Search:
Match:
1 results

Analysis

This paper addresses a significant problem in speech-to-text systems: the difficulty of handling rare words. The proposed method offers a training-free alternative to fine-tuning, which is often costly and prone to issues like catastrophic forgetting. The use of task vectors and word-level arithmetic is a novel approach that promises scalability and reusability. The results, showing comparable or superior performance to fine-tuned models, are particularly noteworthy.
Reference

The proposed method matches or surpasses fine-tuned models on target words, improves general performance by about 5 BLEU, and mitigates catastrophic forgetting.