# Personalized Automatic Speech Recognition Trained on Small Disordered Speech Datasets

**Authors:** Jimmy Tobin, Katrin Tomanek  
**arXiv:** [2110.04612](https://arxiv.org/abs/2110.04612)  
**Date:** 2021-10-09  
**Summary page:** https://www.summarizepaper.com/en/paper/2110.04612v1/

## Key points

- Study focuses on personalized automatic speech recognition (ASR) for recognizing disordered speech using small amounts of per-speaker adaptation data
- Lack of available speech data has been a major challenge in adapting speaker-independent ASR systems for dysarthric speech
- Researchers trained personalized models for 195 individuals with different types and severities of speech impairment
- Training sets varied in size from less than one minute to 18-20 minutes of speech data per speaker
- Word error rate (WER) thresholds were used to determine the Success Percentage, representing the percentage of personalized models that achieved the target WER in different application scenarios
- In the home automation scenario, 79% of speakers reached the target WER when trained with 18-20 minutes of speech data, and even with only 3-4 minutes of data, 63% still reached the target WER
- Performance on test sets containing conversational and out-of-domain unprompted phrases showed similar improvements
- Personalized ASR can benefit individuals with disordered speech even with just a few minutes of recordings, which is significant as recording large amounts of samples per speaker is often impractical and challenging for people with speech impairments
- Previous studies required hours of recorded speech data per speaker for substantial WER improvements, whereas this study shows promising results using significantly smaller amounts of adaptation data
- Research highlights the potential and feasibility of personalized ASR for individuals with disordered speech, offering valuable insights into optimizing ASR systems for such impairments and providing a more practical approach that can be implemented with limited recording times per speaker.

## Summary

This study focuses on personalized automatic speech recognition (ASR) for recognizing disordered speech using small amounts of per-speaker adaptation data. Previous research has demonstrated promising results in adapting speaker-independent ASR systems for dysarthric speech, however the lack of available speech data has been a major challenge. Recent work has explored the potential of personalizing ASR models for individuals with speech impairments. The researchers trained personalized models for 195 individuals with different types and severities of speech impairment. The training sets varied in size from less than one minute to 18-20 minutes of speech data per speaker. They used word error rate (WER) thresholds to determine the Success Percentage which represents the percentage of personalized models that achieved the target WER in different application scenarios. In the home automation scenario they found that 79% of speakers reached the target WER when trained with 18-20 minutes of speech data and surprisingly even with only 3-4 minutes of speech data 63% still reached the target WER. The researchers also evaluated performance on test sets containing conversational and out-of-domain unprompted phrases and found similar improvements. These results demonstrate that individuals with disordered speech can benefit from personalized ASR even with just a few minutes of recordings which is significant because recording large amounts of samples per speaker is often impractical and challenging for people with speech impairments. While previous studies have reported substantial WER improvements through model personalization they typically required hours of recorded speech data per speaker; one study achieved an average WER improvement of 75% on a large corpus by recording about two hours per speaker whereas this study shows promising results using significantly smaller amounts adaptation data. Overall this research highlights the potential and feasibility of personalized ASR for individuals with disordered speech providing valuable insights into optimizing ASR systems for such impairments as well as offering a more practical approach that can be implemented with limited recording times per speaker.