Speaker Personalization for Automatic Speech Recognition using Weight-Decomposed Low-Rank Adaptation

[Conference] INTERSPEECH 2024, Kos Island, Greece, September 2024

Authors:

George Joseph, Arun Baby

Abstract:

This paper addresses personalizing ASR systems with limited speaker data. It explores low-rank adaptation (LoRA) and Weight-Decomposed Low-Rank Adaptation (DoRA) techniques applied to a cascaded conformer transducer model. The proposed approach demonstrates an average relative improvement of 20% in word error rate across speakers with limited data.

Cite:

@inproceedings{joseph24_interspeech,
  title = {{Speaker Personalization for Automatic Speech Recognition using Weight-Decomposed Low-Rank Adaptation}},
  author = {George Joseph and Arun Baby},
  year = {2024},
  booktitle = {{Interspeech 2024}},
  pages = {2875--2879},
  doi = {10.21437/Interspeech.2024-1434}
}

Links:

Proceedings

PDF