Reproduce “Post-hoc Disentanglement of Textual and Acoustic Features in Pre-trained Speech Encoders” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Hosein Mohebbi, Grzegorz Chrupała, Willem Zuidema, Afra Alishahi, Ivan Titov

The authors separate speech-encoder representations into a text component trained for transcription and a task-specific acoustic component trained for downstream prediction, using information bottlenecks. On Wav2Vec2-Large, the text component achieves WER 0.37 versus 0.56 for original states, while the emotion-trained acoustic component has random transcription performance but retains speaker-identity decodability above chance in probes, not establishing complete separation or privacy.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.