51 - Wordly.ai Thank you very much to you, Lakshman, for coming onto this podcast. It's quite brave of you to, like, sit here facing for interpreters. Does it feel a little bit like going into the lion's den? Well, yes and no. I've not worked with interpreters as much. The reason I came in here is to try and understand the interpreter's perspective rather than a technology perspective or a buyer perspective or a end-user perspective. We have always been working in parallel with InterpretBank companies, if they've tried to see what we are doing, but always quite cautious about whether we're going to eat up their market. When we have that kind of fair a situation where people are viewing us with certain doubt, then we are always very careful to tread. So this is this is really good. Welcome to the Troublesome Terps, the podcast about the things that keep interpreters up at night. Now, my name is Alexander Gansmeier. And with us today is the guy who literally wrote the book on machine interpreting versus interpreters, Dr. Jonathan Downie. How are you? Thank you, Alex, I'm fine. I am 100 percent non robotic today, so I'm really pleased. And it is my pleasure to welcome the man without whom there would be no Troublesome Terps. He is like the drummer in a band or maybe the bass guitarist. Oh, careful there. He's like the sound engineer at your favorite concert. This is our favorite tech guy ever. Alex Drechsel. Well, thank you, Jonathan, out of L.A.. I'll take that. But careful with the base's jokes and stuff like that, they're a bit you know, they have a special reputation in musicians circles. But it's great to be here. It's lovely to see you all. Misunderstood. Yeah, totally misunderstood. I know one person who is not misunderstood, and that's Sarah Hickey, who is joining us once again from Ireland. Good evening. How are you? Very nice introduction. Yeah. And our our epistle today is actually going to be particularly troublesome because we're going to address the big threat of machine interpreting. Will the machines replace us when we learn to coexist? And what does this potential future look like? J.D. is our in-house expert on the topic. And to really get into the weeds of it, we invited a very special guest, the CEO of Wordly A Machine Interpreting Solution. Welcome. Lakshman Ratnam. Thank you, Sarah. And thank you all for inviting me to participate in this session. I'm excited. I'm looking forward to discussing machine interpretation today. Fantastic. Great to have you. Maybe we can kick this off with you telling us a little bit more about yourself first so we can get to know you better. Yeah. Well, I will keep it really short. I'm a technologist. I've worked in the Silicon Valley for over 25 years and I worked at many companies and managed teams. I had I had I've managed teams globally as well. And while managing teams globally, I always had this problem of communicating with with the global teams as well. But my Segway into machine interpretation came a little bit roundabout. I was working for a hearing aid company and I was helping with the hearing impaired. And then I realized that hearing aids catered only to a minority of the people that needed hearing assistance. And I. And I wanted to use machine interpretation or machine learning and artificial intelligence to provide them with perhaps pitch text so that it could increase cognitive load for people who could not afford hearing aids or were so hard of hearing aids that hearing aids would not help that. And then I connected this to my pass and said, if you spoke a foreign language to me, I can hear you, but I can't understand you. So essentially, that's what led me to looking into interpretation as a bigger market. That's a really nice way of thinking about it. And perhaps helping bring the world much closer. And it is such a different vibe these days, doesn't it? Yeah. So that's what got me into this whole world of interpretation and getting people communicating across language barriers. I had no background in language or interpretation or in that industry and more technology person. And yeah. And technology person who saw a problem, he wanted to fix That's right. The. As technology to flow into the the really fascinating thing about wordly seems to be that you're pushing into markets that love current machine and tariff thing solutions aren't really looking for. Could you tell us a bit more about who's using wordly and what they're using it for? I'd say sorry to interrupt there, but maybe we can start with our listeners as well. What wordly is first living? Yes. Yeah, so Woodley's. It is as simply put, it allows any speaker to speak in their language of choice. And allows any listener to listen in their language of choice if they have the act, if they have access to work. So what? So we use artificial intelligence and machine learning and help do interpretation, too, without any human interference. I think that's the best way of putting it for general consumption. Ok, thanks. OK. Jonathan, back to your question, maybe. Yes, so Sofus using wordly. What are they using it for? Ok, so what do we have? As you can imagine, there are so many use cases for interpretation and there just but ours is a new paradigm. Just having machine interpretation replace a very human to service interpretation predominantly. So getting into a market is really quite difficult because it's a new paradigm. It needs people to accept it. However, having said that, we have overcome this a little bit and there are early segments have been into conferences or big events where just scaling up for interpretation, human interpretation has been very expensive. And the logistical nightmare of having to provide some boots, custom headsets, infrared transmitters. And at the end of the day, they're more concerned about retrieving their hardware rather than the experience that people get. With interpretation. And so we have found that to be a good market to validate our product. And what we're now seeing is that there is a general acceptance in the enterprise sector. And given the current situation, there are more people willing to accept our solution or at least beginning to engage with us because everything is virtual and scaling for virtual platform to human InterpretBank seems to be an issue. Just in general, with the availability of human interpreters. And I think markets will dictate that. Yeah. Yeah, the market dictates everything these days. Exactly. Could you explain a bit on how it works? Because I know for pretty soon and thereafter as we look at machine interpreting and it just seems like a magic box. Could you kind of break down the basic steps that it goes through to determine a spoken language and one and one language into another? Yeah, so let me just. Perhaps the easiest way for me to do it, to explain this, is to just talk you through the the flow of how the data goes through. And being a technologist, it makes sense to me. But I'm happy to answer questions after that. Right. So what we do is we capture the audio of a speaker. And the important thing is to capture this in a clean fashion. So we capture it through a mobile device. To date, it's real supported. We we run on Android and US devices. So we capture the audio and then we send the audio into the cloud. There we are, proprietary cloud, where we then do the speech recognition, S.R. and then we do the translation as well in that cloud. And what people were what attendees can do is dig at a very secure code for each session that Woodly is being set up for. They can join these each of these sessions and they received the translated text in which would language they prefer. And then we can read that into all of you for them to listen as well. So it's it's it's an all of this is encrypted. So does that help? Yes. So by a Saagar meaning automated speech recognition, which is a technology that takes spoken text and turns it into written text, is that is that correct? That is correct. Basically, Ok. What it does is it takes words and it it it breaks it into phonemes. And then, by the way, it does is that it it recognizes phonemes. Senate understand it. It comes up with the best probable highest probability of what the word should mean in that language. It's extremely fascinating, I've looked into that a little bit, but the whole sort of way this works and how it splits up the the way forward basically into different chunks and tries to put them together is super fascinating. Yeah. Is as fascinating and it's it's fascinating to compare that with the little bits that we know about and therefore cognition as well. But I'm not gonna go into that yet. We'll come to that later. Maybe just a little bit. Well, OK. What did one of the things that has fascinated me, me about machine interpreting, I think, is that the underlying model that you've described has a label. So an interesting research. We call it the conduit model. And it's interesting because the most interesting research has gone on the further away from the content, more the researchers have gone, especially it started off in interpreting and say medical settings or in courts and things even on actually, as time has gone on, we've realized that the conduit model explains some things really well, but doesn't explain things like why interpreters choose this word over that one or why they choose this phrasing over that one, or why the how managed to self correct. And so it's fascinating to see machine interpreting, using a model of interpreting which has really been questioned and challenge this 1990s and research and to see what the the results out of using the conduit model to do machine interpreting vs. the model that most and pterosaurs are trained to know, which is called the triad model, which is more about cooperation and understanding what's going on in the situation, that rather than concentrating on what that person just see. And it's fascinating to see the difference in how the tech world understands interpreting vs. side. The interesting research world understands and terrifying. This might be very out there. But is it fair to say that the Kondrat model is a bit like the Cascade model, like the stand up or the old model of HSR, if you will? Yes. Well, it's Just Basically In the way The That Athie. It's chronological, I guess, sort of sort of Print Out. Print came off. It's more like the best analog to the content model is the ESR to machine translation to split the audio out to the end. That is classic. That is the kind of model in a nutshell. So it was the entire four. Here's the process they see. And it was always the only thing the content model ever dealt with was the language. And so there have been lots of studies since the 1970s on the entire first chunk. The interface was turned phonemes into words and the interpreters turn words into sentences. How do they know when a sentence is finished and so on and so forth. And I believe in German that's very difficult because sentences in German never end. No idea what you're talking about. But That The. This is is one of the things, and especially actually German has been a really useful way of studying that, because with it being very pinel so often, then negation, fatal Negation final, yes. Negation final as well. So this is where the conclusion that researchers have come to since nineties is that interpreters are using their social awareness on their world knowledge as much as they are the language that they're heeding. And it's interesting because when you gave us a demo of wordly before this, my brain was going, oh, I can see the content more working. You can actually trace issues of that word because that's the right content more than words. But an interpreter would never choose that word because we would have context. And it's just fascinating to see the different MO that was working. I wonder, I led to make come to this conclusion at the end by wonder if the two models mean that the two solutions are good for different things. Yeah, so I I tend to agree that there are areas where machine interpretation exists and there are areas where human interpreters will exist. And when I look at the big picture, this is how I see it in the long term, perhaps short term is very different. But if you look at any piece of colonization there, you can break them down into two oh, two areas. One is the emotional content of that speech. And the second is the information, context of that speech. And we believe that today or for the long term, the information content machine learning is ready for this. However, the emotional content cannot be captured as easily because just the way that people behave, it's so. Depending on region, depending on personality. What kind of emotion they bring into their speech. Even their tonality, the way they speak is very different from person to person. And it's very hard for a ride to capture that in a in a in a in a large platform. So I think that's rare. That's a limitation. Today. Does your system capture intonation data as well? Because I know that was a limitation of a lot of systems, is that they would capture words, level data, but they wouldn't capture things like emphasis or intonation switches that humans should enter and go on. He's being ironic. He's being she's being sarcastic or she's trying to be funny. Yeah, I think I think there's some way to do it. We're not there yet. That that is not too far off in my sense. But sarcasm is going to be really hard because sarcasm, a game that different from person to person. I think this is this is this would be the biggest challenge. You can do it on a personal basis, but doing it on a broad scale is going to be very challenging. You make it. The hard part is you make it at 60 percent. Right, 70 percent. Right. But for for general acceptance of the market, you have to have a ninety ninety five percent rate and a margin for error. There is much lower than what people think. I just have a quick follow up question. Following up on what Jonathan just said on the air, you can see the machine think or change things according to how the sentence develops. As far as I can tell, a lot of machine translation systems or speech transaction systems actually show you the transcription results in the original language. And wordly, if I remember correctly, doesn't do that. Was there any. Why did you decide not to do that? I suppose the others do it for transparency or. I don't know. Wait, why did you decide to go a different route? I think. The are our way of doing this thing was to make sure that from a human machine, human behavior model, people don't have time to look at two transcripts and follow the conversation at the same time. And and when you look at our product, it's meant for continuous conversation. And the way we have designed our product is to make sure that you can have a continuous conversation without having an end, like if you had an interpreter in between who is doing sequential interpretation. Right. Then I have to speak a sentence. Wait. And then let the other person speak. What we're doing is make sure that there is a seamless interaction between people. Now, when you do that, there's no point in showing two languages to the other person. However, what we do is when I'm speaking, let's say I'm I'm using wordly and I'm speaking. I can see what I am speaking as a transcription. And what that allows me to do is that when I see something that is transcribed that is incorrect, then I can I can almost predict for 100 percent that the translation is going to be incorrect. Trouble ahead? And so I said so. So I can I can I can repeat what I said. So the translation comes up correctly. Plus, what you're saying is, well, with not showing both the original and the translation before, I'd say the person who receives it. I would say yes for us analysts. Maybe that's interesting if I can see English and German or for interpreters. But let's say if it's if I'm attending a conference center, then I don't know Swahili. It's not of no use to me and to see the script. You know, I mean, it might be interesting to see what kind of words come come up in that language, but it's not like I can do a comparison anyway. And then I just want to understand what's being said. Right. So. It's one of the bigger things within tariff thing that we don't think about in office, the fact that if interpreters are being called or machine and tariff thing is being used, it's because people don't understand one of the languages Exactly. And certain. But this is a fundamental thing, and it is one that we forget so easily, is that there's different two things. One, there's a relation of trust, of implicit trust that we should never take for granted. And two, without some kind of solution that doesn't really a lot this physical communication, which kind of happens. But, you know, the real communication isn't going to take place unless there's some solution there. And I think once we understand that, we can understand things like quality judgments, we can understand why some clients, even for a big conference, would say we just want machine interpreting rather than seeing, you know, how horrible you are. It's like, well, put yourself in their shoes and look at the problem that they've got. And actually, that solution might make sense for some conferences. For others that definitely wouldn't. There are certain meetings. Would I imagine that they would. In fact, I've been in meetings where they call human and teraflops because it's a prestige thing. It's horses for courses. No, I totally agree, I think that that human behavior, again, going back to human behavior. Procedures. You know, there are people that like to on. They like to go for the high level meetings. They like to go with their interpreters. And that also allows them to break the conversation and discuss internally and give some thought process. We don't we believe that that level of dynamics will exist and continue to exist in the future, because that is that's a mode of business operations. But the you know, the the in general, the number of I think there are more meetings that are happening today that would need informational current interpretation. And I think that's where machine interpretation is going to perhaps take a bigger. It may not eat up the revenue, but it may it definitely may have some penetration in those areas. That's exactly the question, right? What? What's the overlap and how's it going to play out? Yeah. That's the interesting bit that we don't know yet. It sounds a lot like so when I started, I first translated before I interrupted and I got a few documents that I translated and they were labeled for information only. I don't know if anyone else has come across stuff like that. Context on the 20. Yeah, seriously, it is for information only of the other one which we dealt with in class and I had one and it was annoying was translation for the file, but that kind of stuff. Translation for the file and tariff thing, that's just you need to do this so long as it's not information that's going to get anyone killed if it goes wrong. It makes sense to have a machine. It's life or death, profit loss where there is some danger in error, then I think people will probably go for humans. But this is, again, a whole conversation to have is and we didn't even have this in the next election. Traditionally, conference Interpreters' Help wouldn't have dozen General have been quite scared and seen machine interpreting as a rival on some of the PR from the speech. Translation community hasn't been encouraging in that regard. There's been a lot of talk of replacing interpreters of, you know, in five years that we know more humans left. Is it anything you would want to see through that controversy? What would you want to come tear therapy today? And would you want to get as anxious to make us all interpret better? What would you input to that controversy be? Yeah. So, well, I think I mean, I can I. Let me put it this way. I empathize with their concern. But I believe that the market is really big for both to coexist. And here's what I my research has said has shown. Interpreters are very hard to come by, just by nature of the business. You know, if I if I am in Middle America and I need a good interpreter in the language, I need to fly them in from either the East Coast or the West Coast or, you know, they're just hard to find. I mean, that's nature of the business. The industry by itself, interpretation industry is fragmented. And if you don't have a platform to unfragmented this, it is going to be very difficult. And what that will lead to the end of the day is that more and more people will not use interpreters, which we actually see. Only 15 percent of the meetings that actually need translation are using interpreters. There's a long tail where they don't use it. They just get by with in a body language and or repeat meetings or whatever it is. It just affects productivity. And that's that's the last part of my. That's the biggest worry. Right. Is that the Long-tailed languages, if you want to preserve them in the long run, you need to do is the dominant language. This will just take up and then people's identity will be lost in these participation. And that is a bigger picture that people have to look at than just look at it as interpretation, as you know. Well, I, I, well, I think it's not I don't want to trivialize it, but not just look at interpretation as a source of income. But how do we look at the bigger picture and preserve languages? I think that is the bigger picture. One has to look. And I would add to that that in addition to preserving languages. In the end, all about providing language access. And like you said already, that in some scenarios or in some locations, this can be extremely challenging or next to impossible. I also think that there's definitely scenarios where people right now, they won't get interpreters either because they can't or they can't afford it. So something like a machine interpreting it can actually, I think also like open up the market even further. It's already a big market with lots that's left unfilled and it might open up even more fields for interpreting and for language access because it's possible this way. That's what I was thinking, too. I'm thinking a lot of those meetings that can be then interpreted with machine interpreting. If more people see it more of the time, it could also help raise awareness. Just generally for the need. And for the the demand that is out there in the meetings, which would then in turn also potentially benefit human interpreter. This is a case that I've been making and more so more recently. There was video that I brought about four months ago called Neumeister Robot. You can have my job, but you can help me do better. And it did feel that way. I've got a couple hundred views. And the interesting thing about that is people are latched on to the you can't have my job. And no one latched on to that. But you can help me do mine better. But one of the use cases for machine interpreting, which I think has been under way, under explored and undervalued, is to use the same technology that underlies machine interpreting and use it to make human integrity more accurate by doing things like speeding up terminology. You so when an interpreter is working simultaneously. You know, we work in teams of two or three, if you're fancy and there's often an interpreter doing the terminology searches for the one who's active. Because when you act, if you don't have the time off of me with the term and there's no reason why this technology couldn't be repurposed to see. Well, the term I recognize isn't in the top ten thousand words in English. I'll check your terminology database. And if not, I'll check the ones that I know you. And that's every I.T. person I've talked to said. Yet the technology exists to do that. And my response has been, well, we need it, because that's the kind of thing where you can get the two sides working together and benefiting each other. And I'm sure from the data of, you know, the terms that interpreters look up most often or have interpreters use machine interpreting that can improve machine interpreting, too, especially on the language recognition side and their scope. Yes, there's scope for rivalry. We should never deny that. But there's also scope for corporation. Well, like I said, I agree with you, I think they both need to coexist. And you're one of the. Yeah. I mean, what you said is quite accurate is that we could feed off of each other. Right. Human interpretation and mission interpretation. The more. Human interpreters use machine interpretation to improve their their quality or their delivery of interpretation. It only speeds up their interpretation and it reduces the cognitive load on them. And what that can help on the A.I. side of this equation is their inputs can actually make the probe. The algorithms much better and provide better outputs or better translation quality that can improve machine translation. So I think they can help each other out. And at some point, I just don't see the scaling for human interpreters being as many just given the nature. I just think that this human interpreters don't scale to the global requirements that their interpretation is needed to be. But that's also kind of related to specific languages. Or would you would you say that's something that's a trend across the board, or is that something that you see specific to certain languages, the language groups, maybe the scalability and availability? Of course, along the availability of interpreters for Long-tailed, languages are fewer and fewer. And that's why even now to their conferences, they probably do. Translation for one or two languages, not just the expense, but they just are not able to get the hours for the long dead languages. I think that over time, the Long-tailed languages. Destruction of data to for the local languages to have presence brought him here with interpreters as well as vaccinia into patient today. But I think it'll catch up soon. Yeah, and I think just following up on that, that language is that you have available. I suppose also constrained by the availability of training data and material to defeat the algorithm, as it were. Absolutely, I mean, that's that is by nature of the quality of machine learning with it. We just need more data sets to be able to do it. The second thing that we are also focused on now, at least being an early stage company, is we focus more on business case. It's easier for us. We can we can get the data for long, three languages, but it's all a return on investment and where we put effort and how we prioritize. I mean, what differentiates what you're doing from the solutions offered by, see, Google or recently Wab's or the other people in the same space? Yeah. So one of the key things that we do is. And as you saw in the demo earlier, we do what we call the continuous translation. Number one, I can speak. I can continue speaking. And, you know, it translates into real time. What we do also is that we translate from one language to multiple languages. We don't require special hardware. You don't need a special headset or a special earpiece to listen to the translation. Your mobile device is good enough. Your computer is good enough. We make it super easy. You don't have to be in the same physical location. You can be virtual. And I think, you know, and our quality. At the end of the day, we have to say is quite reasonable or the delivery depending on the language. Language barriers that we use is fairly good. No, I just had a very practical question, because on your Web site, it says that you're offering it for, among other things, global meetings, conferences and briefing centers. And then in the demo that we saw, it was obviously just you talking. So it was basically all just continuous chat bubbles, if you will, with your text. So if you were to be in a meeting and let's say you and I have a conversation with The View, then adapt or because you're doing this continuous and seamless stream of speech, if you will, like, does it all show up just, you know, as one kind of blob of speech bubbles, or do you do like literally like a chat view where like I say something and then you say something and it shows. So we. So what I showed you earlier today was one speaker and many listener, so it's a one too many mode. And we do support the many too many as well. So if you and I are speaking, if you had the application on your side and I had the application. So what you would see on the screen, is it a Choa Laxman if you put a small initial bubble and say this is what iceberg? And then would say, Alex, for this. And so, you know, we'd have a different breakdown by the speaker who spoke and with with a small initial next to that. But if I had to see it, I will see only what is being said in my language that I choose. Whereas if you had to. If let's say you're listening in German, then you would see everything in German and be able to hear everything in German. Does it label changes of speaker, if just continuously during the same language? Does it say, you know, the speaker said the speaker said that? That is correct. We allow we do a speaking breakdown of that. So did they then, so if you've got a standard conference set up where, you know, you could have three speakers who are all speak using the same microphone, so you have three speakers by the back who speak three different languages. Does it also detect language or does it have to manually change? OK. There's a sentence where he's speaking and. Yeah, yeah. So today for Fat Week, we have the technology to do auto detect. But today what we do is we do it manually. That's what we have in the in the market. A product today that we have to switch it manually. And then would you manually would someone manually label, you know, this is a new speaker, but they're using the same, you know. This is Bob. This is say it. This is Ferrall. Yeah. So what we do is that on the typically in the conference, what we do is there are multiple outputs to the to the microphone channels. So we speak. We use different channels for different speakers. And so Ok. It's automatically label. Yeah. All right, OK. Because one of the things that you often have at conferences is you get, you know, speaker after speaker on the same podium and interpreters have have tricks of the trade that we used to make sure that someone's. Well, I mean, the nightmare for us. I wouldn't say nightmares is actually quite funny is when you get an open forum or a Q&A session or you get a panel session. So I have I had one at a couple of years ago, I was interpreting from the Scottish Parliament and it was the consul general of France, the French consul general to Scotland was interviewing a very famous French speaking musician. And they are doing the interview entirely in French and I'm interpreting entirely in English. And I had to very quickly come up with a solution so that people knew who said what and when. And that's the thing as a profession, I would expect and I'm a consultant interpreter. So if I've got and tariff doesn't my team, I would know that they would know how to cope because I would expect I would you know, I would only get people who would know how to cope. But I'd imagine with machine interrupting, you would maybe have to pre label the makes or you would have to have a very good person sitting, going, this person, that person, the stress and that person. Right, but that's why I ask that question as well. And there's Luxton, who is saying, if you and I have the conversation or if multiple microphones are being used on stage, you can simply pre label that or the device tells you anyways. And that's also the quent. That's why I ask that question as well. Because, you know, if you're having a conversation or if you're having an interview or a panel discussion, you need to as a human interpreter, you need to use those tricks of the trade to make sure that people can differentiate. Whereas with machine interpretation, at least if you look the transcript, you can simply label them. And it's it's, you know, dead easy. Yeah, and which is what we do, right? I mean that I mean, the the the hard part is labeling the Q&A part and that that does that is hard. But we we typically call that as a guest. Right. And which is what perhaps human interpreters do, as they say. What does the audience. That's. Question from the audience and just don't label the person. We have to keep some checks of relief. That's right. But just a quick follow up question on that, because I'm guessing not everybody always follows the transcript. Some people, when they're at a conference, they also just want to listen to the speech synthesis. So if I don't know you and Jonathan and Sarah, you guys are the three of you are in a panel discussion, and I'm not following the transcript. Is there a change of voice? Because I saw in the demo you could also choose between three different voices. Yeah, so. So we have we have choice, four different languages, some different voices, so you can pick different voices. You can you can add different languages, come out of different voices. So you can definitely tell if you're listening to people. Then if let's say I'm listening in Hindi and it it'll come out, of course it will come out in one voice. But you can. But it'll be labeled as different voices like I. I see your point. If we have to do different. That's a that's a good idea for us to change. I like it. Sorry. I like that. I like that. That's a good idea for us to include. That's awesome. But Yeah. I mean, this is this is the thing as well, so where I would quite happily turn to machine interpreting is I wrote once about the kind of people that you meet at a conference and all the interpreters. And therefore, we're going to know when I say this, when you get when the chain of the session says, does anyone have any questions? And there's a Rambler in the audience with a tape, written script that doesn't have any questions on it that no one actually needs to understand, because It's It's More of a comment. Always a mar. It's always, always a man. I don't know why. It's always a man. They always have. They always have size. A tight script. And just. Do you think they're finished? They hold the papers up from what they do. They do that. They tap the papers on the table. Then you turn around and start from the back. And in those situations, when you're just speaking for the sake of speaking it, be nice to have a button on the console that goes missing. Can take that one. Well, I I think the hard part are the similarities between machine translation and human translation. Interpreters is very similar fate in this part, please. So if the speaker is unintelligible. A human interpreter finds it hard to translate. And the same is true for machines as well. If they can't understand, if the machine cannot understand what the speakers say, it is just going to be that much harder for us to recognize as well. And this in this. Garbage Yes, In. Garbage out. It Yeah. Is Yeah. Hard is the weakness that does. That does remind me I launched the challenge, which no machine and terrifying company has ever taken me up on. And that is to pit machine interpreting lie against remote human and toes and human interrogators in the room and to do a blind a double blind test where the audience would be single blind, really. The audience don't know which speech, but they're free to just switch between them on specially rigged up where they could use one mobile phone right now to text the other. They could just do on a mobile phone. So every channel is unlabeled and the humans just scroll through and they just get asked afterwards what they thought. But of course, if you're going to do that, you have to make it is like a real conference as possible. And I am a crazy Glaswegian. So, I mean, let's get some crazy The Glaswegian there. Second invite your fishermen from Aberdeen, Inverness, nevertheless. So so what, Laxman doesn't know the story. I've done a few deep sea fisheries policy meetings and the SEC first or second, when I did the chair, who was a several very senior civil servant in the Scottish government, came up to me and said, Are you two Scots? And because of a two Vesalius were proper professionals. And I was like, Yeah, know we're both Scots. It great. We've got Aberdonian fishermen here and I don't understand a word they're saying. Can I listen to the French and said. And the four score something that was two Scots and the boos that they we were completely fine with it. And and my brother may even manage to do a nice little interface interpreting from Garlock because the guy decided to open the meeting and garlic. Thanks. We just went with it and it was fine. But one fisherman at one point went back in his chair away from the microphone and said, But anyway, I come across, I need the waist up. And we were we interpreted the problem. But you could see all the all the English speaking. David, it's going almost. But just at the risk of giving you more free ideas. I'm wondering I'm wondering if you're also looking in other sort of related services, I guess, because there's so much attention on machine translation oftentimes that this the other services or things that an LP can do gets sort of overlooked like, you know, just transcription, just monolingual transcription and no emotion extraction or extracting to do Eitam, stuff like that. Is that something you're looking at or is that still too far away? Or maybe not interesting for you? No. Oh, definitely. I mean. All of this at the end of the day, getting that adding additional information into interpretation just adds value, right? I mean, that it is it is something that we have looked at. So the first part of it is how do you extract or do a summary and how do you extract keywords? And then with those keywords, how do you actually provide, you know, some sentiment? And people have talked a lot about sentiment analysis. But the key to getting the right sentiment is to have context behind the conversation and then use that context to then derive the sentiment. And since we would have the data in different languages, it's easy for us to do something like this. So these are natural patterns for us as a future. To be able to, like, provide to companies. So, yeah, definitely in that roadmap. To be fair, I know some humans aren't very good at sentiment analysis. So, you know, you are a step ahead of someone. That's where they'll get us. I'll tell you. I was going to say I. I can see which job they go in to buy would get in trouble for mentioning that job. It doesn't start with I. Does it? No, no, no, no, no, no, I was thinking of a job beginning with T. This. To go into something a little different again there or actually pick up something from a little earlier when you were saying that with A.I. and technology, you could only really be successful once you have once the qualities that something like 99 percent range. So it's just I've I've spent a bit of time in Nimdzi as well, of course, as, you know, looking into machine interpreting. And we spend as a company a lot of time on A.I., you know, on all the technologies in our industry that come with that. And a question popped up there for me as well as to say, what do we want from A.I. in the end and what are our expectations that we have towards a AI and towards humans? Because we always make the comparison, of course, AI and humans. And whereas for humans we say, well, it's only human to make mistakes. Right. But for AI, you don't have that. People don't make that excuse. It's like it has to be perfect or it doesn't work. It's like there's very little leeway. I think that we're giving. There's some bias there. Yeah. There's this there is some bias there. So definitely there is some bias. So let me. There are a few ways that there are few errors that machines do. I mean, let's put it this way. Neither of the interpretation schemes are 100 percent accurate. Humans have not hundred percent accurate machines. Not that machines are 100 percent accurate. So today what machines do and are not so much in the interpreter space. I want you guys have more experts on these and I can ask some questions on that. But the the machines, the the biggest error is all driven by the human speaker. So let me put it this way. If I am speaking generally, people speak and phrases. People don't speak in complete sentences. Now, when they when they speak in complete sentences, the recognition that the machine translation works really well. Now, when they speak and phrases, it works well as well. Now, if they speak grammatically incorrect sentences, then machine translation has a big problem machine translate. And then if I speak something and then I change my sentence made midsentence, I, I correct myself. Human interpreters do really well in that area. We just don't do so well. The biggest thing for a human for machine interpretation is proper and proper names. Right. So people are people's names are very hard to come by. I mean and people have names like summer and winter that have meanings and so on. So, you know, that gift literally translated. And over time, I think that's where we are today. I think some of these can change over time and we can add glossaries and we can bias the output to for it to be a noun, which is, you know, something else and not translated. So there are things that we can do some magic on the back end. But those are areas that are the similarities and the differences between human education. But the question I have for all of you is how do you measure quality of human interpreters today? You know, some are good. Some are not good. And then how can we then equate that to regime interpretation, not with just, say, machine interpretation is not acceptable. Right. Please stop stumbling into stuff and doing research on what e.. Legal question. We'll cut But. You off. Jonathan, Ok. It's I Fine. Want to say good first, because I am going to run. I was I Just Should Gonna Wait Say My That, Turn Yes. To run. Actually, I was just recently asked by Renato, our CEO at Nimdzi. Also, like, how do you you know, how do people get tested for interpreting or how do you check interprete quality? And I think it also something I forgot the context of with how you can do it fairly quickly or something, or who checks interpreters on the on the market. How do you check their quality output as like you you kind of don't you know, I was saying that institutions as the head of Buth who checks in maybe with their interpreters every now and again and all that listens and and then you just have your qualification at the beginning where lots of people check. No one just listens to the output. And one day, let's listen to both and all this kind of stuff. But even there is still, of course, very subjective. Even though I might be very high level is still also subjective. And that's why I think in general, this is something we talk about a lot at Nimdzi when it comes to quality of both interpreting or translation, there is no definite quality standard. And yeah, and it That's Doesn't True. Exist, you know, and everyone keeps saying, don't keep saying my quality doesn't matter because everyone says they have the best quality. It's not that it actually doesn't matter, but it's not a differentiator because if you ask in a room, how do you know who you provide the best quality. Everyone's right, Hatton's. Yeah. So then you face that if we say we're the best interpreter. How do you then take machine interpretation and say machine interpretation is just I mean, I can claim today that we have bested patient quality. Does that you know, how do we measure it? Right. I will give you a quick and I promised Alex that I will be quick. Otherwise, he will weigh the post. OK. Two things. The standard quality measurements for machine translation so I can remember what the meteor and the blue score, they are highly pro unde. We as a machine translation expert will explain very clearly why they don't. They're highly problematic. My there has been lots of research from quality in interpreting the data. Basically of the best we can get to at the moment is quality isn't is almost entirely to do with context. So instead of talking about quality, I talk about in both of my books, I talk about adding value and I ask the question and this is a good question to ask both for machine interrupting. For human interrupting. What difference does interpreting make that wouldn't happen. Different sampling wasn't there. And then you are then the next question is what risks were there that were either made worse or were mitigated by the entire thing? That's a new one. I've been working on. So, for example, quality and medical interpreting a base level as the patient didn't die unnecessarily and they got the right treatment. That's what quality comes down to, quality in quality. If you've got a really artistic speaker comes down to did the audience experience it as well as the source language audience. So it's a really dangerous area to get into by would cancel anyone doing machine interrupting or automatic speech translation, as I'm calling it, because of content model. I would counsel anyone in that space to either talk to guys like Andy, we who are creating new machine translation quality measurements or to very be very careful about how they use the Q word and stay away from quality and talk about use cases. I would rather we talk to use cases than quality. But I have devised a quality test that we can try out, which is that pitted against each other. No. No, well said. I mean, that's that's exactly what we what we are. Our platform is as well as, you know, at the end of the day. Do we add value? And if we do not have interpretation, do people walk away from having understood nothing? Or with the interpretation? Did they walk away having understood something? And is there a net gain? And I think actually, I think even in medical use cases, machine interpretation is actually prob perhaps relevant because most doctor patient interactions. Doctors typically ask the same question five to ten different ways to make sure that they are actually coming to the right conclusion. If they say you if you go and say I have a headache, they ask, is it your left side? Oh, you know, when I touch here. Does it pay? You know, so they ask the question 10 different ways to, like, make sure that that the answer is correct. Are their conclusions are what they think it is. And the same can happen with machine interpretation as they are already trained to do something like that. Of course, we don't want to get into a case where they're going into the surgery room or asking, is it right leg that I need to take off? You know, right I left I. When they're doing surgery. But in the in. There are areas of the medical profession where mission interpretation will work. And there are perhaps areas where we do need human interpreters. So I have a follow up question on what you just said, because you said the doctor is usually have five to 10 different ways of asking a question. And I think this hearkens back to something. I forgot who said it earlier, but I think it was about the. I think it was actually you, Laxman, who said it. But the names, you know, summer and winter. And I know, for example, for a lot of the dictation software out there, like Draga and or whatever have you, you can prime it. So you can actually say, OK, we're gonna be doing a dictation on automotive engines. Here are the main words that you need to know. And then you type it into the program and you basically prime the software. Is that something that you guys can do as well or that you're doing already? So if you're having a a meeting on. Who knows. Whatever. On my new iPod. Like, can you prime it on my on my iPod terminology. So, yeah. So the short answer is yes. The long answer is that one of the things we have done is we built our models are built on very big knowledge base, so broad base. So what we then typically ask is that if you have domain specific acronyms, especially many companies now have acronyms Yeah, For different things. We know. Ah, yes. And acronyms are very hard to recognize. So we we ask for acronyms. We ask for names that are unique, such as wordly, you know, corporate names that are Atlassian or, you know, that are not in the dictionary. Product names that come up. And we ask for all of that and we actually create a special dictionary for that particular session. And then we can build it for that particular domain of company. We can continue to build on that for them. So that's how we do it for domain specific or company specific requirement. That makes a ton of sense. We do the same thing. I totally was. I totally want to see the machine and and challenge with the wordly No. Well, actually, Jonathan, I was going to add, since you're saying that, you know, you wouldn't tell participants what voice belongs to which scenario. But I think for now, at least with the machine interpreting solutions, they are mostly still very synthetic voices. And you can understand them. But you would make it. But that doesn't mean the solution doesn't work, you know. So Yeah, I Well, Think. That just been going back to quality, there has been quite a bit of research on voice quality and interpreting. And my argument always is you take the situation as it is and you give it a try as as. And I think there are some cases I mean, I was talking to people recently about kind of emergencies and misinterpreting or, you know, to just go into a library in a foreign country. There are situations where, you know, if an ambulance and someone if a paramedic needs to get a patient's name and work like we saw, we use whatever solution lets them do triaged. And then when they're actually getting treated, you get you know, you get the right solution for the time. So, yeah, I'm I would like to test things ises because I know a technologist well enough that the more feedback you get and the more that you can see in the wild and people reacting to, the more you can make a difference and also the more humans, human interrogators can adjust. I think we've given human interface as far too easy a ride for far too long. And chapters 10 to 13 of my book are a well-deserved poke. And I think if human interrogators don't get a well-deserved poke with war machine and therapy speech translation can do, then I think we deserve to get in trouble if we take it as OK. This is this thing. This is what it can do. Here's how we need to differentiate ourselves if we'd react the right way. It could be good for everyone if we react the wrong way. It's our own damn fault. Jonathan, you should send me a link to your book after the session. I'm curious to read it now. Just you have to promise me that you don't get annoyed for my simplification of how compute the computer models work, because I'm. Oh, I love simplicity. That is perfect. I love simplicity. I drew I drew a diagram. I drew a diagram. That was you, Drew. Well. I believe all of it, yes, all of the diagrams are hard, our hand drawn because corporate commissions are hard, wouldn't go there. I want to talk about some of one of the things that you brought up earlier, Jonathan. So we did have a conference where there was a human interpreter for certain languages and they use wordly as well. And I'm not going to tell you which which conference this was or it was, but it was a big conference. There are a lot of people attending. And then what they did during that session was something that was very revealing to us. Number one is the the organizers were able to audit us in real time. And they were able to audit the quality in different languages. They were asking, hey, how is this interpretation coming? What's color? How is it good? The second thing is that for the buyer who's buying or purchasing the interpretation, they're able to see real time. How many people are actually using our product in real time, in what languages that they're using and all of it. So the openness and the ability to audit in real time is quite unique. With mission interpretation, with human interpretation, they don't seem to be able to tell how many people actually use the headsets or how many are actually currently using it in real time or they just take it, or how many people did not take it. And, you know, it was just they were not able to provide the metrics These In real Are metrics Time. That on the IRR systems, you can't get by on your you. So there is no there is no interpreting into delivering the human interpreting into mobile apps. There's no reason why the metrics shouldn't be available there. And I think this, again, is a can of pork for human and tariff is that we are not used to thinking metrics. And yet there are some very important metrics that we should be thinking about. And so I was thinking, you know, I say by side comparison with the same language, with all the metrics available. You know, the technology to transcribe what the interpreters are saying that's widely available, that, you know, that's on my phone. I can do that just about it stumbles with Glaswegians sometimes. But, you know, as a as well as an interpreter, I'm an independent research it and interpreting studies. And the biggest issue with interpreting studies historically has been access to data. And so if we can find a way of getting access to data and getting these kind of metrics, I know researchers who would love to get their hands on that kind of data because it would make our lives a whole lot easier. But I see that as someone who's currently writing a paper questioning some of the issues with transcribing, interpreting. But we'll get to that on another show. And Jonathan actually wanted to come back to something you said earlier, you were mentioning this, I think really, really good example of, for example, using machine interpreting for that trial situation in emergency situations, in hospitals, you know, just for that initial bit of information, then afterwards, maybe for the consultation and you get a human interpretation. And Laxman, this brings me to a question for you again about what you see, like the main use cases for machine interpreting and where for humans out in real life. I know you already made the difference between the emotional stuff and information, but like, if we apply this, where do you see those, like Nicias, that maybe machine interpreting cells versus humans? So if you look at high density cities, especially for hospital, it's I mean, I'll come back to the case of the hospitals. And even with the covered situation that we have currently going on today, people are trying to find interpreters for patients and having to call an interpreter to do even the initial triage or the initial consultation before going for further treatment. They're finding it to be a very easy logistical challenge. The other is the second part of it is it is quite expensive. I mean, if you try and call through to in the US, you know, AT&T offers. You can call into an interpreter, language interpreter at any time, but it's quite expensive. That's a minimum fifteen minute or 30 minute requirement for any of that. So that we want to we want to be able to make it easier for the patients as well as the doctors because their time is quite valuable. I think there is a place where all of this can be done seamlessly with machine interpretation. And given what Alex said, just Alexander said just earlier, if we keep terminologies to, OK, this is a pediatric and I mean, this is a cardiac environment. This is a pulmonology environment. We can even narrow down the scope in that to provide a better ability to interpret. And the nice thing, as we do with that, may be an overkill, because when you when you talk of doctor patient conversation, doctors are typically conversing with patients in a very simple language. They're not domain specific language. They're talking general speak. And for that, we don't need to go into a big corpus. It's actually a really good point, because I always think of, you know, these like medical interpreting and also illegal, of course, maybe for legal, it's more applicable, but like for for medical interpreters, like, oh, my God, you know, the terminology that comes with this, a potential terminology anyway. But, yes, it's very true that when that might be true for, you know, doctors talking to each other. But when it comes to the doctor patient conversation, of course, you have to break it down for you. The layman. Though it's fairly simple language then in comparison. I guess then the only difficulty that you have left, and it's one I watched, I coached some sign language interpreters who were working with the health service for a few months and when they were doing their coaching and weren't working and so on, they found that the cultural aspect becomes even more important than the terminology. So in some cultures, a male doctor examining a female patient just is not done. And it's very difficult. It's easy enough for a human interrogator to interrupt and explain. It's much more difficult to set up a machine to realize, oh, hold on a minute. I see a man. I see a woman. I know this culture. That's not going to work. That's it. Again, it's the difference between the text and the context. Right. But also, I mean, that's an area, right, you're probably looking at the consultation, whereas we were, for example, the example we mentioned, it was more the initial diagnosis with a triage situation than you could bring in a human in later. I guess some of it could be done by machines later to in consultations. But, Jonathan, what if you don't have access to an interpreter? What do you do in such situations? It's not it's not always the case. Well, there's a mandate for people to have interpreters in hospitals. There have been situations where they don't they don't they don't allow. I mean, they're just not available. The second part, I'll give you a little bit of a personal experience. There are cultural issues where the patient, if they're don't know the language, they don't speak back to the doctor because they just don't know what is being said. Right. And they may understand what the doctor is saying in English. And case in point is my mother, she speaks. She understands English. She doesn't speak English so well. But so when she has to explain to the doctor, she doesn't give the full explanation to the doctor because, you know, she she has the conversation, she listens and she doesn't question back. But if she were speaking to the doctor in Thommo, which is her language, she would definitely be able to, like, speak up and be able to question. And so if you'd had an interpreter available all the time or if you had a machine interpreter, any either machine or a human, I think that conversation could be much a fear. And I think the sad thing is it is challenging and terrifying to look at access issues and to say, well, if they don't interpreters nearby, a machine is going to be far better than nothing in every situation. And I think that the sooner we admit that and the sooner we say, well, if we don't want to lose our percentage of medical interrupting, then training is over. The issue is not the job of V.I. to say, oh, we'll hang back until you get enough. And they're frustrated. That's that's not feasible. Now and then you have the market element as well. And the language axis element and I also like added expense. I think I mentioned this before. I was in the hospital in Spain when I was on vacation and had to go to the emergency room. And even though I speak some Spanish, it is not nearly good enough to go to the emergency room with ADD when you already. I was really out of it as well. I was sort of rated dehydrated and I wish I would have had some kind of a device like a machine interpreting device to at least help me along in that situation to get a little bit more for the doctors were saying and vice versa, because there was nobody who would even speak English in the emergency room. I had to wait until the next day for someone to speak English to me. So, you know, I got by. They treated me really well. I cannot complain at all about the Spanish system. It was fantastic, the care they gave me, but I could not communicate with them. So it was a bit frightening in the beginning. Yeah, this is this is true, especially in big cities and, you know, in Silicon Valley where I mean, this is close to home for for me. But every region of the world is represented here and getting access to interpreters for different languages from across the world for sure. It's easier to get the the the most used used languages like Spanish and English and perhaps French. But the other languages are much harder to come by to find interpreters who are qualified to do medical interpretation. Absolutely. I have an article coming out about this next week of Nimdzi, looking at the number of languages spoken in the US and how many languages are, you know. We look at interpreter certifications and most certifying bodies only provide certifications for like 20 languages or think for translators as about 30. But there's like hundreds. Three hundred fifty something languages being spoken. So there's a big gap there. Yeah. I can see Alex the Goring, it's time to take the coffee to bed. Yes. Yes. Yeah, I was just thinking it's I think we've come at a big question. I think we could certainly go on for for That A lot They. Longer and maybe we can follow up on that. What I thought was interesting is that we're we're basically revisiting the topic that we had in our very first episode, which which lacrimal? It was called Dictionary on Legs. And we were sort of riffing In our On, Very You First Know, Episode. What the situation back then was in machine translation. And I think this was when the Waverly Lapps pilot first came out, I think, or was about to come out or whatever. So, yeah, it's been it's been really interesting to come back to that. Also, you know, in light of recent developments, shall we say. So it's probably not a good idea to to, you know, talk about what the future will bring because nobody saw Coronado's coming. So if we can if we can really say what's going to happen next, I don't know. The. Unless you want to share anything cool. You're working on that. Yeah. There's not much I can say about it, unfortunately. That's what I thought. I think we just have to wait and see. But all I can say is that, you know, there's increased acceptance of people. You will see and hear more and more about us. And we expect we expect this to be a good platform that will scale quite well and can coexist with human interpreters. I definitely see space for boats to exist. I think that's that's one of the key takeaways. Would you agree, Jonathan, as the person who literally wrote I The book on this? Would definitely agree, I would definitely agree and I would make one appeal to interpreters, please spend more time reading about the engineering side of speech translation and learn to distrust the marketing side. It's just standard critical thinking and two and two companies who are making speech translation solutions. If you're not if you're going to say to us that you're not trying to replace us, I would beg you to say the same to journalists, especially tech journalists. Because it's damaging to everyone in the short to medium term to start talking about machine interpreting, taking over, because then less interrupters will train, the availability issues will get worse, and we won't have enough in terms of human interpreters to deal with the work that needs human interpreters. No, until the machines get a lot better. So can we please please call a truce? But gentlemen, I think one thing to keep in mind with that is even though I agree is somewhere, it is just also marketing, you know. There is marketing. I believe I wrote a chapter called Machine Interesting's marvelous but misleading marketing is. I think it's not just marketing, it's it's I mean, there are just some constraints, as you know, you need to raise money. You need to pay the bill. You need to invest money in research. So I think it's not quite that simple, really. And it's not entirely false to say that the solution does replace human interpreters for you event. Technically, for example, Yes. Yeah, it's not false. It doesn't say or replace them forever. For every It's Scenario. It's not false by where I think the thing is, is. Yes, no, marketing doesn't tend to land. We know. And, you know, V.C. guys tend not to take nuance. We all know that. But I think an understanding of seeing, you know, we are looking to provide a solution for or we can see a market gap in is a much more honest and a much more, to me, a much cleverer way of marketing than to say we're going to replace all interpreters. And then the very people who you might want to work with to improve your solution. I'm just gonna fall for dramas as say not so. Yeah, that's Yeah. A good point, you don't want to alienate interpret us, because, like we said earlier, we can all benefit from one another. And it would also be nice, I think, if interpreters were scared of this and would rather. I think I think that is that is a big thing, is that that's historically true, right, is humans are always afraid of some machine or something else taking up their job. But try to end with what Jonathan started is that how can it help improve and help them do something better, their life, if it's not InterpretBank, maybe something else or improve their quality of interpretation itself? I think those are the ways for people to look at it is not look at it as a threat. But, hey, look at that as something that can help them improve their quality of life as well as other people's quality of life. So can I. Can I virtually shake hands on a truce? I'm quite happy to. A virtual Spqr. Truth handshake. And There you I Go. Tell you what, one of these days I'm doing a set of five interviews on my own YouTube channel. I will keep your e-mail. Waxman and I will be inviting you for a live interview, because I think it's time we had a truce and we did stuff together. All right, super, yeah. Yeah, no, absolutely. I mean, I think this is where the the technology and people meet and and that's. Yes. And definitely I think there could be a good coexistence for both.