Speechmatics says dramatic speech recognition bias reduced with unlabelled training data
Speechmatics has developed a speech recognition model that it says improves accuracy in understanding groups that have typically been challenging for natural language processing (NLP) systems, such as African Americans and children.
The new Autonomous Speech Recognition software applies the latest deep learning techniques and introduces Speechmatics’ breakthrough self-supervised models, according to the company announcement.
The company refers to Stanford’s ‘Racial Disparities in Speech Recognition’ study, which showed overall accuracy rates below 70 percent for African Americans using both Google and Amazon models, and says its new software delivers overall accuracy of 82.8 percent for African American voices. This difference equates to a 45 percent error reduction, Speechmatics says, or roughly three words per sentence on average.
The lack of manually tagger or labeled data from the above groups for algorithm training has resulted in a lack of accuracy in speech recognition being commonplace, according to Speechmatics. The company says it has avoided the issue entirely by training its algorithms on huge volumes of unlabelled data from the internet, such as podcasts and social media content. Speechmatics collected 1.1 million hours of audio to replace a dataset of 30,000 hours.
Speechmatics says, based on the open-sourced Common Voice project, that its technology was 91.8 percent accurate at understanding children’s voices, compared to 83.4 percent for Google and 82.3 percent for DeepGram.
“We are on a mission to deliver the next generation of machine learning capabilities and through that offer more inclusive and accessible speech technology,” comments Speechmatics CEO Katy Wigdahl. “This announcement today is a huge step towards achieving that mission.
“Our focus on tackling AI bias has led to this monumental leap forward in the speech recognition industry and the ripple effect will lead to changes in a multitude of different scenarios. Think of the incorrect captions we see on social media, court hearings where words are mistranscribed and eLearning platforms that have struggled with children’s voices throughout the pandemic. Errors people have had to accept until now can have a tangible impact on their daily lives.”
A report from Speechmatics at the beginning of 2021 highlighted the importance of voice biometrics for getting the most value out of speech apps.