Thorsten-Voice · Coqui TTS models
German text-to-speech models of my voice ("Thorsten"), trained with Coqui TTS. This repository is an archive copy of the models originally published via the Coqui TTS model catalogue, so they stay available even if the original download location disappears (Coqui AI shut down in early 2024).
🇩🇪 Deutsche Kurzfassung weiter unten.
Models
| Folder | Model | Dataset | Notes |
|---|---|---|---|
vits/ |
VITS (end-to-end) | Thorsten-22.05-neutral | No espeak-ng needed (uses gruut) |
tacotron2-DDC/ |
Tacotron2 DDC | Thorsten-Dec2021-22k | Sounds slightly more natural, needs espeak-ng and the vocoder below |
hifigan_v1/ |
HifiGAN vocoder | Thorsten-Dec2021-22k | Vocoder for tacotron2-DDC |
All models output mono audio at 22,050 Hz and run on CPU; an NVIDIA GPU speeds things up.
Why these models are interesting for commercial use
The models were trained exclusively on my own voice recordings, which I released under CC0 (public domain). No third-party voices or restrictively licensed data were used. The models themselves are released under Apache 2.0, which explicitly permits commercial use.
Note: espeak-ng (only required for tacotron2-DDC) is licensed under GPL-3.0. If you plan to distribute a product,
check what that means for you, or use the vits model, which does not need espeak-ng. This is not legal advice.
Installation
Python 3.10 or newer. Use the maintained community fork of Coqui TTS:
python3 -m venv thorsten-coqui
source thorsten-coqui/bin/activate
pip install "coqui-tts[cpu,codec,languages,server]"
Use cuda instead of cpu for NVIDIA GPUs. The codec extra requires FFmpeg
(macOS: brew install ffmpeg, Debian/Ubuntu: sudo apt install ffmpeg).
For tacotron2-DDC, also install espeak-ng (brew install espeak-ng / sudo apt install espeak-ng).
Usage
Option A: via the Coqui model catalogue (as long as it is available)
tts --model_name tts_models/de/thorsten/vits \
--text "Hallo, ich bin Thorsten. Schön, dass du da bist." \
--out_path output.wav
tts --model_name tts_models/de/thorsten/tacotron2-DDC \
--text "Hallo, ich bin Thorsten. Schön, dass du da bist." \
--out_path output.wav
Option B: from the files in this repository
hf download Thorsten-Voice/Coqui --local-dir thorsten-coqui-models
VITS:
tts --model_path thorsten-coqui-models/vits/model_file.pth \
--config_path thorsten-coqui-models/vits/config.json \
--text "Hallo, ich bin Thorsten. Schön, dass du da bist." \
--out_path output.wav
Tacotron2-DDC with HifiGAN:
tts --model_path thorsten-coqui-models/tacotron2-DDC/model_file.pth \
--config_path thorsten-coqui-models/tacotron2-DDC/config.json \
--vocoder_path thorsten-coqui-models/hifigan_v1/model_file.pth.tar \
--vocoder_config_path thorsten-coqui-models/hifigan_v1/config.json \
--text "Hallo, ich bin Thorsten. Schön, dass du da bist." \
--out_path output.wav
Important for the vocoder: In
hifigan_v1/config.json, the entryaudio.stats_pathmust point toscale_stats.npy. Set it to the absolute path of that file on your machine before running the command above.
Add --device cuda to use an NVIDIA GPU. For a browser interface, run tts-server with the same
parameters and open http://localhost:5002.
Links
- Website & installation guide: https://www.thorsten-voice.de/en/guides/coqui
- About the recordings: https://www.thorsten-voice.de/en/about/recordings
- Coqui TTS (community fork): https://github.com/idiap/coqui-ai-TTS
🇩🇪 Deutsche Kurzfassung
Deutsche Sprachsynthese-Modelle meiner Stimme („Thorsten“), trainiert mit Coqui TTS. Dieses Repository ist eine Sicherungskopie der ursprünglich über den Coqui-Modellkatalog veröffentlichten Modelle (VITS sowie Tacotron2-DDC mit HifiGAN-Vocoder). Das DDC-Modell klingt etwas natürlicher, VITS kommt dafür ohne espeak-ng aus.
Trainiert wurde ausschließlich mit meinen eigenen Sprachaufnahmen unter CC0. Die Modelle stehen unter Apache 2.0 und dürfen daher auch kommerziell genutzt werden.
Ausführliche Installationsanleitung auf Deutsch: https://www.thorsten-voice.de/anleitungen/coqui
Thorsten Müller · https://www.thorsten-voice.de