Instructions to use bosonai/higgs-tts-3-4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use bosonai/higgs-tts-3-4b with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-to-speech", model="bosonai/higgs-tts-3-4b")# Load model directly from transformers import AutoModelForSeq2SeqLM model = AutoModelForSeq2SeqLM.from_pretrained("bosonai/higgs-tts-3-4b", device_map="auto") - Notebooks
- Google Colab
- Kaggle
License question: using Higgs TTS 3 output to train a speech recognition model (non-commercial research)
Hi — I work on low-resource speech: Mandarin, Taiwanese Hokkien (Southern Min), Hakka, and code-switched speech. Emotional and code-switched data for these languages is very scarce, which is why we are looking at synthetic data.
I'd like to understand how you read Section IV(b)(i) before we commit to an approach.
What we would like to do
Fine-tune Higgs TTS 3 for these languages, and use the resulting synthetic speech as training data for an automatic speech recognition model — speech in, text out, used as a research baseline. It would not generate, synthesize, convert, or clone audio, and it would be a standalone model rather than a component of a larger system.
The work is non-commercial academic research. We would want to release the ASR model publicly as a research output.
The question
Section IV(b)(i) prohibits using the Materials or their outputs to train models intended to generate, synthesize, clone, or transform speech or audio, with a carve-out for non-generative classifiers, evaluators, safety filters, watermark or provenance detectors, and benchmark tools. Speech recognition is not named in either the prohibition or the carve-out.
Would you regard training an ASR model of the kind described above as falling inside or outside Section IV(b)(i)?
If it is permitted — how we would release it
We note that the definition of Derivative Works covers models "trained based on or derived from a Model or its outputs". On that reading the ASR model would itself be a Derivative Work. We are happy for it to be released on that basis: under the Boson Higgs TTS 3 Research and Non-Commercial License, with the license and NOTICE file included, the "Derived from Higgs TTS 3, licensed from Boson AI USA, Inc." attribution in the model card, and a model name that does not lead with "Higgs" — rather than under an open-source license. We are not trying to produce an unencumbered model.
If a separate written authorization is needed rather than a reading of the existing license, please let us know what process to follow.
Thanks — and thank you for releasing Higgs.
Hi, thanks for your feedback. Could you please send an email to contact@boson.ai about your request? Even though I've already let people who corresponding to this know. But it will be better if we could do discussion over the emails. Thanks!
Hi, thanks for your feedback. Could you please send an email to contact@boson.ai about your request? Even though I've already let people who corresponding to this know. But it will be better if we could do discussion over the emails. Thanks!
Thanks for letting me know! I just sent the email over to contact@boson.ai with all the details. Looking forward to hearing from your team.