These announcements have the potential to unlock enormous economic and social value.
It’s no secret that over the years, technology has consistently eliminated marginalized groups and perpetuated the prejudices that exist in society. While acknowledging that people are relying more than ever on technology to do their jobs, if technology only understands certain voices, it will strengthen the workforce with diverse groups of people. The problem becomes clear. While economically savvy, the government’s benevolent commitments will fail unless we work and, more importantly, significantly improve how inclusive our technology is.
One of the fundamental foundations of inclusiveness is accuracy, which is proven by the speech-to-text industry. The industry is projected to be worth about $45 billion by 2032. Speech recognition technology is already being used in a variety of core services, from banking to healthcare. The need for it is growing so fast that it will soon become the primary interface for all technologies. It is also a noteworthy application from the point of view of returning to work for over 50s, as more than 40% of his over 50s employees are deaf or hard of hearing. This makes the need for accurate transcription even more important. This is especially important in environments with high background noise, such as call centers. This is an area where Big Tech is lacking. According to Google’s own research and testing, Google comes in last when it comes to understanding people over the age of 60, generating about 20% more errors than the under-60 population. This is a very concerning result when more than 1 billion people around the world use his Google products and services.
If people are not understood in the workplace by everyday technology and if we allow Big Tech to continue to innovate around marginalized communities, then we will create a truly economically viable and competitive workforce. I know the answer, but what is the solution? I would argue that it may be somewhere between effective regulation and revolutionary scale-up.
the problem comes from the data
Today’s AI or machine learning (ML) technologies are very good at identifying patterns that exist in data. The problem is that these algorithms often identify patterns in data, including social and historical biases. For these reasons, the datasets used to train these models are often at the heart of the issue. Diversity and appropriate representation of historically marginalized groups in data is essential to developing inclusive technology that works for everyone. This is never easy. For speech-to-text conversion, for example, one interesting approach is to leverage self-supervised learning (SSL) techniques to reach a wider range of speech. By learning patterns from large amounts of unlabeled multilingual data, SSL learns fine-grained representations of acoustic features, enabling it to understand a wide range of speech.
looking ahead
Against the backdrop of a tricky and competitive economic environment, governments are focused on extracting value from previously overlooked groups. However, we need to work with the UK’s thriving tech community to ensure that goodwill is communicated. Unless the technology that underpins our professional lives is truly inclusive, policies, commitments and surface-level actions risk overwhelming impact.
Benedetta Cevoli, Senior Data Scientist at Speechmatics.
