English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Zendy Plus

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Zendy Tools

Zendy Open

It is well-known that neural networks can unintentionally memorize theirtraining examples, causing privacy concerns. However, auditing memorization inlarge non-auto-regressive automatic speech recognition (ASR) models has beenchallenging due to the high compute cost of existing methods such as hardnesscalibration. In this work, we design a simple auditing method to measurememorization in large ASR models without the extra compute overhead.Concretely, we speed up randomly-generated utterances to create a mappingbetween vocal and text information that is difficult to learn from typicaltraining examples. Hence, accurate predictions only for sped-up trainingexamples can serve as clear evidence for memorization, and the correspondingaccuracy can be used to measure memorization. Using the proposed method, weshowcase memorization in the state-of-the-art ASR models. To mitigatememorization, we tried gradient clipping during training to bound the influenceof any individual example on the final model. We empirically show that clippingeach example's gradient can mitigate memorization for sped-up training exampleswith up to 16 repetitions in the training set. Furthermore, we show that inlarge-scale distributed training, clipping the average gradient on each computecore maintains neutral model quality and compute cost while providing strongprivacy protection.

Unintended Memorization in Large ASR Models, and How to Mitigate It