From Siri, to Alexa, to Google Now, voice-based virtual assistants have increasingly become ubiquitous in our daily lives. So, it is unsurprising that yet another AI technology – speech recognition systems – has been reported to be biased against black people.
A scientific study has found that the speech-to-text tools developed by the five major tech companies – Google, Apple, Amazon, Microsoft and IBM – suffer from large racial disparities in speech recognition technology. In other words, these technologies recognise white voices better than others. For instance, the rate of misidentified words by these systems are at 19% for white people but rises to 35% for black people. When it comes to the unreadability of audio snippets by these systems, the study shows that only 2% of the snippets from white people were considered unreadable, while this unreadability increased to 20% with black people. While not all applications run on these five speech recognition systems, it is worrying that this is a pattern across the majority of big tech companies.
A reason for these errors can likely be found in the lack of diversity in data sets; and the false assumption that different ethnic groups are properly represented by these companies. However, improving diversity in datasets (or hiring) may not be a priority for these companies despite what they claim, nor the main point. It is crucial to remember that bias cannot be rid by technology. Technology could actually exacerbate, or distract us from, the main problems we need to address.