Overcoming bias in speech recognition with Speechmatics
Increasingly embedded into people’s professional and personal lives, speech recognition software can be found in voice assistants, driverless cars and in contact centres.
Speechmatics, global experts in deep learning and speech recognition, provide autonomous speech recognition technology that has the ability to understand all voices. The company looks to develop technology that is as unbiased as possible.
Last year, a study revealed that Speechmatics’ software outperformed tech giants including Google and Amazon when it comes to addressing bias in speech recognition.
Commenting on the importance of this software, particularly in business, Speechmatics’ Data Science Engineer, Benedetta Cevoli spoke to AI Magazine, she said: “The core value of speech recognition, especially in a business environment, is the rich understanding it provides. Businesses of all kinds are looking for ways to understand what their customers want: using speech recognition can play a major role in this, giving breadth and depth to understanding what clients are saying, whether it be by accurately recording language or deriving insights from the text produced.”
Issues of bias in speech recognition
With many artificial intelligence (AI) and machine learning (ML) technologies, speech recognition technology relies on the data it is trained on for its quality. As the data for speech recognition has been limited as it comes from a small section of society, the technology can contain significant bias if not tackled.
This bias can massively reduce the context in which speech recognition works well. Having in-built bias can mean there are many real-life scenarios where minority voices are not being understood effectively, with real-life consequences.
“Take, for instance, an emergency call and the need to transcribe key information - being misunderstood at that moment could be life-affecting,” explained Cevoli
The Data Science Engineer also outlined a way to overcome such issues, she said: “By using self-supervised learning techniques we can mitigate against this: they take vast amounts of unlabelled data and use some property of the data itself to construct a supervised task, without the need for human intervention. Using this method, we can massively increase the amount of data machines learn from and thus give them a significantly more accurate representation of all voices.”
Tackling bias in speech recognition
Monitoring is also a crucial step in the fight against bias in speech recognition. Cevoli explained how this can be achieved: “We need to build systems that can zero in on problems as soon as they arise and then formulate strategies to assess how best to tackle them.”
She concluded by outlining how to tackle these biases as soon as it has been recognised: “From the earliest stages of development, right through to the late application of these technologies, regularly questioning whether biases are being addressed is crucial. Questions like: are we thinking about the problem in a uniform way? Is the approach we're taking inclusive? Are we subconsciously marginalising a particular group of people?”
“As soon as we start designing any new technology, we need to build in ways to evaluate the role of diversity and inclusion at regular intervals and marry this with continuous research into ways in which we can enhance this evaluation process. It's also of critical importance that the teams working on these new technologies are diverse within themselves,” Cevoli added.