#291 Developments in Speech AI with Alon Peleg & Gill Hetz, COO and VP of AI at aiOla

Hi, I'm Maggie, VP of Curriculum at datacamp, and I know firsthand how overwhelming it can be to keep up with the rapid changes in data and AI. But staying ahead doesn't have to be hard. That's why our curriculum team has curated the best learning content to help you and your team stay ahead, covering everything from data and AI literacy for busy executives to cutting edge generative AI for developers, and everything in between. Whether you're upskilling yourself or your team, we make learning practical, engaging and flexible. It just works. Power your data and AI transformation journey on datacamp.com.

I believe that every machine interface is going to be voice interactive. We are going there, but we are not there yet. And I think that as long as we can integrate the jargon, as long as we can integrate accent, as long as we can get the background noise, we are in the right path to achieve it. When was the last time that you called to a center and a human really answer to the phone? It's never happened, right? We don't want to get into a point that we are talking to something and we don't even know if this is a person or a machine, right? So for that, again, what we have, at the same time we're building models to identify those models because you can still replicate voices, but the way that they're generated, you can still see some ways to to differentiate it between a human to a machine.

Welcome to Data Framed.

This is Richie.

Large language models have stolen a lot of the attention recently, but typing isn't always the best way to communicate. The fact that you're listening to a podcast means I know you know how useful speaking and listening are. Today we're going to explore the latest developments in speech AI that includes speech to text AI where voices are transcribed, and text to speech AI where you generate voices from a script. We've had speech to text AI for a long time, but I want to know what progress we've made with dealing with background noise and strong accents and languages that aren't English. Similarly, text to speech AI has come a long way since Stephen Hawking's robot voice. I want to know how close we are to having completely natural sounding voices and what the consequences are of AI being able to deepfake anyone's voice. Finally, it'd be nice to know what are the best ways of adopting speech AI at work.

I have two guests for you, both.

From speech AI company Iola Alon Peleg is the Chief Operating Officer where his focus is on aligning Iola's speech AI tech with customers needs. He's been a technology manager for 25 years with stints at Cisco, intel and Wix. Gil Hetz is the Vice President of research. He works on improving the speed and accuracy of Iola's speech AI models and products. He joined Iola from the petrochemical AI company QRI after being a researcher at Texas A and M University. So between them, we've got expertise on both the business and technology sides of speech AI. Let's find out about the cutting edge of this technology.

Hi, Alan. Hi, Gil. Welcome to the show.

Hey. Hey, how are you doing, Richie?

Life is good. Really great to have you here. So I guess just to begin with, can you give me an overview of what speech AI is? Gil, do you want to start?

So, you know, everybody's talking about AI, probably the most used and abused letter in English lately, which referring to artificial intelligence. Speech AI usually is the process of taking, understand, generate human speech and understand it. The key components here usually are the ASR model, the nleu, the ASR, basically taking the speech, speech itself and convert it into text. The nlu, which basically understands what you are saying, and then the tts, we generate the response for you. So that's basically the three components that build the entire process of the speech AI.

So you can go in both directions between speech and text and then you can also have, well, I'm just understanding what is that speech without necessarily having to switch into text as well. Okay, cool. So maybe we'll cover all three of these things. Let's talk about some of the use cases. So what are the business use cases? Let's start with speech to text.

Then when we are talking about generally speech to text, a lot of the time when you are talking about Siri, they need to understand what you are talking. Alexa, all the virtual assistants, call centers, healthcare, education, everybody's basically, when was the last time that you called to a center and a human really answered to the phone? It's never happened. Right. So the fact that you can speak to somebody and they understand or at least translate it to text and then you can take this text and start working on it. And probably we're going to talk about that deeply in the next sections. But the idea is really taking the valve that I'm actually making and translate it to text information. But at this point it's unstructured. It's unstructured. And if you want to make sense of it, you need to process it.

Okay, yeah. Call centers just seem like an incredibly important use case of this. But it seems like anytime you want to transcribe text, then this is where the automatic speech recognition, I think it was, it was ars, you said this is the acronym. That's where that's going to come in useful.

Yeah, ASR, Automatic Speech recognition. And you know, when we are talking about that just a little bit, because we're probably going to go deeper on that, there are challenges when we are talking about speech. AI and I would basically focus and in this conversation we'll try to talk through the main three topics that really are the challenging ones, which are the accent. Right. I have my accent, you have your accent, Alon has this accent. And you need to understand that there is the background noise. Okay. What is happening when somebody is speaking behind you, there is noise behind you. And the last one, probably the most important one, which really changed the entire way that we are looking and generating results is the jargon, the complex English or any language that you are talking that general ASR does not understand.

Okay, yeah. So certainly a lot of challenges and I definitely want to get into these in more depth. So you mentioned things like accents and background noise, things like that. So we'll cover those. Before we get to that, I'd like to talk a bit about like what are the use cases going the opposite direction to when would you want to generate from text?

So we talked about like I'm taking speech, you are talking to me, I'm generating the text. And then if you want to talk about text to speech, where you basically taking the text itself and generate response that sounds like a human, like I'm going to type in a sentence and you're going to try to the machine and it's going to generate basically speech. So you know, in this case, a lot of the customer interactions that we have, always somebody is sitting behind the scenes and just talking to you. I mean the machine is talking to you, but you get the response for that accessibility. Any assistant, virtual assistant, usually getting the response first in text and then generate the speech from it. Any education, training, elearning. So if you are trying to learn a language and to talk in the right accent, usually you are using text and then it goes to some of these machines that really pronounce the things the way it should be. And a lot of the content marketing ideas that people have, they just throw in text and then generate it in a way that people really attract the ear of the audience and make them at least pay attention. And that's what marketing is all about. If I can get your attention for a few seconds Few minutes I get something.

Absolutely. I have to say that education example is really interesting because if you're generating the speech from text with a machine, then it's going to be more consistent. And this seems really important in the educational use case because you want it to have exactly the same accent every time, you want it to have the same cadence and whatever. Whereas humans are going to vary a bit in how they talk over time.

Okay, cool.

So actually maybe we just need some motivation in terms of success stories. So Alon, can you talk us through, Are there any businesses that have made good use of speech AI and had some success?

Of course, and I'm happy to share a few. Let's start with the food retail industry. They have a lot of regulations. They need to for example, check the temperature in every fridge at least twice a day or once in a shift. It takes them a lot of time. From one of our customers, it took them around one hour. And with voice they just going or even running through the store speaking and the system is capturing all the temperature fridge in the store. If there is a deviation of from regulation, immediately an alert being sent to the manager of the store to take immediate actions and to make sure that is comply with the regulation. So the time it takes today with a speech AI system like ours reduced by 55% and for this industry where the margin is not too high, maybe 3%, it's a use saving both in terms of the cost of the worker as well as the reducing the regulation risks. And that's one example. The second example is from a different world. It's from a pharmaceutical manufacturer. We have two examples there. One is the technicians. The technicians are very busy and they have a legacy system that is sitting on a PC somewhere in the warehouse or in the floor. They are expected to report every issue they fixed to give more description, as much description as possible. But they're very busy. So okay, I fixed something. I fixed maybe 10 things in the last hour. Then I need to return to my office and type all the information of what we did. We, we saw that to record a task they made and the average words in their description was actually 4, 4 words, 9 minutes to report with voice. First of all, they will report it real time. They don't need to go and type anything. They will just speak to an application. The application will record exactly the description of what they did. They even can take a picture real time of what they fixed. They can also scan a barcode and then they know if the part that they fixed is under warranty or Not. And we found out that they cut by half the time to report tasks by half. And the amount of data they capture is in average 11 words. So from 4 words to 11 words in 50% of the time. So it's a huge value to the technician department. On the same factory. They have also quality control departments that are doing changeover. Every time they change the products that are being manufactured, they have a process. They reduce the changeover process from 2 hours to 14.5 minutes. And they did it mainly by reducing wait time when there is an issue. First of all, again, they don't need to go to a corner in the floor and they can report it real time. The inspector that is coming immediately because real time is getting. Hey, there is a problem that was reported now on the production line. Come to fix it now. So the wait time was reduced dramatically on every shift. The amount of pictures that they are taking, which is a record and visibility to what's going on on the floor increased from 10% to 40% in the event of an issue. So they reduced dramatically the time in almost 90% and they increased the amount of data that has been captured. Let's proceed to a jewelry manufacturing, a large one in Hong Kong. They are doing quality control for products. It took them 175 seconds to complete a check. And they were Reporting in average 4.5 returned product. Product that has an issue and need to be returned. Every miss like that will cost a lot of money because it will go to the stores with damage. So they, they have to do a very thorough work and complete the inspection and to maximize the amount of bugs or issues are finding in the items. So they reduced the inspection time from 175 seconds to 57 seconds. And the amount of returned items, which is a good thing, grew from 4.5 on average to almost 7. So the quality perceived by their customer will increase. They can cover much more in terms of. To complete much more product in every shift. So. So that's insane in terms of value and ROI for this company. Then we can move to fleet management that they have like about 100,000 cars to check every morning to verify they will not be stuck on the road because then it's very costly. It took them with paper. It's actually paper. It took them about 40 minutes to complete every inspection with voice. It took them. And this is the name of the check. It's one minute check. And it took them one minute check with voice.

Yeah, it's called 60 seconds.

Yeah.

They have 40 items they need to go it really in 60 seconds. And the only way to do that in 60 seconds, as Alon just said, is by using voice. But I think that Alon just went through four different cases. We have many more, but each one of them in different language. Each one of them is really what we are doing as a. We bring the speech AI to the fingertip of the people in the ground. And that's what is important. That's what is nice. The only thing you need is your smartphone. You have everything you need to speak to that and collect all the information and really make it happen. And you know, you have heard the numbers. The numbers are crazy. The first time when we saw it, we couldn't believe because Sachin improved just by switching the way that you are collecting the information.

I want to finish with two examples that just to give the audience a point to think of, think about the cultural change that you are making in an enterprise when you are allowing every employee to report a safety issue by voice. You have your own internal application. I'm an employee in an airline. I'm going through my gate. I see some coffee on the, on the floor. I don't need to open a form or to write a letter. I'm just opening my internal application and I'm reporting. Hey, I'm in Newark. I see in Gate 90D or C a glass of coffee on the floor. And it's been captured and immediately been sent. It's actually changing the culture of the company to A. Everyone can and should report very easily safety issue that can risk our brand. So let me end with that. I have many more, but let's keep keep some time for other questions.

Yeah, lots of examples there and I'm sure there are many more. But I think the consistent theme here sounds from these examples is that these are places where you haven't got hands on keyboards. Like, I guess with my day job, I spend a lot of time at a laptop. It's like, well, I don't really need speech AI so much. But if you've not got your hands free, if you're in a shop or you're in a factory or you're in some sort of manufacturing place, then yeah, you want to do things without having to go and find somewhere to type things. So this seems incredibly important.

I want to give you even an example because I'm doing is more like text to speech. Okay. I have a, a big presentation tomorrow for investors. Right. I written all the points I want to cover and I know that I have a commute of two hours back Home and I cannot open a paper and practice. Right. In few seconds I translate. I changed the text to speech and used leverage two hours to just listen to what I want to say tomorrow. Think how much even us that we are like mostly using our PC. Even to us it's very useful too.

Okay. Yeah, I like that. Just if you've got a script or something, have it turned into speech, then you listen to it back and you you're probably going to decide does this sound right or not. It's much easier to determine whether it's good or not compared to just reading it. Okay, that sounds like another cool use case. Okay, I want to. Since we were talking about sort of shop floor use cases, what happens when you've got background noise? How do you deal with that? Is it still a problem or is this something that's been solved?

No, no, it's a problem. And you know, I think the main challenges as I mentioned are three of them. And background noise is definitely one of the things that you need to make sure that your model can handle. I mean there are different ways to do that. I'll talk in general about our model for a second and then we can talk about how other people try to at least approach to this problem. So as I said, I think the number one problem that we are actually dealing with is a journey. This is the problem that it's not just English that we are talking right now. It's English in a specific industry, in a specific site, in a specific slang. It can be different from one site to another even if they are working at the same company. And that's the number one problem that we are dealing because when we are going to general speaking English off the shelf models can do a good job in terms of word error rate that they are producing when they are translating to text. That's the word error. And then when you go into the jargon you suddenly get down to about 40, maybe 60% error rate that you cannot deal with when you are trying to report and non checklists that really that important in terms of reporting that something that you cannot get less than 90% accuracy. It's not good, it's zero or one. Now that's the first thing that we are dealing with and that's why we have actually two models that working together. One is the ASR model which is general and the second one is keyword spotting when we are injecting the model all the jargon. And this is a zero shot. It's basically looking at the way that you are generating the word itself, the signature of the acoustic sign and try to pronounce it right. So even if I use a word that makes no sense in English, but I type it in in the right format, I will get it from this model. And this is an item that you might do. So this is the first thing. The second one is the accent and background noise. When I'm dealing with area with a very high DB noise, I really need to include it when I fine tune the model. The way that we are dealing with it is we are generating, we have our own TTS model that can connect the background noise so we get a sample of the background noise, we have the text that we are generating and we are generating the voice with respect to the background noise. So if you are just taking the background noise and what you generate and does do the super position, it's not going to be good enough. Because when I'm talking loudly, there is a background noise that probably louder behind me, right? This is the reaction that I do because there is a background noise. I'm increasing the amount of DV that I'm generating when I speak. So that's what our model does. And then we use that to fine tune the ASR very quickly. So that's how we are dealing with the background noise, the jargon. And last thing is the extent. And again for that we have in the tts, the voice cloning, it is taking your own voice, include those specific samples and generate the response. And that's how we really dealing with different challenges that we have. We've seen them in different languages, actually in languages that are not English, we see a huge improvement because of the fact that those models, to start with, they're not good enough. But when you are improving them, you are getting much better results.

Okay, so lots of different challenges there. So you talked about background noise, you talked about jargon, you talked about accents, languages other than English. I'd love to get into some of these in a bit more depth. So maybe with the jargon one, because you said that was the biggest issue and certainly like once you start talking about very technical terms, quite often these sort of just off the shelf AI models just don't have a clue what they're talking about. Certainly whenever we transcribe these data frame podcast episodes, anytime we have technical terms, it's just absolute nonsense. So if you know what jargon you're going to be talking about, can you just provide a list of words and say these are the things that we're going to be talking about? Does that Improve things or what do you need to do to get better results?

Exactly. So the way that we are doing that is we go to a specific site, we collect the forms that they have, we extract the information, the keywords that they are using, the jargon itself, really the words that the name of the items, the name of the people, sometimes the name of the events that they are reporting on, and include it in our keyword spotting model. This model is in charge of integrating the jargon itself to the general ASR. The fact that we are actually having a specific model that is zero shot. Zero shot, meaning that you provide me the list no matter the language, no matter what kind of size of the list and I'm going to reproduce those words. I know I'm going to tell the ASR. In this sentence, I think you use this word which could be a word that does not make any sense in English or in any language. It's just going to be the signature of the acoustic sign of the pronunciation. And that is going to go into the ASR model and increasing the probability because all those models its probability. What is the chances that I use this word and this is going to increase the probability for a word that actually came from the jargon that you are talking about.

It's important to say that it's not a process of months or years. We are generating such model in a matter of minutes.

Yeah, it's zero shot, meaning you provide me the list, you got your own model, you provide me the list in English, you provide me list in Spanish. You get it in Spanish.

That's very cool. So it's not some kind of extensive process of like I need to go and fine tune my own model. It's just like these are the words and then.

Exactly. So I think that's what I said. I did not touch it right now. But what people are doing in general, they are taking examples and fine tune their own models with some example from the jargon. Now this process takes time and usually you need to get some data, you need to generate data. And that's why I think our solution in that sense make the advantage that we have in dealing with jargon.

So it sounds like jargon is sort of like a largely solved problem. So the next one you mentioned was accents. Now I'm wondering are there any accents that are particularly difficult to deal with? I'm thinking like in Britain there are a lot of like cities with very strong accents. So like Glasgow is maybe like the most common example of things that foreigners struggle to understand. So are there Some hard accents to deal with and some that are easier.

Yeah. So the fact that English is actually, I think is a great example because there are so many people that speak English but in so many different accents. And what we are doing, we are trying to collect samples from people and we don't need a long samples, we just need a few second samples from specific people in the site and then we can generate text, any text. The fact that we actually have our own TTS model that take this audio of me speaking with the text, I can generate so many different examples to fine tune the model for a specific access.

I wanted to add that we have also examples from French Quebec, which is difficult and we have evidence that French people don't understand French in Quebec and we have customers there and we are reaching very, very high percentage of accuracy.

Yeah, you can say 98% accuracy, that's pretty much the number which is phenomenal for language that is not that common in terms of the accent.

So 98 sounds amazing. Is that, is that good enough for every use case? I don't know whether like, are there some use cases where you need like really close to 100% or is 98.

Good enough for most purposes, usually 98 is good enough. So again, you know, when we are talking about speech AI, we are talking about three different components or maybe two main components. The first one is taking the speech and convert it to unstructured text. This is the ASR model getting 98% and then you need to have another model that, the NLU model that process the information and put it in a structured way so you can really use it to make decisions. Now for us, from what we have seen, when we are getting to north to 95% we usually are in a good position to move to process the data. It's just like me and you speaking and I'm sure that in this specific conversation I made couple of mistakes, but you still could understand what I said. And the fact that you could understand it because you don't need the 100% accuracy because you understand the general language and you know what is the context of the conversation. And that's pretty much the same process that we are having when we are talking about models.

I suppose that's one of the good things about speech is you can stutter a bit and say the wrong words sometimes and still it's understandable what you're talking about.

Exactly. Let me give you another example. I'm sure that. Are you using ChatGPT? In a way, yes. Everybody, that's okay. That's another way to get your AI to your fingertips when you are typing in something to chatgpt, sometimes, you know you have mistyped something. Are you going back to and change it? Not really. I mean just send it, it will understand. And I mean I used it many, many times. I made mistake by typing and I still see that I get what I need. So that's what I'm saying about the AK studio.

Absolutely, yeah. I think my ChatGPT instance knows very well how badly I type. So yeah, it's good that with all these mechanisms the AI is smart enough to sort of correct mistakes. Wonderful. All right. So related to accents, I suppose the other thing is how well do different languages work? So I mean English is just spoken so much, there's so much available that it seems the models work pretty well. I don't know how well it works for other languages that are less common.

So English, of course, the more examples you have, the better accuracy you get. As I said at the beginning, we know how to bring the jargon to a very, very high level of accuracy in different languages and our ability to run a quick fine tune on the specific examples that we are generating from the field actually bring us to a very good result. We had that in Chinese, we had that in Portuguese, we had that in French, cubic French, which is differently from French. So we have seen that. And the fact that we understand the jargon actually allowing us to get very high accuracy in languages that are not English.

So most of the languages you mentioned are sort of like, I guess, top 10 most popular spoken languages around the world. Like how deep into unpopularity can you get before it stops working?

In Indonesia?

Yeah, we also had Indonese, which is really not easy languages. Thai, we also got very, very good results there. Hebrew, another one you don't really have.

You got a vested interest in making that one work.

As I said, you know, when you are working in a systematic way that includes the jargon because you know what you are looking for when you are including the background noise and you can fine tune to the way that really people pronounce the language itself, you can get a very good result. And that's what we do. I mean this is the magic that we bring to the table.

That's very cool that it's opening up to a lot more languages and speech AI is going to become accessible to many, well, I guess a lot more people around the world once, once all the different languages work. And since translation is one of the big use cases of this, having it work in different Languages is going to be incredibly important actually. Yeah, do you want to talk me through that? Are there some good use cases around translation here? Going from speech in one language to speech in another language.

First of all, if you're looking at trying to fine tune your models, you definitely want to have in a way your TTs in a very good shape because you are going to use synthetic examples that you are generating to fine tune your ASR models. Now when you are going through the process of trying translation, this is usually happening in the middle box which is the NLU process and the task that you are trying to solve. Test this specific sentence and convert it to the same sentence in the different language and then run it through the TTS model to generate the translation that you are running. We are as Alon just mentioned, we are working with companies all around the world and at the end of the day they are collecting information from different countries and they need it in a very systematic way to save it and to have it general. So they need a lot of the time to translate everything to English even though they are speaking different languages. And we have this capability which basically happening in the NLU model. I know I'm still using the ASR to understand the people on the ground. I take it same way it's happening in your head. Translate it to the different language and then processing in this language that I.

Want to send the information feature for global workforce. So if you go to a car manufacturer in Singapore, it will have their employees from all over Asia, they can speak natively on with their language. The backend will probably data will be captured in English or in any or Korean or whatever they decide.

Yeah, this seems incredibly useful. I was just sort of thinking if I go on vacation somewhere, I just want to be able to say things and have it automatically translated into whatever the local language is and vice versa. So I can understand what's going on. But yeah, in a workplace it does seem even more important, particularly once you've got a lot of people from different countries with different languages trying to communicate in the same place.

So just to comment on that, this is what's actually happening real time going through your ASR in your language, NLP or nlu basically translate it to the language that you want to and then run through the T.T.S. generate your voice. And if you have your voice, your example in your tts, it's even going to pronounce it in the same voice that you have. Accent is probably going to be a problematic unless you speak the accent provided. But in general it is going to be okay.

That's very cool. That just sounds a lot easier than spending hours on Duolingo trying to learn the other languages have automatically converted. Okay, cool. So I guess on that last note you mentioned about having my own voice there. So it does seem like natural language or more natural sounding voices have become a thing in the last year. Can you talk me through where we're up to here?

Yeah, absolutely. So, you know, in general I'm going to focus on the last block, which is the tts. And for that you have heard about, I guess everybody on voice cloning, basically taking voice that is natural and then use it to generate TTS models. So basically you constrain your model, the tps to the text that you are providing and the voice, the audio voice that you are providing, and then generate the response we see today. And again, I'm going back to call centers that we all talk. I mean, when you need some service, if I go back five, 10 years, it was too robotic. You know, it was like, hey, how are you doing? Please press one. And now when you talk to people, you can even hear them the way they are actually breaking up the words and think over words when they are speaking. So I think there is a huge improvement there. We spend a lot of time on creating our own model that can integrate that specific voice or any voice, including background noise, because we need it to train the models. But also in general, people want to use that and have it as part of what we are producing. So we made a huge improvement in the last few years. And I think very little time from now we are going to be in a position where it's going to be hard to differentiate than a person to a model when it comes to tds.

Okay, this is pretty interesting that we're going to get to the point where it is difficult to tell whether it's like a real human voice or whether it's AI generated. So I suppose this leads to a natural problem in that you've got voice cloning deepfakes there and it's unclear what happens there. So do you have any concerns around this?

We all should have some concerns about that in the general view because we don't want to get into a point that we are talking to something and we don't even know if this is a person or a machine. Right. So for that, what we have, at the same time we're building models to identify those models because you can still replicate voices, but the way that they're generated, you can still see some ways to differentiate it between the human to a machine. Think about the emotions or breathing or natural flow that we know how to generate models and we know what kind of pattern we are expecting to get from models compared to people. So that's kind of things that models can help us. And at the same time that you are producing those models that generating great voices, we need to generate models to make sure that we can identify those voices generated by the machine.

That's interesting that you have to then have the I guess forensics models to determine what generated the thing. So you talk about like its patterns, speech and I guess is there a breathing noise in the background? Okay, can you do fingerprinting on this? I know with like images you can add like a fingerprint. Say this was AI generated and it's like just slight changes to the pixels where humans can't detect it, but a machine can. So do you do the same with voice? Can you put some.

I'm trying to. I mean at the end of the day there is a very clear way of how we are generating those voices so we can identify when and where we generated that. For you as a person, it could be exactly the same, but as a model that the only purpose of this model is to identify this was generated by a person, a real person or a machine. Those models probably do better than what we do as people. So when you get a call and if you are not sure if this is a person or not, never ever give your credit card.

That's a, that's a very good idea. You know, I used to work in debt collection on the phone just asking people for money and almost everyone would just give out their credit card details, whatever. So yeah, it's, it's tricky stuff. So I'm wondering whether there are limits, current limits to how well the speech AI can work. So we talked about becoming very natural sounding. Are there any areas of speech that are difficult for AI to replicate at the moment?

We are dealing with many challenges, specifically the three of those that I talked about. I think background noise is tricky. Specifically where there is not. There is no clear signature or pattern of, of the background noise. And that's kind of thing that at the beginning we're having problem. But if you can build a platform that collects information and learn same way that we do as people. Right. We try to learn from mistakes and improve ourselves. You're probably going to get into a point where the models that you are generating are I would say, better than what people can understand.

That's kind of cool that most of the Problems. I mean, it sounds like there are challenges still left to be faced, but most of them sound like they are at least solvable in the near future for us.

I think we are right exactly in the right time. The potential of voice is not there yet. At the end of the day, I believe that every machine interface is going to be voice and we are going there, but we are not there yet. And I think that as long as we can integrate the jargon, as long as we can integrate accent, as long as we can get the background noise, we are in the right path to achieve it.

That's interesting to say. Basically every machine will have some sort of speech or voice interface. I'm actually slightly concerned about what happens in open plan offices. So just for context, Daisy Gap has an open plan office. Joe from Sales. It's naming is very loud. If he starts talking to his AI all day, I'm going to murder him. So what do you do in that kind of context? Do you think it's going to take over like offices as well?

I don't think so. I think that we as people that are working in offices, we are going to learn how to take advantage of it and to improve ourselves. But I don't think it's going to take over what we do. We're just going, you know, it's not about writing 20, 30 years from now. It's not about writing the model. It's more about how you are operating your models to make sure that it does the things that you want it to do.

So basically just don't be dumb and put speech everywhere. Just make sure that it's in a sensible place where it's actually going to benefit you. Cases when you don't want to type stuff for communicating instead. All right, I guess just to wrap this up, do you have any advice for businesses that want to make more use of speech AI then Liz?

I believe that first of all, in any change you are making in your organization, you need a champion that can run a change. What does it mean? It means that he needs to understand the use cases, for example, that he wish to prioritize where he will be able to he or she will be able to show value. To gain confidence. This champion will need to pilot before he runs. He needs to walk. The champion will need to make sure that there is budget because hey, you are building a new operation. The champion needs to decide if he is going to replace existing systems or he wish to integrate to existing systems with voice. A lot of decisions that need to be done and they Cannot be like random decision. It's a decision that the organization is taking. You need to involve the relevant employees who will be impacted from that. And you need to show them why we are doing the change. You need to choose the right processes to allocate budget, to analyze the ROI, the KPIs. So it's a program. In order to make it successful, you need the right planning and the right champion to run it.

Okay, wonderful. So this sounds like it's going to be very similar to like a lot of digital transformation programs. You just really need to be mindful about like what are your goals, how are you going to measure performance, who needs to be involved and all that kind of stuff. That's cool. All right. And is there anything that you are most excited about in the world of Speed Reach AI?

I'm excited and all the. If you look in the company, you see everyone are excited because we are making something big. It's the right timing. We see the usage growing dramatically year over year. We see agent coming and joining this big wave. And agent will use voice a lot. We are going to change the way employees are using machines reporting data. We are going to be part of a change of how unstructured data in those industries, across, you know, industries that are. You speak about AI, they are so, so far away. We are going to work with those industries, collect unstructured data and give real time insights that they never had before. We're going to bring voice AI to every employee. It can be employee in the supermarket, employee in the airline, employee in the automotive. Every employee will start using voice and conversational AI. And we are in the middle of this, maybe in the late beginning, but we are feeling the energy in the market. So every employee here in our company is feeling that we are making a big wow and part of a big story. And yeah, we are excited.

Yeah, it just seemed like exciting times. And Gil, what are you most excited about?

For me, this is the way of living. I think at the end of the day, as I said, every interaction with machine is going to be by voice first. And we are there revolution we are going to have in the AI revolution.

Yeah, definitely exciting times. And it seems like there's a big change in the way that humans and machines interact in progress. Wonderful. Thank. Thank you for your time, Alan. Thank you for your time, Gilbert.