Cerence Inc. (CRNC) Earnings Call Transcript & Summary

Good morning to everyone. This is Richard Yerganian, I'm Vice President of Investor Relations for Cerence. Welcome to our first technology showcase. This event will be very different in several ways from past events. First of all, that will be hosted by Prateek Kathpal, our Chief Technical Officer. Sanjay Dhawan, our CEO; and Mark Gallenberger, our CFO, will be in the audience, just like the rest of you for this event. Secondly, this event is solely focused on helping analysts and investors gain a better understanding of Cerence's products and our innovative technology. So financial-related topics will not be covered during this call, just so you're aware of that. And finally, because of time constraints, we're not taking any questions during the event. However, if you do have any questions, please feel free to follow up with me afterwards. We will make sure to get an answer to whatever question you may have. Before starting, I just want to caution everyone to our safe harbor statement. Please read through this at your leisure or any of our SEC filings as well. So let's get going. I'd like to hand the call over now to Prateek, our Chief Technology Officer. Prateek?

Thank you, Rich. Well, hello. Welcome, everyone. I'm Prateek Kathpal, CTO at Cerence. We, at Cerence, are involved in creating the world's best in-car experience. We're an interactive artificial intelligence company, and we use AI and deep learning to provide the experiences beyond the voice. So if you look at Cerence today, we have about 325 million plus cars on the road. And 1 in every 2 cars shipped globally in 2019 had Cerence technology. Now we have received several requests from you guys over the past few months that you would like to learn more about the technology, understand how it comes up together and how the solution gets used. So Cerence today, as an organization, we have more than 1,800 developers, and these are speed scientists and experts worldwide. They are involved in various stages of the vehicle development, architecture and a wide -- and have a wide array of the knowledge and experience in AI and deep learning and neural networks. Doing AI itself is not that simple, but creating a delightful intuitive experience and bringing it together using voice, touch, gestures, emotions and gaze adds a whole new layer of complexity. Now when you add on top of that, 70-plus languages support with a moving vehicle, where the environment is typically not within your control, you can have, let's say, the engine noise, the road noise or various external elements, you're talking about a very, very complicated system working together to create a human-like experience. So the purpose of today's session is to educate, to showcase some of this technology behind. What you guys see in your daily life in the cars you're driving on the road, we'll talk about how those cars are powered and what's the technology behind from the Cerence side in some of those. So with that, we'll talk about what we are going to talk about today. So we have a very full agenda today for the next 90 minutes or so. I'll start by talking about how AI is humanizing your car. Followed by me will be Udo, who'll talk about the building blocks of conversational AI. After which, Christophe will talk about multimodality components in interactive AI. George is then going to dive into how these interactive AI blocks of conversational AI and the multimodality components, how all these can be assembled into turnkey solutions. Then Nils is going to talk and discuss about some of the applications that are based on top of this technology. Now no solution is complete without a great UX, so Adam will talk about how we invest into creating amazing experiences for the users in the vehicles. We'll then have a guest join me for a virtual chat. We'll have Charan Lota, Chief Engineer, Connected Technologies at Toyota North America, join us for a discussion on the role of AI in the future-connected vehicles. And finally, in the end, we'll have Rick, the CMO at Cerence. He'll talk about how all this technology and innovation, how it comes up together to deliver a final great experience, and he'll cite a real example over there as well. With that, let's get started. Next slide, please. So -- next slide, please. Now when you talk about a complete in-car experience, several different things come into play. Human machine interface, or HMI, as we call it as well, in its simplest terms, includes any device or software that allows you to interact with a machine. This can be a simple touch screen display mounted on a machine or it could be a very technologically advanced multi-touch, voice-enabled with biometrics, vision controlled, an enhanced control panel, which might have even a connection with a mobile technology such as your smartphones or smart watches. Now our traditional HMI solutions were stand-alone, isolated touchscreens. Those were -- that were deployed by the OEMs. But today, the new HMI solutions, they require a very integrated approach. They're cloud connected. They're IoT enabled, and they provide a very intuitive and a natural way for people to interact. So when it comes to in-car experience, we, at Cerence, are redefining this completely. We're using AI integrally in a multimodal form from voice to biometrics to gaze to gestures, and we're focusing on not only connecting drivers to their cars, but also to the entire digital universe. So we're basically building the heart and soul of the modern car as a part of this. Next slide, please. So a multimodal human machine interface, or HMI, as I said before, it allows users to control their vehicles using various input modalities. This could be the speech, the touch, the gestures and the handwriting. Now it has the potential to minimize the load, the cognitive load on user's brain when performing these complex tasks. And at the same time, makes the driving experience much better and safer. So if you look at the modern car today, we have hundreds of different sensors. They provide -- that can provide vast amount of real-time information. This could be from vehicle data to road conditions, weather information. And then you have the user data, which is your personal life on the phone, your apps, your preferences. And then finally, now the IoT, which is connecting you to your home or, let's say, some other devices. So this is a very complex ecosystem, right? And it consists of a knowledge graph, your home graph, and even your personal preferences. Since AI and deep learning models use all this data along with neural networks and algorithms, so real time, think of -- it mimics this function of human mind and enables you to interact with the system in several different ways. And it also enables the vehicle to respond and take actions without that human intervention needed. So as an example, it enables you not only to adjust, let's say, the temperature, but -- or respond to how much is the gas left in a tank, but it also learns from individual driver's speech patterns and behaviors over time and enables you to safely and very intuitively interact with the car through various multimodality components like gestures and gaze and emotions. Next slide, please. So if you look at the profile on your phone, it is very different today than the profile on your car. You might be using, let's say, a phone to dial into your Zoom or your Teams meeting. You might be checking your WhatsApp messages or listening to your music on your phone. But when you sit in your car, you do want to keep your phone in the pocket. But because of the disconnect between a car profile and your phone profile, you're forced to pick up your phone while driving to maybe, let's say, reply to that phone call or reply to that message or maybe even an e-mail. So we at Cerence, as we started defining the car of the future, we did not just start with only how to improve the experience, but we also took over the goal and the objective on how to bring in your personal life seamlessly into the car. So the improvements that we are making are more towards how we can connect these 2, your personal life and your time in the car and not how we can create a new profile of yours on the car. And it's actually very different, right? So you're connecting these 2 things as opposed to recreating them. Now the -- let's say, the songs that you're listening on your phone or the music library that you might have on your phone that becomes very easily then accessible in the car. And then if you, let's say, share your car with your spouse or another user, their personal life, their preferences becomes primary and not yours. Now with these innovations that we'll be bringing in the future, we're creating a single digital identity inside and outside the car. So we're focusing on creating a digital twin of the car on the cloud. Think of it as a virtual model of the connected vehicle. The digital twins, they allow you to capture the behavioral and also the operational data of the vehicle and result in a very personalized experience for the users. So next slide, please. Now if you look at the core foundation of Cerence, it is this hybrid edge and cloud-connected AI platform. This is what enables to create a simplified human experience for drivers. But in the back of it is an extremely complicated stack and defensible set of technologies, all working together. So it starts with several, I would say, inputs and AI technologies, including automatic speech recognition, natural language understanding, text-to-speech and beyond voice multimodal interactions as well. You'll hear some of the presenters today who'll go deeper into these blocks as I walked you guys through the agenda as well, and how they're all packaged together to bring an overall solution to the market. And how the Cerence cloud, which is a key element over here, is scalable and handles 1 billion-plus connected transactions in a year. Now the Cerence cloud, it acts as a data brokerage, right? So between the third-party ecosystems, between your data, the user data, the vehicle data and also the environmental information. So it uses this data to real-time enable deep learning, which we can then also use to further train our AI models. But then at the same time, the OEMs can use this data to improve the experience for their users. Next slide, please. Now when we talk about the Cerence road map for immediate future, it's very clearly focused on 2 things. Enhancing the in-car experience is obviously the first part of it, and we enable the OEMs, as a part of this, to create a great brand-able experience and at the same time, create a great end-user experience as well. And the second over here is to innovate with the cloud. So as more and more cars become cloud connected, it's possible now to offload some of the in-car edge competition to the cloud. And it just opens up and enables quite a lot of new use cases that were never possible before. So I'll give you an example of an AI-powered infotainment system will know exactly what movie to play for your kids in the back seat or, let's say, play the song that you like after you had a long day based on your calendar or wake up time. And it enables you to have a much safer journey, and at the same time, reduces, as I said before, the cognitive load on your brain while driving. Next slide, please. So the next-generation interactive AI is not just a simplified user experience. It also requires the intelligent data to support it. And it needs to be environment-aware. It needs to be ecosystem-aware, and it also knows about your preferences and connects to your profile seamlessly and enables multimodal decisions and interactions, all at the same time. So with this, I'll pass on to Udo to talk about the building blocks of conversational AI, which is a key element over here in the interactive artificial intelligence stack.

Thank you, Prateek. So I'm Udo Haiber, and nice to talk to you all. I'm heading R&D within Cerence, and I'm in the auto and speech industry now since more than 20 years. So I'm excited to have the opportunity, during the next 10 minutes, to build a bridge that allows you to get some insights and previews on what is cooking in our technology labs. Next slide, please. So I'm going to start off with a high-level diagram on what are the core building blocks and how to use the blocks and how they are being connected to each other. So -- but before I talk to the diagram, I wanted to mention that all the technologies in those blocks are used on both ends, the edge and the cloud. That is also why we call it hybrid technologies. So using the same engine in both places is a big advantage, as we are now able to run large deep learning networks also on the edge in real time with state-of-the-art processors. First cars with this new technology will hit the road already next year. By the way, even though we talk about engines and hybrid and automotive, it has nothing to do with powertrains, and our engines also do not consume fuel as they run purely electric. So now let's move back to the diagram. On the left side, you'll find the user input. In this case, a driver may have said, "Find a charging station on my way." First thing we do is to clean the speech signal in the so-called speech signal enhancement engine, SSE. We remove noise and echoes or deal with multiple microphones, for example, to find out from which place in the car an utterance was spoken. So after that the signal splits and goes into 2 automatic speech recognizers, one in the cloud and one in the car, and ASR does convert speech into text. This text is then processed by the natural language understanding step, the NLU step and delivers a language-independent semantic representation of what the user said. This can then be followed by content retrieval on the cloud and is then sent back to the car where we decide which of the results to use in our hybrid arbitration module. The dialogue engine then finally decides on what feedback to give to the user. And in this case, will play a prompt using our text-to-speech engine, TTS, and may also display a map on the screen with the charging stations there. So with that, I move on to the next slide. Now that you have seen how the building blocks work together, I want to dig a little deeper to understand what is inside each block. As you can see, maybe also on the many more acronyms on this slide, all our engines are powered by deep learning modes, and the respective models are built on the compute grid in our own deep learning training toolkit, which is based on TensorFlow and others. So the technology [indiscernible] specific networks to solve specific problems of spoken language. Starting again with our signal [indiscernible] and [ we'll know ] more about this technology in the next slide. But the next in the list is speech recognition, which was the first tech in-house to adopt deep learning. [indiscernible] acoustic modeling and language modeling. On the NLU, or natural language understanding, we [indiscernible] and last but not the least in the list, there is the text-to-speech, TTS...

Excuse me, Udo. We're having a hard time hearing you. I think there's some connection bandwidth issue. So maybe what we could do is we'll move on to the next section of Christophe. And if you could dial back in, and we'll come back to yours in a minute, okay? So can we get to the -- to Christophe's section, please? Thank you.

Okay. Thank you, Prateek. My name is -- good morning. My name is Christophe Couvreur. I'm Head of Product at Cerence, and I work firstly, with Udo on our AI building blocks. And Udo is going to cover -- was covering the main traditional building blocks of a speech recognition system and conversational AI system. However, as those systems evolve towards even more humanized interaction, there are more building blocks that come into play, more multimodal, things that go beyond speech because we are humans. We have been wired for face-to-face communication over millennia of evolution, and we pay attention, consciously or not, to more than just the words spoken. We notice how people speak, their body language, the context of the interaction, what's happening around us, et cetera. So a really natural interactive AI should take those other things into account, too. Next slide. And starting with the speech output as an example. As you're going to see in -- when listening to a voice, there is more than the words, the tone of the voice, the accent of the speaker, like mine, French. The style of speaking matters to the listener. We notice them, and this impact we perceive what is spoken. And for a long time, computers have used a very flat neutral voices that were technically easy to build, things of HAL in 2001: Voice -- 2001: A Space Odyssey. But now that deep learning arsenal that Udo started to cover is at our disposal, we can create voices that are much more expressive and can achieve near-human quality in all aspects. And to infill more subtle like the emotion they convey, the style they use. An example of that is our latest Cerence Reader product. It's a virtual newscaster that can read news the way a human would read news, including deciding on how to read them because you're not going to read, say, the sports results, the way you would read an obituary. Well, it probably depends if your favorite sports team lost, maybe you would. But normally, you would have a different way of reading depending on the content. And that's an example that you're going to see here of what the system can do. Can we play the example? [Presentation]

So as you could hear, the virtual newscaster picked a style, slightly upbeat. It paused in the right places. You could even hear some breathing sounds and this makes the voice a lot more natural than traditional TTS voices, which, in turn, make them a lot easier to listen to for a longer period of time, like on a news text. Next slide. Next slide. Thank you. So as I just said, speech is more than just the words spoken, the tone of the voice matter. But beyond that, what else matters is what it conveys towards the system. So Udo was explaining that speech is converted from [ speech ] to text and then processed, which means that any extra information like the emotions, the stress that may be present in the voice has been removed by that stage, but this is also important information. So what we are doing now is we are building new components that can extract that from the speech and other modalities and add them as additional information provided to the system. We have built an emotion classifier that operates next to the speech recognition system, the SR system, and classifies the emotional traits present in the voice of the speaker. And with that, the dialogue component can decide how to react and maybe adopt an upbeat tone if the speaker is down or maybe a more soothing tone if the speaker is stressed. If there are cameras in the car, by the way, monitoring the occupants, this is possible to combine the video signal to do a video-based emotion detection and merge that with the audio information to come to an even better decision. Next slide. So the emotion is one extra information that your voice may carry. But beyond the words and the emotions, your voice also carries a lot of additional information about who you are. And using voice biometry, which is again, a technology component powered by deep learning, you can identify the characteristics of the persons speaking. Are they male or female? Are they old or young? Are they who they pretend to be? And based on that info, you can adjust directions of the system. For example, if the system identifies who the speaker is, it can respond to a comment like, "Load my profile," by loading your profile and not that of your child or that of your wife. It can also be useful to authenticate the identity of the person speaking. Are they who they pretend to be? And that's important if you want to validate monetary transactions by voice, and Nils will tell you a bit more about that later. Once again, if there's a camera in the car, what we can do with the voice, we can also do with a video signal. So we can do face biometric next to voice biometric. And combining that with additional information about the car and the car data in the biometric suite, we can do a fusion of those information sources to come to a more reliable and less intrusive experience for the drivers. For -- an example of that would be, if the car is speeding on the freeway and the doors are locked, you know that the driver is not changing or it shouldn't be. So you can make a decision about the driver, about who's present in the car that is much more reliable because you know it's not going to be somebody from the outside that you have to reject. Next slide. Now I've mentioned a few times that we have cameras in the car. And those cameras are put in usually for a good reason. You may have been wondering where does that little pop-up icon on your dashboard of maybe your Mercedes coming from. That little coffee icon sometimes that advises you to take a rest or to go grab a coffee. Well, it's using an inside-mounted camera to monitor the driver for signs of drowsiness. Those cameras are put in the car today mainly for safety reasons. But once they are in the car, with the AI software, we can enable more advanced multimodal features because we can use those cameras to detect elements of human interactions like gaze, body language, gestures. And when we look at something as we speak, we provide information. We may say, "What's this?" by looking at someone. We may be pointing at a thing. We may be doing a gesture. And a real humanized virtual assistant that would act like human should be able to understand that body language, take it into account in the way it's interacting with the users. So those cameras that are put in for driver monitoring can also be tracking the gaze of the driver. That gaze alone is not sufficient. You need to combine that with additional information. And that's exactly what your Cerence Look product does. So it's taking the gaze information. It's combining that with information about the surroundings of the car, the map data, what is around it, what is visible from the car, which is calculated using a visibility model of the surroundings of the car, the speed at which the car is moving, the location of the car, the timing of the query. And taking all of that into account, it can resolve a request, "What is this restaurant?" into, "Okay, the person was looking at this restaurant at that point in time, that side of the road," and fetch from a point of interest database the information about the restaurant and provide that back to the user. So this will give you a very precise answer without having to touch your phone, without having to look at anything just by using that information from the car, combined with the queries you made to the virtual assistant. Next slide. So far, we have been talking about human interaction. But in fact, the same family of building blocks can bring capabilities to the system that are human capabilities and sometimes superhuman capabilities that go beyond what we, humans, do. So an example of that is emergency vehicle detection. We, human, can detect things like noises, other sounds beyond just the speech. And we can tell which direction the sound is coming from. This is a very useful capability that virtual assistant and cars could benefit from. So the way we have enabled that in the car is having a module that uses the signal of the microphones inside the car or outside of the car to detect sirens, police cars, firetrucks, ambulances. It's deeply integrated with the audio system, which allows it to use speech signal enhancement techniques that audio -- that Udo mentioned, to remove the echoes from the loud speakers on the microphone signal, the background noise and then detect the siren sounds even when the radio is playing at full blast or the road conditions are very noisy. Once it has detected the siren, it can warn the driver or it can lower the music volume to make you hear the sound of the siren appropriately; or if your car is in autonomous driving mode, it can tell the autopilot that there's an emergency vehicle approaching from the back, needs time to request the driver to take control, for instance. And this -- those are very useful capabilities to add to the AI in the car. I have a short video illustrating that, that we can play now. And those are 2 cars, 1 equipped with a siren that you see on the side and the black one. We're inside the black car and the car with siren's in front at this time. And you can see at the bottom, as the car passes by, how it's tracking the direction of the siren of the car bearing the siren and it keeps tracking it even as we move further away. And we stopped hearing the siren, yet the system is still able to detect it and hear it. What you were hearing here, by the way, was the German Police siren. It's one of the things we have learned working in this field. There's a lot of different type of sirens. And if you want to quiz us on what does a firetruck sound in Taiwan, we can answer the question. So beyond the sirens, there are a lot of other running sounds that are important for the driver or for the car to know about, things like beeps of backing up trucks, alarm bells of railroad crossings, which we're adding to the system. So that's it for me. I mean I have covered some of the additional AI building blocks we have been adding to the system. Those combined with those of Udo give a broad range of capabilities of the system, but that's to be assembled to have a single experience and that's what my colleague will be telling you. George will tell you about packaging now, how those can be quickly and rapidly packaged to deliver a full solution. Thank you.

[Audio Gap] mobility solutions, IoT technology [ build ] and Great China. So I've spent more than 17 years in SAIC Motor, which is biggest car OEM in China, has a joint venture with General Motors and Volkswagen. So I was Chief Architecture of the digital technology strategy in SAIC Motor headquarters before I joined Cerence last year. So I am going to talking about how to use the blocks to building the solutions. Next slide, please. So we call the solutions as modularized turnkey solutions because we need to address the problem or pain point of our customers. So today, the feedback from our customers are always talking about a time to market matter. And car companies realize they need AI digital solutions very quickly and sell them because of the digital transformation. So however, to order a [ seamless ] solutions in the market is very, very difficult. So sometimes we need a partner, a one-stop shopping partner, to help them integrating and building the solutions. In the market of China, especially the low end of the car, they need out of solutions -- out-of-box solutions and while customization for the brand and model. So our customer wants accelerated penetration and have this penetrated install base to allow future data modernization and also some new digital business. Next slide, please. Next slide, please. So as my colleague expand the technology before and also the product before, it's very mature and highly penetrated across the world. Every 2 cars, have 1 car is using us, the Cerence technology. In fact, we built the experience for 30 years, building very complicated AI solution in the car. The car, what I believe is this is the most complicated IoT consumer product. So I -- we invited to build ARK, which is AI reference kit. This platform is not the software-only. So you can see it has 4 components: software, best practice, data and operation. So I will describe it later. So our mission is to accelerate out-of-box solutions to mobility solutions that may be beyond the car. So it can mean any kind of assets like people on large vehicle. So we look at elevators, trains, 2-wheelers, buses and trucks and many other industries. And most recently, we have taken the ARK into the 2-wheeler very successfully, and 70-mile in scooter, and it's got a great feedback from the market. I will show the video later. So next slide, please. Okay. Next, we -- the ARK platform leverage our core conversational AI stack, and it's completed setting into these 3 module blocks, cloud, client and edge. The cloud allows us to deploy the real features in minutes without downloading the software -- without downloading the app from the app mart or OTA's head unit. The client allows us to take the traditional head unit from consumption to SOP in weeks instead of the months. And also the edge, it allows us to penetrate deeper into industry IoT, where it's low or limited computing power like the TCU, telematics communication unit (sic) [ telematics control unit ]. In order for us to achieve a very rapid time to market, we have put together and prepackaged the user interface, including dialogues that allow us to quickly customize the feature set and solutions. While leveraging our automated quality assurance, it will ensure a very high-quality end-to-end user experience. So this rapid penetration into the IoT devices will generate a large amount of data, which our ARK platform can capture, fuse, learn and provide the necessary insights. Then connected to our customers' operation back ends, they can easily find the new plan to [ monetize ] the data in their digital businesses. Next slide, please. Okay. Here is the slide for engineer. So it's a technology for those who are the engineer on this call. I want to expand the architecture. As you can see, ARK is a multiple operating systems across multiple devices. It leverages the advanced conversational AI stack that Christophe and Udo has discussed. But a couple of things I want to highlight here, I think, is very important. So Cerence is known for the great conversational AI technology for the entertainment, the infotainment in the car. But to the car, the car's future is not just entertainment, but also control and drive. So we have automated control mechanism across conversational AI and at least 200 control features in the car. So you can control the car just by voice without any [ bothers ]. So all cars powered by ARK that you can control your sunroof just using, "I need some fresh air." "I'd like to see the stars." Not just command, like, "Open the sunroof." So as you're driving the highway, with the sensor fusion of the car, you can easily say, "50 miles per hour." "Increase to 65," and et cetera, to start and set the cruise control without worry about the gears and the balance, which is always confusing my wife. Using the combination of those controls, you can use what I call is mode. It's like a smoking mode or reading mode. So one shot for multiple controls basing on user scenario. And this user scenario can be triggered by the sensor inside the car, the IoT sensor out of the car and the weather information, the traffic information, maybe the stock price. So it's modularized ARK. Our customer can adjust the focus on the combination, like LEGO blocks. So we have large ARK in multiple brands in China to enable them to build their own modularized turnkey solutions with the creation of the brand. So -- and also, we keep innovating our new car controls as we think that the car becomes more assisted. Next slide, please. So we are very excited to have brought the ARK to the work, and this is what I think is probably one of the world's first AI platform for modularized turnkey solutions. It encompasses our best-in-class combination of AI stack as well as the best practice of user experience and quality assurance. So we are able to empower our customers to use ARK to rapidly bring state-of-the-art IP solutions, including cloud, client and edge into the cars and beyond, the -- in these realities. It enables them to create their own ecosystem flexibly and operate it for their new digital innovation. So now I'd like to show you some examples. The new scooter from 70mai, a member of Xiaomi, so let's enjoy the video and back to Udo. Thank you. [Presentation]

Thank you, George. I apologize for the connection issues we had while we were going through Udo's section. As we adjust to the new normal of working from home, I guess we're all testing the bandwidth limits as kids work from home and dial in to their online classes and with everybody dialing in at the same time. So let's give it a try again with Udo's section and hopefully, we'll have a better luck this time. Udo, back to you.

Yes. Thank you. I hope the network will work for the next 5 minutes. I would like to resume where I got interrupted. I think we were pretty much finished with that slide so that should talk about what is inside, high level, on each of the building blocks, on the core building blocks. And in the next 2 slides, I have a more deeper view on 2 of those engines. So let's go to the next slide. So the first one is our audio AI or speech signal enhancement product. And this product is basically shipped with almost all car OEMs. It does signal enhancement for speech input as well as hands-free dialing in a car. As such, it also improves CarPlay and Android Auto acoustics. So it is very likely you use this product every day in your own car when making a phone call. The core modules of this is noise reduction, echo cancellation, beamforming, passenger interference cancellation, and the latter also allows to create a private audio zone, where interfering speech or noise from the outside is removed. So for example, you could switch off your kids on the backseat when having a phone call with your boss. Unfortunately, switching off kids only works for your boss in that case, but not for you, as you will still hear them loud and clear inside the car, of course. With deep learning applied to noise reduction, we now also improved removal of nonstationary noises like indicator clicks or wind buffet noise. On that, I have a quick demo for you, where you will hear 3 audio samples. First, the original mic signal, then the current state of the ARK, and finally, our recently launched products. So Rich, if you could play it. [Presentation]

Okay. Thank you. So now I would like to move your attention to the next chart on the right side. It will show you another innovation driven by Cerence, where we solved the problem of interfering speakers. It almost happens each time I want to enter a destination in my car, that my wife starts talking to me while I still talk to the voice assistant. In the past, this was a recipe to fail. Now I am safe. My wife still can't wait to talk, but I get the input entered successfully to the voice assistant at the same time. So in addition to that standard signal enhancement task, we also released what Christophe already mentioned now about emergency vehicle detection, so that is all powered by this audio AI product that we have in the stack. So next slide, please. This is our new high-end text-to-speech product that we launched in June, with great improvements due to transformer-based networks and generative voice nets. As you can see the chart on the right side, it is the best core TTS, text-to-speech globally right now in the category long-form reading, also compared to Google, Amazon or Microsoft. News agencies have approached us already to talk about newscaster product, and we did sign a deal with a car manufacturer. Responsible over their set, I can't stop listening to this new TTS voice. And you heard a sample from Christophe already. So the software architecture of this TTS engine is divided into a front-end on the left side and the synthesis part on the right side, both with different network architectures. The front-end does prosodic and linguistic analysis like pronunciation guessing or pause prediction using word embeddings. It also includes encoding of various speaking styles like news documentary. And the synthesis part in itself is now completely generative and not unit selection anymore, so completely generative using wave recurring neural networks way far in it, okay? With that, I'm moving on to my last slide, please. All those basic, core tech building blocks form our conversational AI platform, which is clearly the most complete stack with the broadest language coverage compared to any competitor out there. A user request enters the system on the input side by speech. And we get what was said, in which language it was said, who said it and how it was said even, including the sentiment. So on the natural language processing, NLP side, we understand what was meant. And who was addressed in case there is multiple assistance active. We will interpret all that, of course, as well in a context as part of the conversation. And the conversation engine then is there to predict what the user wants to do next instead of asking a lot of questions as the best UI, user interface, is the one which has least interaction with the user and still get the task done. So for example, we can predict that a user wants to have a covered parking place in case it is raining even though he did not mention it. It learns such behavior from all the data we get and collect. So in formulating the reaction to the user, we use, again, natural language processing to generate a prompt, based on this context. And finally, the output is synthesized by the text-to-speech in an appropriate speaking style, matching the emotion. In addition to those pure speech input/output, you heard already from Christophe that we can give feedback in different modalities to the user. And -- for example, some feedback that Christophe did mention is what we also implemented now in the new S-Class for Mercedes. If you say there, I am stressed, we'll simply turn on the message seat. That's also a kind of feedback for the user then. Okay. And with that, I'll hand it over now to Nils, who will talk about new applications.

Thank you, Udo. This is Nils, pleasure to meet you all and speak to you. I joined the company actually 25 years ago after completing my PhD in AI. I now have the BU for applications that we started recently. And typically, when we talk about these apps, we talk about the new business models attached to them, like SaaS models or revenue share. Today, I'm actually going to talk about something different. Next slide. I'm going to talk about what I call the yin and yang of successful apps because if you want these business models to work, you need user adoption, right? You want people to pick up these apps and use them. And one important thing for that is a great UX or user experience, as we call it. That means a great design, but also a process leading up to it with surveys, usability studies, testing prototypes and all that good stuff and my colleague Adam Emfield will talk about this right after my talk. I'll focus on the second aspect here, which is high-tech actually baked into the apps to make them defensible and make them work actually. So we look into this now with a couple of examples. Next slide. So the first step I'm going to talk about is called Car Life. It's basically a companion that works with you throughout the lifetime you spend with your car. Actually starting before by helping you to select a model, booking a test drive. And once you have your car, it proactively tells you what the great features of your car are and how they work. It allows you to ask the car questions about its operation, a smart car manual. It integrates with dealerships for service appointment scheduling and all that good stuff. And it also allows OEMs to push offers and new products to the car. So it's really a great thing. But there is also a good technology baked into it. And I'll pick 2 examples here. I'll talk about the car manual questions and about the proactive feature teaching or onboarding. Next slide. So starting with answering questions drivers might have about a car. It's relatively easy to pick the frequently asked questions, what we call here the short-tail or head of the curve. These are questions that are asked again and again. There is a limited list of those, and you can compile good answers and question-and-answer pairs for them relatively easily. But what about the long tail? The many, many more questions some users ask, and you still need to answer in order to have good experience. And therefore, we bring in deep learning techniques. To come out with a technology we call question answering, which is basically passage retrieval from unstructured texts. So basically, we take texts like the car manual or other texts provided by the OEMs. And we use technology to find the relevant passages or relevant paragraph and sentences that answer the question of the user and answer those. And that allows us to combine these 2 elements to frequently ask questions and the long-tail question answering to basically cover all the questions and good quality and have a good experience. Next slide. The other aspect I mentioned is proactive feature teaching. By that we mean imagine you enter a highway, you have a new car. In 3 months, the car tells you, "By the way, I noticed you've never used lane assist as you're entering a highway. Do you want me to teach you how to do it?" And in order to do it, you need to have this proactive behavior in a smart way. You need to come up with when and what, right? You need to decide what to teach you about and when is a safe and good moment to do it. And we have built the machinery. It uses the same reasoning capabilities that both Udo and Christophe mentioned that we use also for smart domains and make this reasoning about covered parking that Udo mentioned. We use the same to make decisions about the right moments and the right content. And what you see here is a visual relation tool we built, in order to explain the machinery. You can see how we use factors like center input, like is the driver occupied, what is the emotional state. You heard from Christophe how we can measure that. You have information about the driver. You have the content you want to teach. And with all those inputs, we select the right content at the right time. Next slide. Here is the overall architecture. You will find the Q&A engine and the FAQ module in the center and a bit to the right, as I covered before. And a proactive module with the AI in a reasoning framework you see in the bottom, and together with lots of other components from our standard architecture plus components, specifically for Car Life, like integrations with dealership databases for scheduling appointments and so on, we build you a whole Car Life experience. And you can see here how a relatively easy application, easy to use and easy to understand what it does, underneath, has quite complicated architecture in a bit of high tech baked into it. Next slide, please. Okay. So the next app I'm going to talk about is called Cerence Pay. And the usual punt I do hear is it's all about payments. But everything we do is not the payment itself. Yes, so this is centered around something like an e-wallet payment gateways that we believe already exists in the OEMs or we can bring in from partner. What we build is the experience around. We bring partners that have content like parking providers that we announced in the recent product launch here. We bring the cloud and hybrid architecture we already have in place. We're already in the car, so to speak. We are connected to the cloud and the car. And we bring the user experience that makes it a great experience to pay by voice. This is what it's all about. You can pay by your voice, which is different from starting an app manually on your smartphone or even on the head unit. Instead, you can do it all by voice while still driving, and you will see a nice example in the video that Adam is going to show you. I'll focus here on one aspect, top left, the authentication bit. Next slide. You already heard about voice biometrics from Christophe. Again, here, we use it for the purposes of authenticating a person. So to know that a payment is legitimate and they are who they claim they are, you can see it on the bottom left. And the overall picture here is an architecture we came up with Visa to address a problem that all of Europe has. There is a new payment services directive coming into effect by end of this year, early next year, which means all payments you do by mobile methods on your smartphone, in a car or other means have to have strong customer authentication, which breaks a lot of the current user experience as you know. Instead, if you don't do anything, you have to go back to the issuer of the credit card. They have to authenticate you. And then you go back to the shop you're using, et cetera. It's very cumbersome. And in a car, it's especially cumbersome because you cannot send a pin or something or a token to a smartphone. You cannot pick up your smartphone and a enter pin. Instead, we solved this by voice biometrics, and under a scheme that Visa called delegated authentication. Actually, the OEM in the car can authenticate the driver using our voice biometrics and the second factor, which can be the ownership of the car, and then there is no step up needed to the issuing bank, et cetera, which solves this problem. It shows how quite a simple feature is enabled by technology. Next slide. As I mentioned, the whole payment experience is centered around voice. And again, we have productivity here. I mentioned productivity in car life, where we select a moment in time to teach you features of the car. Maybe use productivity to actually make you discover that you can pay by your voice. You don't have to discover an app or it's somewhere in a menu in an app. Instead, we start with requests like navigation requests, which still account for 40% of the voice utterances people do in a car. And we nudged people into understanding. They -- when they navigate to a restaurant, they can make a reservation there. If they navigate to a gas station, they can actually start paying for it while they're driving. And why that works and how people like it, you will hear now from Adam in the next section.

All right. Thank you, Nils. And hello, everyone. It's my pleasure to be able to talk to you today. I wish we could do it in person, but we'll make the best of this as it is today. So I'm going to be talking about user experience at Cerence. My role is to lead the user experience and user interface design and research teams. We have a team called the DRIVE Lab, just the design, research, innovation and in-vehicle experience team. And our job is to be the voice of the user at Cerence and to be able to help us work towards the products that will create these experiences that users find endearing, they want to come back and use again and again. Next slide, please. All right. So we talked a lot about artificial intelligence and the technologies we build today. And if you -- we were presenting this at an AI conference, someone would inevitably get up on stage and go, what does AI mean? We can't agree on a perfect definition. We're all here for an AI conference but we disagree on what it means. Well, my job is to ask what it means to the user and what they want us to do about it. And when you think about what it means to the user, they start saying things like, well, AI means you have a system that can learn about me, and I can do something with that. It has access to the cloud. It has access to big data. It has access to all of this information that allows it to make these decisions that might be too complicated for me -- myself to handle, and I may be able to solve things that I don't even know I need solved. It may be able to get to know me better than I know myself and help to be proactive and do things like teach me things that I need to know about my car, as the example you saw before. And so click for me from this side once. But if I really want to summarize what it is that users come to when they talk about AI, it comes down to them saying, I don't really know what it is. It's really, really complicated. And for us, it's not a bad thing at all because the users don't need to know what's under the hood necessarily as long as they have a great experience with the way it comes together. Next slide. So AI introduces a huge opportunity for complexity. We're talking about all these building blocks today, all these parts of our technology, we're talking about ones that our content partners have or that the automakers might bring in. There really is a risk of having so much complexity that you end up with clunky experiences. And if you think back decade after decade, you've seen this in the evolution of technology, where things felt like they were bolted together, often crudely, and sometimes this even seems subtle and wasn't clear at the time. If you think back to the automotive industry, going from some of the oldest automakers in the world like Daimler and Ford, you'll see that even the addition of a radio in the earliest cars had some user experience issues. They took 2 great technologies, fused them together, but ended up with a problem where you have all these identical knobs that are next to each other, and you're driving. And so what happens if you want to turn the radio up and you accidently grab the knob to tune the station differently? Again, it seems like it's subtle, but it didn't take into account the fact that while users are driving, they're going to be fumbling with their hands, hopefully, with their eyes on the road. And that in and of itself was just a small issue that later was fixed. And there are countless examples we could go through, but coming to where we are in the new way today, fortunately, for the end user, the vast majority of technology is moving more into a user-experience-first mindset. This doesn't -- this includes the automotive industry as well, where people start to think around, okay, how -- what does the user need first? What is the user journey going to be like? And how are we going to make that a reality? And the way I like to talk about how we do it at Cerence, in particular, is to bring art and science together. The art of the pleasant designs, the nice sounds of the voice, the pretty user interface you put in front of you, and the science of the technology. But further than that, as a team with people that come from backgrounds, from communications to psychology to engineering, we also want to understand the science of the user. We want to know how they learn information effectively. We want to learn what captures their attention and what doesn't, especially when they're driving. So our job is to try to bring that all together to create these better experiences for the user. Next slide, please. So let me give you an example of some of the ways we get at that, and I'll talk about some data we've got from this as well. It's really a good idea to test an idea before you invest in the engineering time and money to make it a reality. And one of the ways that my team does that to support our product ideas at Cerence, is that we will mock up something; and see here an example of a restaurant reservation prototype, where we can do studies such as Wizard of Oz, which means we're behind the curtains, pulling the levers that are making the machine do actions and see how users respond, find flaws in our ideas to throw out bad ideas or find new ideas from them, without them ideally even knowing it's a baked system. There are many ways we go about this. We can test things, but just wanted to give one example of how we take pride in putting these concepts in front of users as early as possible and then continuing to refine based on feedback from those users to make sure that we aren't overwhelming them with complexity by creating more simple experiences. Next slide, please. So as a quick case study. Cerence Pay is no different than the other products where we want to lead the design process and engineering process through research and surveys and usability studies and that sort of thing. In the case of Cerence Pay, I just want to share 2 graphs here that can give you an overview of some things we've learned that help us understand that we believe that we're building a product that users want, they report that we are, and help us prioritize how we're going to tackle it. So on the chart on the left, the -- for the preferred payment method, when we help educate people around what the process looks like, what the benefits are for being able to pay by voice and potentially any trade-offs with any of these, we find that the majority of people say they would to make payments in their car for various things, you'll see on the right, by voice, more than any other method. And if you look at those particular methods, paying for parking by voice, gas, drive-through, et cetera, you'll notice that there's a good chunk of those where if you look at just the blue colors alone only, the likely and very likely, we're talking huge margins of people report that they're likely to do this. And because of that alone, we have some guidance here on which directions we would start with Cerence Pay and where we would go next after that. And it's something we'll continue to ask and find out and refine as expectation changes and the product evolves. Next slide, please. So let's bring it all together then. As I wrap this up here, I want you to see a vision of what we think Cerence Pay should be. That was built off a great clever design but also built off of research and will be put together with some of the best technology we have available to us. Do you want to go ahead and play that for me? [Presentation]

So even if in-car payments aren't the first thing you think about as to where a company may go to improve the experience, I hope that shows that we're looking for all different ways to create these experiences and simplify them for the user in the end. So with that, I'm going to hand things back over to Prateek.

Thanks, Adam. I look forward to using that in my next vehicle. It's pretty cool. So with that, we'll move on to our fireside chat section. Joined with me today is Charan Lota, Chief Engineer Connected Technologies, Toyota North America for our discussion on role of AI in the future connected vehicles. Thanks, Charan, for your time today, and we're honored to have you as a guest here today for this virtual chat.

Thanks, Prateek. And interesting presentations, I'm really enjoining it. Thanks for having me.

Great. Great. Well, first of all, before we get started, as a disclosure, I want to state that my first car was a Toyota Corolla. And since then, I've owned a Lexus and a Camry hybrid in my life, so I'm surely impressed by the technology, the quality of the work you guys are doing, so just as a background. So Charan, as we talk about connectivity and electrification, which is the connectivity and connected car is the buzzword today, as a chief engineer for Toyota, how critical is connected car to Toyota's future products?

Yes, good question. Connected cars are basically not an expectation, right? They're no longer viewed as an optional feature. I think when somebody goes to buy a vehicle, they fully expect it to be connect and if it's not, there's going to be a disappointment, right? The reason is that connected car, it enhances countless user experiences, everything from safety, freshness of data as well as providing convenient time-saving features, just like the video that we just showed previously. I was watching it, to me, who wouldn't want a feature like that, right? So overall, I strongly, strongly feel that connected car is key to the success of our brand, Toyota. So yes, for sure.