For human lip readers, context is key in deciphering words stripped of the full nuance of their audio cues. But a technology model for lip-reading developed at the University of East Anglia in the UK has been shown to be able to interpret mouthed words with a greater degree of accuracy than human lip readers, thanks to the application of machine learning tech to classify the visual aspect of sounds. And the kicker is the algorithm doesn’t need to know the context of what you’re discussing to be able to identify the words you’re using.
While the model remains a piece of research at this stage, there are scores of potential applications for technology that could automagically transform visual cues into accurate speech — whether it’s helping people who have audio impairments, or enhancing audio-less security video footage with additional speech data — or even to try to figure out exactly what charged word one footballer spat at another in the heat of a match…
Such a tech could also be applied as a fallback for poor audio quality on a mobile or video call. Or for automating subtitles. Or even perhaps to power a front-facing camera-based mobile ‘voice’ assistant which you wouldn’t actually have to speak to but could just discreetly mouth commands at (how cool would that be?). Safe to say, the list of applications-in-waiting for machine powered lip-reading is as long as the dictionary is deep. So there’s bags of future potential if only researchers can deliver the goods.
The UAE team behind this new machine learning training model for lip reading have been looking purely at visual inputs — so training their model on the shape of the mouth as certain sounds are spoken, without any audio input cues at all.
“We’re looking at… visual cues and saying how do they vary? We know they vary for different people. How are they using them? What’s the differences? And can we actually use that knowledge in this particular training method for our model? And we can,” says Dr Helen Bear who created the visual speech recognition tech model as part of her PhD, along with Prof Richard Harvey of UEA’s School of Computing Sciences.
“The idea behind a machine that can lip read is that the machine itself has got no emotions, it doesn’t mind if it gets it right or wrong — it’s just trying to learn. So in the paper… I’ve been showing how we can use those visual confusions to make better phoneme classifiers. So it’s a new training method,” she adds.
Dr Bear notes that a lot of current research in the lip reading field is looking both at audio and visual cues to try to improve the accuracy of machine lip reading. So the UEA model stands out on merit of focusing solely on visual speech to try to boost machine-powered lip reading.
“We were effectively pretending that that audio signal is not there at all,” she says. “The idea being you can either have a lip-reading only system or it could be used in an audio-visual system that maybe one day hopefully it would be nice if it could jump in, do the visual signals only until the audio comes back in, for example, if you’re on a Skype call and the audio goes out but you can still see somebody.”
The core challenge for lip reading techniques in general is there are — at least to the human eye — fewer visual cues than there are acoustic audio sounds humans make. Examples of sounds with confusingly similar shapes when seen on the lips are ‘/p/,’ ‘/b/,’ and ‘/m/’ — all of which typically cause difficulties for human lip readers. However UEA’s visual speech model is able to more accurately distinguish between these visually similar lip shapes.
“It turns out there are some visual distinctions between ‘/p/,’ ‘/b/,’ and ‘/m/’ but it’s not something that human lip readers have been able to achieve,” says Dr Bear. “But with a machine we are showing that those distinctions are there, they do exist and our recognizers are much better at doing it.”
“If I was to try and build a classifier to recognize just the /p/ sound what I would have done is it’s first trained on all the sounds that look the same. What we then do is we then refine that training by doing some more iterations of training which are only on the /p/ sound,” she says, discussing the training technique.
“We’re actually learning and understanding what all these visual units mean and why they differ between people and we’ve used that knowledge in order to change the conventional lip reading system and make it better. It is a significant step forward,” she adds.
‘Much better’ is still relative — with the accuracy level for lip reading remaining low. Accuracy at the word level for the model stands at between 10 and 20 per cent (i.e. for correctly identifying a word), according to Dr Bear — albeit she stresses that’s still much higher than guessing. Over a sentence it of course becomes easier to distinguish sense from an entire transcript, she adds.
“In all honesty we’re not 100 per cent sure [why it works],” she tells TechCrunch. “We just know that with our particular classifiers if we train them in the right way, with the right data, they’re not biased towards anything.
“The complexity is that understanding the science of why visual speech is as complex as it is is a much harder question than can we use machine learning to get better results. We know that machine learning is evolving all the time, and we’re getting different types of classifiers… But actually asking the hard questions of what it is they’re learning and how visual speech is and how much it varies and how we’re going to control all those variables, those are the harder questions.”
Asked to hazard a guess on how far out the research might be from being usefully commercialized in an application, she jokes: “If I worked for Google probably a lot sooner!”, before adding that any commercialization is likely to be “a fair few years away yet”.
“We’ve still go things we need to learn and understand,” she says, characterizing the research as just one piece of an interlocking series of linguistic models that will be needed enable machines to adroitly and accurately pull speech data from the twists and turns of human lips.
It’s also worth noting that the UEA model was also solely focused on the English language. So the scope of the challenge ahead to deliver on the promise of lip-reading powered applications is not to be underestimated.
Could the UEA model be combined with other predictive linguistic techniques — perhaps machine learning based next-word prediction technologies — in order to further enhance lip-reading capabilities? “That’s exactly what I love to be able to do,” she says. “To have something that robust would be amazing but that’s going to take quite a bit more work as yet. It’s not going to be going to market any time soon.”
DrBear is presenting the research findings at the International Conference on Acoustics, Speech and Signal Processing in Shanghai this Friday when her paper — Decoding visemes: Improving machine lip-reading — will also be published. The research was part of a three-year project, supported by the Engineering and Physical Sciences Research Council.