# A Survey of Multilingual Models for Automatic Speech Recognition

**Authors:** Hemant Yadav, Sunayana Sitaram  
**arXiv:** [2202.12576](https://arxiv.org/abs/2202.12576)  
**Date:** 2022-02-25  
**Summary page:** https://www.summarizepaper.com/en/paper/2202.12576v1/

## Key points

- The paper presents a survey of over 40 papers on multilingual models for Automatic Speech Recognition (ASR), focusing on models built with cross-lingual transfer in mind.
- Cross-lingual transfer is an attractive solution for low-resource languages to benefit from higher-resource languages either through transfer learning or being jointly trained in the same multilingual model.
- Recent advances in Self Supervised Learning (SSL) are opening up avenues for unlabeled speech data to be used in multilingual ASR models, which can pave the way for improved performance on low-resource languages.
- Multilingual models perform better than monolingual counterparts trained with the same amount of data for a single language.
- Combining the data of all languages available leads to better generalization and improves performance across all languages.
- The paper distills key findings from research in multilingual ASR to describe factors that influence cross-lingual transfer and SSL, including the size of pre-training datasets, language similarity, and domain adaptation techniques.
- The authors provide best practices for building multilingual models from research across diverse languages and techniques, discuss open questions, and provide recommendations for future work.
- This survey highlights the potential benefits of using cross-lingual transfer and SSL techniques in building more effective ASR systems for low-resource languages.

## Summary

This paper presents a survey of over 40 papers on multilingual models for Automatic Speech Recognition (ASR), focusing on models that are built with cross-lingual transfer in mind. The majority of the world's languages do not have usable ASR systems due to the lack of large speech datasets to train these models. Cross-lingual transfer is an attractive solution, as low-resource languages can potentially benefit from higher-resource languages either through transfer learning or being jointly trained in the same multilingual model. Recent advances in Self Supervised Learning (SSL) are opening up avenues for unlabeled speech data to be used in multilingual ASR models, which can pave the way for improved performance on low-resource languages. The paper addresses questions about whether multilingual models are indeed superior in performance to monolingual models for both low-resource and high-resource languages. From the papers surveyed, it is clear that multilingual models perform better than monolingual counterparts trained with the same amount of data for a single language. Combining the data of all languages available leads to better generalization and improves performance across all languages. The paper distills key findings from research in multilingual ASR to describe factors that influence cross-lingual transfer and SSL, including the size of pre-training datasets, language similarity, and domain adaptation techniques. The authors provide best practices for building multilingual models from research across diverse languages and techniques, discuss open questions, and provide recommendations for future work. Overall, this survey highlights the potential benefits of using cross-lingual transfer and SSL techniques in building more effective ASR systems for low-resource languages. By leveraging existing resources from higher-resource languages and utilizing unlabeled data through SSL techniques, researchers can improve ASR performance across multiple languages while reducing resource requirements.