LoRA adaptation shows mixed gains on Whisper ASR models

A new paper explores the effectiveness of lightweight low-rank adaptation (LoRA) on Whisper-style automatic speech recognition (ASR) models. The research compares zero-shot Whisper, pooled LoRA, and group-specific LoRA across Common Voice, LibriSpeech, and VoxPopuli datasets. Findings indicate that adaptation gains are not uniform, being more pronounced in challenging Common Voice groups and consistent on VoxPopuli, while diminishing on the low-error LibriSpeech dataset. Notably, pooled LoRA frequently matches or surpasses group-specific LoRA, suggesting that observed personalization gains should be evaluated against pooled adaptation. AI

IMPACT Findings suggest that pooled adaptation methods may be sufficient for many ASR personalization tasks, potentially simplifying deployment.

RANK_REASON The cluster contains an academic paper detailing research findings on AI model adaptation. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• alphaXiv
• arXiv
• CatalyzeX
• Common Voice
• DagsHub
• Gotit.pub
• Hugging Face
• LibriSpeech
• LoRA+
• ScienceCast
• VoxPopuli
• Whisper

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →