Let’s study learning trajectories in self-supervised speech models! 🔊 Do they reflect the hierarchical organization of spoken language? We have analyzed a lot of training checkpoints to find out 🌠 Preprint: arxiv.org/abs/2604.02043 ⬇️
Tracking the emergence of linguistic structure in self-supervised models learning from speech arxiv.org
AI Weekly's analysis
→
- Researchers probed six Wav2Vec2 and HuBERT models trained on spoken Dutch across layers and intermediate training checkpoints.
- Different levels of linguistic structure showed distinct layerwise patterns and learning trajectories inside the same speech models.
- Higher-order prediction tasks using iteratively refined pseudo-labels induced greater parallelism across the models' internal layers.
Read full analysis →
This should be a reference to a peer commentary (e.g. de Heer Kloots et al. refers to arxiv.org/abs/2512.14506) which aren’t published by BBS yet (only some of them are preprinted by the commentary authors atm).
Linguists should learn to love speech-based deep learning models arxiv.org
Now out in BBS, as commentary on @futrell.bsky.social & @kmahowald.bsky.social's "How linguistics learned to stop worrying and love the language models"! Humans learn much of spoken language structure from speech (not text), & we can study models that do the same. www.cambridg…
Linguists should learn to love speech-based deep learning models | Behavioral and Brain Sciences | Cambridge Core cambridge.org
↻
Marianne de Heer Kloots reposted
@avsp.bsky.social
Linguists should learn to love speech-based deep learning models www.cambridge.org/core/journal... "Once the bottleneck of text can be replaced...modelling more human-like linguistic processes..[& handle] the vast majority of local & regional language varieties that are rarely…
Linguists should learn to love speech-based deep learning models | Behavioral and Brain Sciences | Cambridge Core cambridge.org
View on Bluesky →
↻
Marianne de Heer Kloots reposted
@avsp.bsky.social
Linguists should learn to love speech-based deep learning models www.cambridge.org/core/journal... "Once the bottleneck of text can be replaced...modelling more human-like linguistic processes..[& handle] the vast majority of local & regional language varieties that are rarely…
cambridge.org
View on Bluesky →
All related commentaries are out now! www.cambridge.org/core/journal...
How linguistics learned to stop worrying and love the language models | Behavioral and Brain Sciences | Cambridge Core cambridge.org
