Dr. Catherine Breslin is a machine learning scientist specialising in speech and language technology. She has a distinguished academic background with a degree from Oxford and post-graduate qualifications from Cambridge. Her career has been about the application of speech and language research to real-world problems. She has worked for Toshiba and then Amazon, where she played a major role in developing Alexa onto other platforms and now works for Cobalt Speech.
Education - University of Oxford
Catherine Breslin attended secondary school in Yorkshire, having moved there with the family. Maths was her favourite subject throughout school alongside the sciences. She earned A levels in maths and physics before going to Trinity College, Oxford to study Engineering with a sponsorship from GCHQ. She says: “I was always interested in physics as well as maths. I really liked taking maths and applying it to things. When I was doing A levels, for example, my favourite parts were the mechanics modules where you could really see the maths and how it coincided with the physical world. You could take things in the physical world and model them. I never actually considered doing maths, I was torn between physics and engineering. In the end, I decided that engineering was the more practical of the two so that’s where I decided to focus.”
Through the sponsorship, Catherine worked for four summers at GCHQ doing short projects. She says: “We were working on projects. What was great about this, and all internships I think, is that the chance to take your studies and apply it to actual real world problems; it is a great way to learn more, to reinforce what you’ve learned in university but also to see why it’s useful and how you can use it in the real world.”
During her first year, having dabbled in programming at home as a teenager, Catherine took a computing module and opted at the end of her first year to switch to a joint honours degree in engineering and computer science. She says that the theory she learned during her degree forms the basis of the work she does today. She also studied machine learning as a module during her degree; she explains: “I was most interested in information in engineering which covers topics like signal processing, machine learning, control theory, those computer related and engineering subjects.” Much of the focus of the course was on medical imaging and led her to do a research project in her fourth year on using ultrasound to detect breast cancer, she says: “So, that was my first introduction to using computers to do these vision tasks.”
Education - University of Cambridge
Following on from her degree, Catherine went to the University of Cambridge to do an MPhil in Speech, Text and Internet Technology which was jointly run between the Computing and Engineering departments, she explains: “I was interested in the application of computers to some of these problems. I looked at vision in my undergraduate course, and specifically medical imaging, but then I thought, well what about speech, how does that work, how do you get computers to understand what people are saying. The main thing was just curiosity, and Cambridge happened to have a course focused on speech and language processing which was almost a direct continuation of my studies.”
Having achieved her MPhil, she continued to do a PhD in Engineering and automatic speech recognition, and took an interim internship with Toshiba which had a relationship with the engineering department in Cambridge. Toshiba contributed to the funding for her PhD along with the EPSRC, enabling her to research and write her thesis: ‘Generation and Combination of Complementary Systems for Automatic Speech Recognition’.
After initially contemplating an academic career, Catherine took a role with Toshiba; she explains: “My inclination is not towards necessarily research on its own but applied research, so taking what we know about how things work and using it to build in the world.”
AI and machine learning
Catherine explains the difference between AI and machine learning. She describes AI as the field of making computers do things which seem smart. It encompasses a whole host of different approaches. She says: “It’s not just how you do it but also it might include the engineering around that like the design of the system and the user interface. I think it covers more than just the core technology.”
Catherine describes machine learning as a group of algorithms which allow you to learn from data. She explains: “They’re a set of algorithms where you have a set of data and you learn from that set of data how to do a task. Machine learning has come on a long way in the past few years and it’s one of the drivers that has made artificial intelligence systems more capable.”
She concludes: “So I think of AI as being an umbrella, and machine learning being a key set of algorithms underneath that umbrella.”
Many versions of machine learning are now entering the exploitation stages and Catherine gives the example of supervised learning saying: “you can take a database of labelled images and from that labeled data you can predict. If you get a new image it is categorised by which of those labels it falls into. Supervised learning is well proven in research and, I think, is fuelling a lot of the exploitation in industry right now”. Tasks that fall under supervised learning include speech recognition, image categorisation, machine translation, language understanding. Other forms of machine learning such as reinforcement learning, and unsupervised learning are still in the research stage.
While the devices we use today have changed the ways in which we communicate with the introduction of messaging, texts, emails etc., Catherine believes that we will not lose the need or desire for face to face, voice communication.
Catherine joined Toshiba in Cambridge at the research lab in 2008 as Research Engineer in speech recognition. Toshiba was one of the few companies during tha time that was interested in maintaining their research into machine learning. She explains: “I carried on doing applied research into speech recognition systems there. One of my favourite projects was when we built a microphone array that sits on a table, and we used it to record our meetings. So, we’d have group meetings and would record them. Then we built our own speech recognition system which transcribed those meetings. It was pure research; it was our internal research platform and I think it was ahead of its time because it was before the current wave of voice products.”
In 2012, after four years at Toshiba, Catherine took up a fixed term postdoc role as a Research Associate at University of Cambridge, working on dialogue systems to allow her to expand her research. She explains: “That’s the system that will have a dialogue between a human and a computer. You need speech recognition to understand what the person said, but then you need to interpret it and then understand what the computer should say back to the person. That iteration keeps going for a few steps so that you’ve got an actual conversation going back and forth. It was a great chance to learn how that technology worked and to link it up with the speech recognition.”
In 2014, Catherine moved to Amazon as a speech recognition scientist working on speech products in the engineering side of Amazon at a lab in Cambridge. In 2016 she became Manager, Applied Science, Alexa AI.
She joined Amazon just before Alexa was launched and worked on Alexa’s speech recognition. Following Alexa’s launch she worked on how to take it from the Echo device to other devices like the Fire TV. She says: “It was great to be part of Alexa at the time it launched because I think it was a much bigger success than people realised it was going to be, and just seeing how it grew and being able to contribute to that was great.”
She also enjoyed the opportunity to innovate around products that are used by lots of people, she says: “There are very few places that you can go in the world and work on a product that lots of people are using. One of the products I worked on was the Fire TV; a box that you plug into your TV that streams video from the internet. Lots of people have one, so the chance to build systems which help all these people is something you don’t get in many other places.”
For Catherine the challenge of continuing to update and innovate products that are already in mass use without breaking them has been really interesting, she adds: “You have to keep updating and innovating because you’ve got to build the next generation of what people are going to use, but also not break what they’ve already got, and fix all the issues that people have.”
Having worked on the speech recognition of Alexa and her Fire TV project, Catherine started working on the language understanding side of Alexa. She explains: “If you take a system that’s going to talk with somebody, like Alexa, Siri, Cortana or Google, you have a number of different technologies in there. You start off with speech recognition; you talk to Alexa and it understands what you’ve said, and then it has to work out from what you’ve said what you actually wanted, so it has to understand the intent behind what you requested. There are many, many different ways to ask for the same thing. That’s the language understanding. Then there’s the understanding and answering. If you’ve got a factual question that you want the answer to, you have to be able to take a lot of different knowledge sources to find the answer.”
Looking to the future of voice devices such as Alexa and others, Catherine believes that refining the context of requests will be the next challenge and development.
On the subject of where devices like Alexa and privacy intersect, Catherine believes that companies need to be transparent with their consumers in order to develop trust and be better able to provide the best service.
Catherine grew the team in Cambridge and took on management responsibilities taking advantage of Amazon’s management training courses along the way. She enjoyed managing people saying: “I think it’s nice to see people grow as they start to learn, and to be able to deliver and to get more confident.”
Having spent four years at Amazon, in 2019, Catherine decided to move on and has joined Cobalt Speech as a Director, Solutions Architect. Cobalt is a relatively young company founded by Boston based Jeff Adams; a fellow speech and language technology specialist. The company was founded to help businesses develop speech and language recognition technology, but which, due to the relatively small numbers of experts in the field, don’t have the resources to do so.
Catherine is working directly with companies who have ideas for how they want to use speech and language technology, and helping them convert their ideas into the actual technology.
Diversity in technology
Catherine believes that we need to ensure that not only do we attract women into tech careers but they are also able to progress, she says: “The number of women working in technology is low and those numbers go down the higher up you go, so the number of senior women working in technology is very low. I think there is definitely work to be done to change that and it’s twofold. Encouraging girls and women to consider taking technology courses and going into STEM fields is important. We need to work with schools, with parents, to show what these careers can do is a good thing. Also, I think companies have to look at how they treat and how they retain their women, such that they can advance to senior positions in organisations, because without both of those things, things aren’t going to change.”
In giving advice to those thinking about a career in technology and machine learning in particular, Catherine says: “I think there’s a lot of interesting things going on here in machine learning and artificial intelligence right now. It’s a great area to look at because it spans different areas; maths, computing, user interface design, project management – a whole bunch of capabilities which means that you don’t have to be somebody who is going to dive right into the algorithms to be able to work on the technology. You might have a preference for interface design, or for software engineering, and they are also things that you need to be able to build these. There’s a large choice of different career paths you can take now.”
Catherine also points to the many online courses on the subject that are now available and adds that “You can learn a lot about machine learning independently of university, but I think that the technical skills you learn from something like physics, maths and computing at university are very useful. Equally people come from different angles too. Somebody might come from a linguistics or psychology background and gradually move in different ways into the field. Some of the skills that you bring from those other fields are very useful and complementary to some of the other more technical physics and maths subjects.”
For those who are thinking about joining a start-up, Catherine advises: “I think a lot of people are very interested in working in a small company, and I think you shouldn’t necessarily be put off if you have no experience. You have to look at yourself and what it is that you want from your career because, if you are not the kind of person who copes very well in a chaotic kind of environment and you need more guidance to get started, then maybe a start-up is not the right place. But if you’re happy with that level of ambiguity and that level of chaos, then it can be a great place.”