Private Speech Classification without Collapse: Stabilized DP Training and Offline Distillation

Yadi Wen † \dagger, Tianxin Li † \dagger, Enji Liang, Rong Du ∗ *, Yue Fu · 2026

Read Private Speech Classification without Collapse: Stabilized DP Training and Offline Distillation by Yadi Wen † \dagger, Tianxin Li † \dagger, Enji Liang, Rong Du ∗ *, Yue Fu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-02718.txt.

Description

Yadi Wen and four co-authors study private training of speech classifiers that must be released as audio-only models. During development, richer side information may be available, but the shipped model can use sound alone and must carry a formal privacy guarantee. Training with differential privacy on the private set reveals a failure mode under strong privacy.

On imbalanced tasks, private training can collapse toward predicting a single class, a fault that plain accuracy hides. The authors stress balanced scores and a simple collapse check. Their two-stage protocol first trains a private teacher, possibly using side information, then distills an audio-only student on a separate, recording-disjoint dataset using one round of teacher outputs. Only the student is released, keeping the privacy guarantee auditable.

Catalog record

CREATORYadi Wen † \dagger, Tianxin Li † \dagger, Enji Liang, Rong Du ∗ *, Yue Fu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE27 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.02718

Browse the full library · About the Tower of Babel