Dear all,

Over the past four years, researchers worldwide have entrusted the ExELang team led by Alex Cristia with their raw audio recordings (either through bilateral Data Transfer Agreements/Memoranda of Understanding, or through scientific archives), which allowed the development of several computational models. Today, we are writing to you about one called BabyHuBERT. 

BabyHuBERT is a speech representation model, a sort of “jack of all trades” that can then be specialized to solve a specific task, such as voice type classification. The release of this model raised some important ethical questions. To get a non-technical introduction to BabyHuBERT and learn about our approach to its ethics, check out the slides and video recording (also available on YouTube) of a talk Alex gave recently for some researchers who had shared their data with us.

The technical description of the model has been reported on "BabyHuBERT: Multilingual Self-Supervised Learning for Segmenting Speakers in Child-Centered Long-Form Recordings" (https://arxiv.org/abs/2509.15001), recently accepted at Interspeech. For those interested in potentially using the model, we invite you to visit this OSF Link, which includes:

- A PDF with the documents surrounding the decisions taken regarding the release  
- The BabyHuBERT License v1.0
- An introduction to BabyHuBERT, a FAQ, and a process flow document

Please feel free to reach out to Marvin Lavechin <marvinlavechin@gmail.com> if you have any questions!

Best, 

The ExELang team