Two proven models of my voice (VITS and Tacotron2-DDC), installable as a Python package. Runs without a graphics card on Windows, Mac and Linux – and thanks to the training data, it can also be used commercially without restrictions.
I trained the Coqui models exclusively on my own voice recordings. They contain no other voices and no data under restrictive licenses. This makes them suitable for products, apps and commercial applications as well.
Note: espeak-ng, which is only needed for the Tacotron2-DDC model, is licensed under GPL-3.0. If you plan to distribute or sell an application, check what that means for you – or simply use the VITS model, which works without espeak-ng. This is not legal advice.
Coqui TTS requires Python 3.10 or newer. This command shows which version is installed:
python3 --version On Windows, the command is usually python instead of python3.
Recommended, so that Coqui TTS does not interfere with other Python projects on your computer.
python3 -m venv thorsten-coqui
source thorsten-coqui/bin/activate Windows: python -m venv thorsten-coqui followed by thorsten-coqui\Scripts\activate
The original Coqui project is no longer maintained since Coqui AI shut down. I therefore recommend the actively developed community fork by Idiap. The extras in square brackets matter: “cpu” pulls in PyTorch (nothing works without it), “codec” handles reading and writing audio files, “languages” adds the pronunciation rules for the VITS model and “server” the web interface.
pip install "coqui-tts[cpu,codec,languages,server]" Fork on GitHub ↗ With an NVIDIA graphics card, use “cuda” instead of “cpu”. The download is fairly large (PyTorch) and may take a few minutes. “codec” requires FFmpeg to be installed (macOS: brew install ffmpeg, Debian/Ubuntu: sudo apt install ffmpeg).
The Tacotron2-DDC model uses espeak-ng for pronunciation. If you only want to use the VITS model, you can skip this step.
# Debian / Ubuntu / Raspberry Pi OS
sudo apt install espeak-ng
# macOS (Homebrew)
brew install espeak-ng espeak-ng releases ↗ Windows: download and run the installer (.msi) from the espeak-ng GitHub releases page, then reopen your terminal.
Both models are called via the tts command. The text goes in quotes, and the result is saved as a WAV file in the current folder. The models speak German, so the input text should be German too.
Trained on my Thorsten-22.05-neutral dataset. Works without espeak-ng. Download on first use: about 300 MB.
tts --model_name tts_models/de/thorsten/vits \
--text "Hallo, ich bin Thorsten. Schön, dass du da bist." \
--out_path output.wav Tacotron2 DDC with HifiGAN vocoder, trained on my Thorsten-Dec2021 dataset. Sounds slightly more natural than VITS. Requires espeak-ng. Download on first use: about 500 MB plus vocoder.
tts --model_name tts_models/de/thorsten/tacotron2-DDC \
--text "Hallo, ich bin Thorsten. Schön, dass du da bist." \
--out_path output.wav Starts a small local web server. Then open http://localhost:5002 in your browser, type some text and listen.
tts-server --model_name tts_models/de/thorsten/vits Everything works on the CPU, but longer texts take noticeably longer. If you installed Coqui with the “cuda” extra instead of “cpu”, the GPU takes over:
tts --model_name tts_models/de/thorsten/vits \
--text "Jetzt mit Grafikkarte." \
--out_path output.wav --device cuda TTS is the original package, which is no longer maintained and only installs up to Python 3.11. coqui-tts is the community fork with support for current Python versions. Both packages use the same internal name, so never install both at the same time – remove an old installation first with pip uninstall TTS.
This affects the Tacotron2-DDC model. Install espeak-ng as described in step 4 and reopen your terminal. Alternatively, use the VITS model, which does not need espeak-ng.
The VITS model uses gruut for German pronunciation. gruut comes with the “languages” extra – simply run pip install "coqui-tts[languages]" inside the active virtual environment.
Yes. On first use, Coqui downloads the model once and stores it in your user directory (on Linux, e.g. ~/.local/share/tts). After that, no internet connection is needed.
PyTorch is not installed automatically, only via an extra. Inside the active virtual environment, run pip install "coqui-tts[cpu,codec,languages,server]" (with an NVIDIA graphics card, use “cuda” instead of “cpu”).
Since PyTorch 2.9, Coqui needs the torchcodec package for audio files. Inside the active virtual environment, run pip install "coqui-tts[codec]". If you then get an error about FFmpeg, install FFmpeg (macOS: brew install ffmpeg, Debian/Ubuntu: sudo apt install ffmpeg).
Usually the virtual environment is not active, for example in a new terminal window. Run source thorsten-coqui/bin/activate again (Windows: thorsten-coqui\Scripts\activate).