eigenhand's picture
Articulino-DE-EN-Realtime-Stream-TTS 1.0 (private preview)
0e3f9f5 verified
Raw History Blame Contribute Delete
2.16 kB
Articulino-DE-EN-Realtime-Stream-TTS
Copyright 2026 eigenhand
Licensed under the Apache License, Version 2.0 (see LICENSE).
---------------------------------------------------------------------------
Training data
---------------------------------------------------------------------------
The model was trained on the following corpora. The first three are licensed
under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/), which
requires this attribution. Keep this NOTICE when you redistribute the model.
- Emilia-YODAS, German subset. Haorui He et al., Amphion, "Emilia: An
Extensive, Multilingual, and Diverse Speech Dataset for Large-Scale Speech
Generation" (2024); built on YODAS (Xinjian Li et al., ESPnet / WAVLab).
CC BY 4.0. https://huggingface.co/datasets/amphion/Emilia-Dataset
Only the YODAS part was used; the rest of Emilia (CC BY-NC 4.0) was not.
- Hi-Fi TTS 2, selected English chapters. Ryan Langman et al., NVIDIA.
CC BY 4.0. https://huggingface.co/datasets/nvidia/hifitts-2
- CML-TTS, German. Frederico S. Oliveira et al. CC BY 4.0.
https://www.openslr.org/146/
Also used, without an attribution requirement:
- Mozilla Common Voice 27.0, German. CC0 1.0. https://commonvoice.mozilla.org
- Thorsten-Voice datasets, Thorsten Mueller. CC0 1.0. https://www.thorsten-voice.de
- HUI-Audio-Corpus-German, iisys (Hof University). Audio from LibriVox
(public domain). https://github.com/iisys-hof/HUI-Audio-Corpus-German
No changes were made to the content of these corpora other than resampling to
24 kHz, cutting, filtering and text normalization for training.
---------------------------------------------------------------------------
Design credit
---------------------------------------------------------------------------
The overall recipe (letter-level text, duration predictor, flow-matching DiT
on 25 Hz latents, 16-step teacher distilled with DMD2 into a 4-step student,
separate latent decoder) follows the model card of EMA Lightning by Canberk
Aslan (Apache License 2.0), https://huggingface.co/canberkkkkkk/ema-lightning.
No code, weights, data or samples of EMA Lightning are included or were used.