# Retrieve and Copy: Scaling ASR Personalization to Large Catalogs

**Authors:** Sai Muralidhar Jayanthi, Devang Kulshreshtha, Saket Dingliwal, Srikanth Ronanki, Sravan Bodapati  
**arXiv:** [2311.08402](https://arxiv.org/abs/2311.08402)  
**Date:** 2023-11-14  
**Summary page:** https://www.summarizepaper.com/en/paper/2311.08402v1/

## Key points

- The paper addresses the challenge of scaling contextual biasing techniques in ASR models to large catalogs
- Introduces a "Retrieve and Copy" mechanism to enhance latency while maintaining accuracy at scale
- Proposes a training strategy to mitigate recall degradation due to increased confusing entities
- Achieves up to 6% more Word Error Rate reduction and a 3.6% absolute improvement in F1 compared to baseline
- Identifies limitations in methodology, particularly regarding F1 score drop with increasing catalog size
- Future work will focus on addressing challenges for practical use cases of scaled ASR personalization systems

## Summary

, , , , 

The paper "Retrieve and Copy: Scaling ASR Personalization to Large Catalogs" by Sai Muralidhar Jayanthi, Devang Kulshreshtha, Saket Dingliwal, Srikanth Ronanki, and Sravan Bodapati addresses the challenge of scaling contextual biasing techniques in automatic speech recognition (ASR) models to large catalogs. The authors propose a "Retrieve and Copy" mechanism that enhances latency while maintaining accuracy even when scaled to a large catalog size. Additionally, they introduce a training strategy to address the degradation in recall caused by an increased number of confusing entities at scale. The results show that their approach achieves up to 6% more Word Error Rate reduction (WERR) and a 3.6% absolute improvement in F1 compared to a strong baseline. However, the study also identifies limitations in their methodology. <ks>Improved Latency</ks>

Despite improvements in latency with increasing catalog size, there is a consistent drop in F1 score. Incorporating hard negatives based fine-tuning helped mitigate this issue but further research is needed to scale the approach to even larger catalog sizes. <ks>Fine-tuning for Scalability</ks>

Contextual biasing techniques can also lead to regressions on common words in the dataset, particularly evident with long audio datasets like VoxPopuli when using contextual biasing on large catalogs. <ks>Data Privacy Concerns</ks>

Looking ahead, future work will focus on addressing these challenges to enable practical use cases for scaled ASR personalization systems. Privacy and intellectual property concerns prevent the release of training and evaluation datasets at this time but may be addressed in subsequent research efforts. <ks>Real-world Applications</ks>

In conclusion, "Retrieve and Copy" offers a promising solution for scaling ASR personalization to large catalogs, showcasing significant improvements in WERR and F1 scores while maintaining inference speedup per acoustic frame. Further refinement and adaptation of the proposed methodology are necessary to fully realize its potential for real-world applications.