Does Financial Speech Need Continual Learning Regularization? A Study of EWC and mini-KFAC on Earnings Call ASR
Adapting Whisper-small to Earnings22 improves financial ASR, so EWC regularization is often counterproductive.
The study adapts Whisper-small from general-domain speech to the Earnings22 earnings-call benchmark and finds performance on held-out financial speech improves rather than suffers catastrophic forgetting. Because the tasks are complementary, Elastic Weight Consolidation regularization is counterproductive. When regularization is appropriate, a layer-wise mini-KFAC curvature approximation achieves lower word error rates and more stable optimization than diagonal EWC.
- Adapting Whisper-small to Earnings22 improves held-out financial speech.
- Task complementarity makes continual-learning regularization counterproductive.
- When regularization is needed, mini-KFAC beats diagonal EWC.
- mini-KFAC reports lower word error rates and more stable optimization.
From the original article
1 min read at arxiv.orgEarnings call transcription presents a realistic continual learning scenario where Automatic Speech Recognition (ASR) systems are updated with new recordings but risk catastrophic forgetting of previously learned financial speech patterns.
By Thai Quoc On · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.