# Metric Learning for User-defined Keyword Spotting

**Authors:** Jaemin Jung, Youkyum Kim, Jihwan Park, Youshin Lim, Byeong-Yeol Kim, Youngjoon Jang, Joon Son Chung  
**arXiv:** [2211.00439](https://arxiv.org/abs/2211.00439)  
**Date:** 2022-11-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2211.00439v1/

## Key points

- Authors aim to improve keyword spotting tasks by allowing users to define custom keywords
- Focus on metric learning techniques for training models for user-defined keywords
- Construct a large-scale keyword dataset and introduce a filtering method
- Propose a novel two-stage training strategy based on metric learning techniques
- Demonstrated significant improvements in representations of user-defined keywords and overall performance
- Proposed unified evaluation protocol and metrics for fair comparisons in user-defined KWS field
- System eliminates need for incremental training on new keywords and outperforms previous works on Google Speech Commands dataset

## Summary

, , , , 

In their paper titled "Metric Learning for User-defined Keyword Spotting," authors Jaemin Jung, Youkyum Kim, Jihwan Park, Youshin Lim, Byeong-Yeol Kim, Youngjoon Jang, and Joon Son Chung aim to improve the performance of keyword spotting tasks by allowing users to define custom keywords. This approach not only enriches user experience but also enhances the transferability to unseen terms. Unlike previous works that treat Keyword Spotting (KWS) as a closed-set classification problem, this study focuses on using metric learning techniques to train models for user-defined keywords. The authors first construct a large-scale keyword dataset using an existing speech corpus and introduce a filtering method to eliminate data that may hinder model training. Then, they propose a novel two-stage training strategy based on metric learning techniques. Through experiments, they demonstrate that this approach significantly improves the representations of user-defined keywords and boosts overall performance. To ensure fair comparisons in the field of user-defined KWS, the authors also propose a unified evaluation protocol and metrics. Their system eliminates the need for incremental training on new keywords and outperforms previous works by a significant margin on the Google Speech Commands dataset using both proposed and existing metrics. Overall, this study provides valuable insights into improving user-defined keyword spotting through metric learning techniques and sets a benchmark for future research in this domain.