Posted on

How is speech recognition affected by face masks?

Interview conducted by News Medical

News-Medical speaks to Dr. Joe Toscano from Villanova University about his latest research that investigated the impact that face masks have on speech recognition.

What provoked your research efforts into the ongoing COVID-19 pandemic?

Our research focuses on questions about speech perception and language processing more broadly.

We were interested in studying the effects of face masks, in particular, given the widespread use of masks to help prevent the spread of COVID-19 and possible concerns over how they might affect speech recognition.

Can you give an overview of the different types of face masks available?

In our study, we looked at four types of masks: a surgical mask, an N95 respirator, and two homemade cloth masks (one with a fitted design and one with a pleated design). Cloth masks have been used by the general public during the pandemic, whereas N95 masks are used by healthcare professionals.

We were particularly interested in the effects of the homemade masks since many people use these types of masks every day and there was only limited previous research on how they might impact speech recognition.

How does background noise affect hearing?

Background noise can have a large impact on hearing, though listeners with normal hearing are generally very good at recognizing speech even in background noise. The type of noise we looked at in our study is called multi-talker babble, which consists of multiple talkers speaking at the same time (six talkers in our study). It sounds similar to what you might hear in a crowded restaurant.

Also, it is important to note that, for listeners with hearing difficulty, background noise can have an even bigger effect. In our study, we only looked at effects for listeners with normal hearing.

Can you describe how you carried out your latest research into face masks and speech recognition?

We presented listeners with spoken sentences and asked them to type the sentence that they heard. Sentences were presented in the multi-talker babble noise, in either a relatively easy condition with low levels of noise or a difficult condition with high levels of noise. We recorded sentences while wearing each of the four masks and when wearing no mask.

In all the conditions, the listeners only heard the sentences—there was no visual information like you would get in face-to-face communication. As a result, the study just focused on the auditory effects of the masks. The loss of visual information could play an additional role in face-to-face settings.

What did you discover?

We found that, in low levels of background noise, typical of noise levels in many everyday settings, listeners were very good at recognizing speech produced without a mask, as we expected. They correctly recognized 94.3% of the words in the sentences for this condition.

For speech produced with the masks, they also did rather well, particularly for speech produced with the surgical mask, where they were 93.5% accurate. Overall, masks had relatively small effects in this condition.

In contrast, when the background noise was very high, listeners were only 45.2% correct for speech produced without a mask, demonstrating that this was a very difficult listening condition. Here, we found larger differences between the masks. The surgical mask again led to the best performance among the masks, and listeners were less accurate for the N95 and cloth masks.

Prior research has been conducted into speech recognition whilst wearing masks but has had limited sample sizes and other limitations. How did your research overcome some of these limitations?

We used a large sample size (181 participants) in this study. We also tested conditions that were both relatively easy (low levels of background noise) and more challenging (high levels of background noise).

Previous studies had looked at some of these conditions for certain types of masks on their own, and overall, the effects we found in the current study matched what we would expect based on that previous research.

What are the next steps in your research?

We would like to investigate a wider variety of masks and look at the effects of double-masking. We would also like to know how people have adapted to listening to speech while wearing a mask over the past year.

Listeners are very good at adapting to novel listening conditions, and they may have gotten better over time at understanding speech produced while wearing a mask.

Where can readers find more information?

This material is based upon work supported by the National Science Foundation under Grant No. 1945069. Any opinions, findings, and conclusions, or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation.

Posted on

Mayo Clinic CIO says AI has been key to understanding COVID-19

By Mike Miliard for Healthcare IT News

In his opening keynote Tuesday at the HIMSS Machine Learning & AI for Healthcare Digital Summit, Mayo Clinic CIO Cris Ross enumerated some of the many ways artificial intelligence has been crucial to our evolving understanding of COVID-19.

Way back in March, for instance, researchers were already using an AI algorithm – trained on data from the 2003 SARS outbreak – for “a recurrent neural network to predict numbers of new infections over time,” he said. “Even from the beginning of COVID-19, artificial intelligence is one of the tools that scientists have been using to try and respond to this urgent situation.”

And just this past month, Boston-based nference – whose clinical-analytics platform is used by Mayo Clinic – sifted through genetic data from 10,967 samples of novel coronavirus. Along the way, researchers discovered “a snippet of DNA code – a particular one that was distinct from predecessor coronaviruses,” said Ross. “The effect of that sequence was it mimics a protein that helps the human body regulate salt and fluid balances.

“That wasn’t something that they went looking for,” he said. “They simply discovered it in a large dataset. It’s since been replicated and used to support other research to discover how genetic mutations and other factors are present in COVID-19 that help, both with the diagnosis of the disease, but also its treatment.”

Many other now commonly understood characteristics of the novel coronavirus – the loss of smell it can cause, its effects on blood coagulation – were discovered using AI.

Around the world, algorithms are being put to work to “find powerful things that help us diagnose, manage and treat this disease, to watch its spread, to understand where it’s coming next, to understand the characteristics around the disease and to develop new therapies,” said Ross. “It’s certainly being used in things like vaccine development.”

At the same time, there are already some signs that “we need to be careful around how AI is used,” he said.

For example, the risk of algorithmic bias is very real.

“We know that Black and Hispanic patients are infected and die at higher rates than other populations. So we need to be vigilant for the possibility that that fact about the genetic or other predisposition that might be present in those populations could cause us to develop triage algorithms that might cause us to reduce resources available to Black or Hispanic patients because of one of the biases introduced by algorithm development.”

The profusion of data since the pandemic began has allowed advanced models to be purpose-built at speed – and has also enabled surprise findings along the way.

Sure, “some of the models that are being built that are labeled AI are really just fancy regression models,” said Ross. “But in a way, does it really matter? In any case, [they’re] ways to use data in powerful ways to discover new things, … drive new insights, and to bring advantages to all of us dealing with this disease.”

It’s notable too that the big datasets needed for AI and machine learning “simply didn’t exist in the pre-electronic health record days,” he added.

“Just imagine where we would have been if it was a decade ago and we were trying to battle COVID-19 with data that had to be abstracted from paper files, … manila folders, and a medical records room someplace,” said Ross.

“The investments we’ve made to digitize healthcare have paid off. We’ve learned that the downstream value of data that’s contained in electronic health records systems is incredibly powerful.”