High‑quality French/German zero‑shot TTS built on CosyVoice2. Bilingual adaptation (FR+DE), streaming and non‑streaming synthesis, and a one‑command CLI via the companion PyPI package.

• PyPI: https://pypi.org/project/cosyvoice2-eu/
• Hugging Face: https://hi-paris.github.io/CosyVoice2-EU/
• Live demo: https://hi-paris.github.io/CosyVoice2-EU/

Install the CLI:

pip install cosyvoice2-eu

(or if you use uv: uv add cosyvoice2-eu --frozen, then uv sync)

This will install all necessary dependencies needed for inference with our CosyVoice2‑EU model.

French example:

cosy2-eu \
  --text "Salut ! Je vous présente CosyVoice 2, un système de synthèse vocale très avancé." \
  --prompt path/to/french_ref.wav \
  --out out_fr.wav

German example:

cosy2-eu \
  --text "Hallo! Ich präsentiere CosyVoice 2 – ein fortschrittliches TTS‑System." \
  --prompt path/to/german_ref.wav \
  --out out_de.wav

Notes:

• First run downloads the model checkpoint and caches it locally.
• You can steer style via prompts, e.g. "Speak cheerfully. <|endofprompt|> Hallo! Wie geht es Ihnen heute?".

from cosyvoice2_eu import load
import torchaudio

cosy = load()  # downloads on first use, then reuses the model
wav, sr = cosy.tts(
    text="Salut ! Ceci est une démo.",
    prompt="/path/to/french_ref.wav",
)
torchaudio.save("out.wav", wav, sr)

• cosy_repo/: Local CosyVoice2 code, scripts and notebook for inference and utilities.
  • Notebook: cosy_repo/inference_notebook.ipynb:1 (interactive local synthesis examples)
  • Script: cosy_repo/run_inference.py:1 (command‑line inference with local checkpts)
• dataset: Different scripts to prepare datasets (LibriSpeech, EmoEnet) for training.
• evaluation/: Reproducible evaluation pipeline and configs used in our experiments.
  • Main config: evaluation/eval_config.yaml:1
  • Pipeline: evaluation/run_evaluation_pipeline.py:1
• standalone_infer/: Source of the PyPI package cosyvoice2-eu (packaging‑only).

Here are the links to the baseline models we evaluated against:

• Coqui TTS (XTTS2): https://github.com/coqui-ai/TTS
• Fish‑Speech (OpenAudio S1 / related): https://github.com/fishaudio/fish-speech
• ElevenLabs (Flash V2.5): https://elevenlabs.io (proprietary, closed source)

Our evaluation pipeline supports running baselines independently; see evaluation/run_baseline_evaluation.py:1 and per‑model example configs under evaluation/ (e.g., evaluation/eval_config_coqui.yaml:1, evaluation/eval_config_fishspeech.yaml:1).

• Zero‑shot voice cloning for French and German
• Bilingual FR+DE adaptation on top of CosyVoice2
• Streaming and non‑streaming synthesis
• Simple CLI (cosy2-eu) and local inference scripts
• Interoperable, modular pipeline (text→semantic LM → flow decoder → HiFi‑GAN)

If you’re working in this repository and want to run local inference with your own checkpoints, see:

• cosy_repo/inference_notebook.ipynb:1
• cosy_repo/run_inference.py:1

• Config: evaluation/eval_config.yaml:1 controls language, budgets, metrics, and models.
• Run: python evaluation/run_evaluation_pipeline.py --language fr --hours 100,250,500
• Baselines: enable or run separately via the scripts and example configs under evaluation/.

• Base models: FunAudioLLM/CosyVoice2‑0.5B, Qwen/Qwen3‑0.6B, utter‑project/EuroLLM‑1.7B‑Instruct, Mistral‑7B‑v0.3
• Built on CosyVoice2: https://github.com/FunAudioLLM/CosyVoice
• Hugging Face model: https://huggingface.co/hi-paris/CosyVoice2-0.5B-EU

Please cite or acknowledge CosyVoice2 and the respective base LLMs when using this work.

We prepare FR/DE training and evaluation splits from publicly available datasets. We do not redistribute any third‑party data; please obtain them from the official sources and follow their licenses/terms:

• Multilingual LibriSpeech (MLS): https://huggingface.co/datasets/facebook/multilingual_librispeech
• LAION “LAION’s Got Talent” annotations (used for expressive data): https://huggingface.co/datasets/laion/laions_got_talent_enhanced_flash_annotations_and_long_captions
• M‑AILABS Speech Dataset (OOD evaluation): https://github.com/imdatceleste/m-ailabs-dataset
• Mozilla Common Voice (prompt references): https://commonvoice.mozilla.org

See NOTICE for a concise overview.

• Root license: Apache License 2.0 (see LICENSE).
• CosyVoice2 code in cosy_repo/ is from FunAudioLLM/CosyVoice and remains under Apache‑2.0 (see cosy_repo/LICENSE).
• The packaging project in standalone_infer/ is released under Apache‑2.0 (see standalone_infer/LICENSE).

See NOTICE for a concise overview of third‑party components and their licenses.

• PyPI (CLI): https://pypi.org/project/cosyvoice2-eu/
• Hugging Face (model): https://huggingface.co/hi-paris/CosyVoice2-0.5B-EU
• Live demo: https://hi-paris.github.io/DemoTTS/
• Upstream CosyVoice2: https://github.com/FunAudioLLM/CosyVoice
• Coqui TTS (XTTS2): https://github.com/coqui-ai/TTS
• Fish‑Speech: https://github.com/fishaudio/fish-speech
• ElevenLabs: https://elevenlabs.io

If you use CosyVoice2‑EU in research or products, please cite:

• CosyVoice2‑EU (this work): preprint forthcoming. See CITATION.cff:1 for metadata.
• Upstream CosyVoice2: FunAudioLLM/CosyVoice — please also cite their paper and repo.

BibTeX entries (will be updated soon!)

@misc{horstmann2025cosyvoice2eu,
  title        = {CosyVoice2-EU: Europeanized CosyVoice2 for French and German Zero-Shot TTS},
  author       = {Horstmann, Tim Luka and Ould Ouali, Nassima and Arous, Mohamed Amine and Hussain Sani, Awais and Moulines, Eric},
  year         = {2025},
  note         = {Preprint in preparation},
  howpublished = {\url{https://hi-paris.github.io/CosyVoice2-EU/}}
}

@misc{du2024cosyvoice2,
  title        = {CosyVoice2},
  author       = {Du, et al.},
  year         = {2024},
  howpublished = {\url{https://github.com/FunAudioLLM/CosyVoice}}
}