 2000 hours of Danish speech. The model supports zero-shot voice cloning on as little as a 10 seconds audio prompt. It is compatible with the original Chatterbox library, thus being easy to setup and retaining the built-in Watermarked outputs.

The model works very well with the two predefined speakers from the CoRal-tts dataset, Mic and Nic.

Usage Recommendations

We have found the model to work best with the following settings:

Generally, lower temperature and lower top_k gives longer breaks and more stable prosody.

Known Limitations

• The model does not handle longer text inputs natively. Others have implemented streaming for this use case, and we recommend splitting longer texts into sentences.

Evaluation

The model was evaluated using Mean Opinion Score (MOS), achieving a score of 4.01 as rated by a panel of 20 native Danish speakers.

The evaluation used a set of 10 samples for two different speakers, Mic and Nic. Samples were generated with temp=0.7, top_p=0.95, top_k=600. The MOS scale was defined as follows:

Training Notes

Hyperparameters

Training Data

The model was trained on the following datasets:

Creators and Funders

This model has been trained and the model card written by Daniel Christopher Biørrith at Alexandra Institute.

The CoRal project is funded by the Danish Innovation Fund and consists of the following partners:

• Alexandra Institute
• University of Copenhagen
• Agency for Digital Government
• Alvenir
• Corti

Citation

@misc{roest-v3-chatterbox-350m,
  author    = {Daniel Christopher Biørrith, Dan Saattrup Nielsen, Sif Bernstorff Lehmann, Simon Leminen Madsen and Torben Blach},
  title     = {Røst-v3-chatterbox-350m: A Danish state-of-the-art text-to-speech model},
  year      = {2026},
  url       = {https://huggingface.co/CoRal-project/roest-v3-chatterbox-350m},
}