Hello, and welcome to the NVIDIA AI Podcast.
I'm your host, Noah Kravitz. There are roughly 1.8 billion Muslims living on planet Earth.
But only 16% or so of them speak Arabic, the language of the Quran.
This is in part due to the fact that many Muslims struggle to find qualified instructors to give them feedback on their Quran recitation.
Enter today's guest and his company, Tartile.
Tartile, who is a member of NVIDIA's Inception program, I should add, was founded with the mission of strengthening the relationships Muslims have with the Quran.
They're doing it with a fusion of Islamic principles and cutting-edge technology.
It's a story born from traditional values and modern day rituals like San Francisco hackathons.
And I'm delighted to have Tartil's CEO, Enes Abul-Eleven, here to tell us all about it.
S, welcome. And thank you so much for taking the time to join the NVIDIA AI podcast.
Likewise. Thank you. I appreciate it. So let's start at the beginning.
I referenced the hackathon, and I think there's a little bit of origin story before that.
But why don't you tell us how Tartile started out?
Yeah, so the name Egg and, you know, the original idea of Terceal as like an AI company started out in a, well, some hackathon in the Bay Area. where one of my co-founders organized and helped bring about the idea that how can we utilize state-of-the-art technology to solve problems Muslims face in their day-to-day lives.
And so a group of hackathoniers got together and came up with the idea of basically starting out to collect a data set of ordinary Muslims reciting the Quran that can be used to train speech recognition models, which can, you know, in the future be used to build the speech recognition applications, which is what we have today known as Sertir.
The idea of just speech recognition for recognizing Quran representation started about a year or two ago. with also one of my other co-founders, Mohammed Moussa, where he built an app called Iqraq, which is in Arabic means read.
And it started out as just using, you know, Apple and Google speech recognition APIs to search stuff on.
There was no like transcription or anything, just like basic, you know, keyword search.
And then a few years later, got in touch with my other co-founder Abdul Maltid.
They got together, you know, at the hackathon to build up, you know, this initial idea.
And then I joined about a month later to help push forward the MO side of the project.
And that's kind of, you know, the origin story.
And, you know, we worked on Tercel on the side as part of our day jobs when we were out in the day area.
How much was that Facebook? What was that Twitter?
And I was Amazon, Amazon Web Services. Once COVID hit, we, you know, we all went back home basically and we're like, you know, I think this is like a good time to actually focus on, on, you know, starting to deal full time and,
It was around, you know, 20, you know, 2021.
We left our full-time jobs and, you know, we started working at Tifty the whole time.
It's funny in the, I guess it's going on five years that I've been hosting the pod.
I feel like the first couple of years the story was always well there's this big kind of acceleration of of compute available you The hardware really took off.
And then that led to software tools kind of taking off.
And then this whole big... rush of people doing things with AI.
And more recently, it's kind of been, yeah, we were kind of working on this.
And then COVID hit. We had all this time at home.
So it's kind of interesting the areas of development.
So maybe you can walk us through just kind of the basics of what what Tarteal does and whether you wanna talk about what it does now, or if there's kind of an evolution you wanna trace.
And then as part of that, I'm curious, When you guys started out, was there a real lack of Arabic language speech recognition tools and other tools available to, you know, consumers. who wanted to, whether it was Quran related or just to study Arabic or You know, what were some of the problems that you faced as you started developing what became Tarteal?
I think early on, just in general, there was a lack of speech recognition frameworks and toolkits that the general population could use.
A lot of it was, you know, probably closed source used by, like, financial institutions or, like, TV stations for, like, trying to describe Apex in a real-time way.
And on top of that, in general, Arabic and Islamic technology has been kind of on the back burner.
There hasn't been a lot of research and work done towards supporting the language.
It's considered a low resource to language in general compared to other languages like English or Mandarin. which is where a lot of research goes.
And if you follow the money, you can see like, Look at Meta, look at Google, Baidu, you know, they're the ones who are pushing the research and, you know, they're utilizing, you know, the languages that they commonly use to do so.
Great. And so that was one of the main challenges we had to solve is like, how do we bring about this amazing technology that's being out there and A, make it applicable to the Quran, and then B, also deploy it in a manner that is usable by people.
So one of the very interesting constraints that we also had is when you recite the Quran, you want the feedback to be real time.
When you're sitting with like a teacher or an instructor, You don't want to wait like three, four seconds down the line and then the teacher tells you, oh wait, I made a mistake back there.
I'm not going to go back there. And so, you know, there, there was a real time constraint as well.
And so, Compute was a part of it. Frameworks are a part of it.
Data collection was a part of it. So we basically had to solve a lot of problems across the entire ML stack.
And that was like one of the major challenges we faced.
And so as we built out a lot of the tools and the frameworks necessary to do so. we were able to continuously solve these challenges.
And obviously with the help of NVIDIA, a lot of the toolkits, that we use are actually built on top with the technologies provided, such as, you know, Nemo and Reva.
I think, Reva specifically has been like a game changer for us because it's been able to actually give us that real-time feedback for our users.
Did you have to go about collecting data sets?
Because you're talking about some of the tools already being out there. but not being used, you know, following the money, the money not leading back to Arabic necessarily.
And I'm wondering about that process. And then my understanding is of Arabic is limited, but I understand that speech and inflection are quite important and important in reciting the Quran.
And just wondering if that was a data set problem or you kind of had to tweak you know, the frameworks and the technologies and the real time speech capabilities to hone in on those things that might not be as important in, you know, commercial applications of English speech recognition for that matter.
Yeah, no, it was both. So initially when we collected our first data set, we had issues with that.
We learned a lot in terms of like how we should actually be able to annotate it.
So there was a data set collection issue.
And also because Quranic recitation isn't like conventional spoken Arabic.
There's a lot of specific rules and nuances associated with this recitation.
So there are these rules called Tejweed.
Okay. which basically govern the way you're supposed to pronounce the Quran.
And then the way you recite it is supposed to be in a melodic and in a beautiful tone.
And that's actually where the word tersir comes from.
Tersir is to recite the Quran in a beautiful and melodic tone.
Oh, got it. Yeah, and so the data set had to be annotated in a certain way.
You know, there's certain... Like I said, to read rules that we had to annotate accordingly.
There's certain diacritics, basically battles that we also had to annotate properly.
Generally, English is, you know, the vowels are like the A, D, I, O, U, so they're separate characters.
Versus with Arabic, you can have multiple characters.
And then different vowels, different pronunciations can actually change the meaning of the word completely.
Like if you say ah instead of ooh, Like you, you basically said a completely different word.
It's a different word. Exactly. So it was, it's a hard, it's a hard problem to solve.
Like, especially we even argue internally sometimes it's like, When we're evaluating different models, it's like, did the model get it right?
Or did the person say it right? It is very nuanced.
So we had to solve it on the data level, on the annotation level, but also the models, the tokenization.
There was a majority of factors or changes, but certain architectural decisions had to be made in order to be adapted for this decision in this case.
And so when did you and your co-founder sort of decide to, I know that, you know, during the COVID lockdown, there was, more time and bandwidth to spend.
When did you guys decide to leave your, you know, big tech jobs and focus on turning Tartiel into a company?
I think it was a combination of two things.
One is when we were confident that the technology to build out is possible.
But B, also we had some form of traction.
And so the app version was free. And, you know, we had like, you know, uh wait lists and stuff for like building out correction and people were always talking about it And so when we saw that there was some market validation basically saying like, okay, this is what people want.
And I mean, we kind of knew it ourselves internally, but the question was like, as a business is like, will people actually pay for something like this?
And so really when we had some market validation, basically people were like, okay, it makes sense now that you know we need to like focus on this full time and we need to get it to market quickly because also the thing to keep in mind is like While no one has worked on this problem before, if you ask almost any Muslim tech company, person, you know, what they would like to work on.
This is probably like the first idea that comes to mind is like, how can I use speech recognition to correct some substitution of the front?
Because it impacts like Literally every Muslim.
Hundreds of people have the idea. It's just people haven't been able to execute on it.
Governments have, you know, wanted to work on this idea.
Institutes have wanted to work on this. It's just, it's acknowledged either wasn't there.
Uh, the, Team wasn't there. You know, the market timing wasn't right.
And I think we've got those stuff in order now.
And that's when we decided to go about it.
And so the app is available now. It's iOS and Android?
Correct. It's available for free. And just to clarify, because a lot of people have a misunderstanding on this, it's like The Quran and the conventional features associated with the app are available for free just for reading, basically.
But the mistake detection, the unique value proposition, basically, being able to actually identify your mistakes in your recitation and tracking your progress, that's a premium feature.
So do you offer different, I don't know if versions is the right word.
I'm thinking, you know, tech wise version control, but do you have a... different versions you offer or parameters that users can set?
How do you solve for that? So right now, we just solved for one recitation style and one rendering of the prong.
We're keeping it simple. If we do that, it's a combinatorial explosion of... yeah of all the different kinds of you know types of you know books and styles and and Yeah.
So we're sticking to one, keeping it simple.
And then later on, you know, as a quarantine, it's something that can solve for.
Right. But we do also offer different purchasing prices.
So we do localize prices to different regions, for example.
In terms of the actual reading of the Qur'an, we stick to one. is like for example one of the reasons why we don't explicitly like advertise ourselves in Southeast Asia like India and Pakistan because they have a specific rendering called the Indo Park.
And so we know for a fact, if we decide to market there, we're going to get one star reviews.
It's like, why don't you have this reading move?
Everyone uses this type of font, but they don't realize there's like 10 different ways of writing the fonts.
Cool. That was great. I'm glad you brought that up.
So one of those things that was percolating in my back brain, but I wasn't near kind of knowing how to phrase it.
So I'm glad you brought that up. Walk us through it real quick.
How does it work from a user perspective?
Yeah. What I love about our app and what we built is like, Anyone, and we've actually, if you go on TikTok or something, you look up any, like the demos people do, literally they tap a button They start reciting and, you know, it's like magic basically.
Like Tarsil automatically picks up what they're saying.
It knows where they're saying it. and then just highlights and tracks what they're saying in real time.
And so they basically follow along, you know, in the app as they start reciting the Quran. the app highlights where they're reciting and it buzz along with them.
And if they make a mistake, it highlights it in red.
If, you know, they're in what we call memorization mode, people like to memorize the bran, so it hides the page for them.
And it only shows them what they've recited so far.
If they make a mistake or they skip a word, it highlights it.
And so every time someone shows a demo of the app, like the magic moment is literally just the tapping that button and everyone just like, you know, minds blown.
It's like, Holy moly, this thing actually recognized what I was saying.
That's fantastic. And there are some gamification elements?
Yes. So this is actually some of the things we're trying to do is that, you know, Muslims, especially out here in the West, you know, busy people, never jobs, never families and all that stuff.
And so building a relationship with the Quran isn't, the most, you know, not usually, you know, it should be top of mind, you know, especially as practicing Muslims, but it's not always top of mind.
I mean, it's kind of like exercising. It's not like workout.
You should be exercising, but not everyone goes to the gym every day.
So the same thing with spiritual habits is, you know, we're trying to instill, you know, gamification elements to get people to build a habit with the Quran and build, you know, a a more stronger relationship.
Some people call it, you know, materialistic or, you know, you know, you're avoiding the spiritual aspect, but at the end of the day, like you got to do what you got to do to like actually help you. build a stronger relationship with your faith.
And I think whatever works for you, like you should do it.
Like you shouldn't shy away from it. Yeah, it's an interesting thing, thinking about making spiritual practices and sort of say old in a negative way, but you know, long established sort of old world traditions that, as you said, people think of, you know, well, I should be doing this every day, but it doesn't really fit in. into my modern life and, you know, figuring out ways and I'm sure walking some lines of tension between folks who might applaud, you Well, yeah, tap into what people are doing today.
And, you know, any way to get the spiritual practice happening is good and sort of more traditional traditionalists.
As you started, as your team started developing the app and working towards what's now these magic moments of the... of the real-time error correction, what were some of the either biggest problems, hurdles you had to overcome get past, or maybe just things that surprised you from a technical standpoint as you were developing the app?
Well, internally, the magic moment for us was when we got our first model deployed and actually working in the cloud and everything.
That was like a huge, huge win. And so we were just like, wow, like model A can actually recognize Quran and then B can actually be deployed in real time.
So that was... That was pretty cool. What was very interesting actually is that, and this is why, you know, I have a bias against like all these NLT and, you know, uh, recent ML practitioners is they have it easy.
It's like, you know, you guys are working with text and, you know, structured data.
It's like, we're working here with audio.
It's like terabytes and, you know, huge data sets. and it's very hard to annotate and label.
And so a lot of advancements have been made in the language modeling space, the text and NLP space, you know, I think Facebook personally released like, excuse me, better, I like NLLB and all these different models, but there hasn't been a lot of work in a speech recognition space.
So we've had to build a lot of tools and hacks internally.
But the other thing was streaming speech recognition is not like conventional streaming ML that you would see in other applications.
Like when people talk about streaming ML, it's like, And so it makes it like, okay, how do you badge?
How do you, you know, get results back quickly or like via sockets as opposed to like, you know, ADI calls or asynchronous calls.
With streaming speech recognition, the different parameters associated with streaming, such as the chunking size, the padding size, as you provide a window basically of audio you feed to the model. those different parameters can drastically affect the behavior and the end result of the user experience.
And so internally, for example, like we have different modes where Accuracy can be traded off of latency.
And so if you want a model to perform very fast, we give it, for example, a smaller chunk size, but the word error rate might not be as good. versus a model that we want to be more accurate at the cost of latencies.
Like we have a bigger window, bigger chunk size of the audio that we feed to the model and at the expense of latency, but it gives us a more accurate result.
And so, you know, that's kind of like some of the interesting like optimizations that we have to explore and evaluate for the purposes of our of our application.
Are there big differences depending on the device that the end user is using, whether it's a more recent phone, couple of generations old.
I know on the iOS side, things are relatively standardized Apple chips, but Android, you know, there's a million different devices running running at least a few different chipsets.
That's all the ballgame here. No shots at Android.
I just know from my previous life reviewing mobile devices that Android development is a different can of worms.
Yeah, exactly. There's so many devices out there. uh, and then Huawei decided to like make their own set of stores and stuff.
And so that, right, right, right. But no, uh, Right now, everything runs in the cloud.
And so really the bottleneck is latency and us processing the words.
Yeah. When we are exploring those, you know, apples are going to come up with like coromel, coromel plukits.
Mm-hmm. some very interesting frameworks for putting, you know, for utilizing this with the neural engine on device.
And so we've been exploring ways that we've been building the model on device.
We found that phones from five years moving forward can potentially have a real-time factor of like greater than one.
So like basically almost real-time performance.
But we haven't tested much on Android just because, you know, frameworks and everything is pretty diverse.
And so we've actually been exploring on-device ML for iOS for a few months.
And it's something we probably will be pushing forward in the future.
And I think it's possible. I mean, Apple has their own speech recognition toolkit. which runs on device.
I mean, Sierra runs on device sometimes, or we're in the cloud if it has connection.
And so it... The real-time factor is possible.
The question is like, you know, it's just the engineering around it that's needed.
So we touched a bit on some of the technical challenges involved in getting the project started and gathering your data set and training your models.
But when you went to market, I don't know, what are some of the challenges you faced and whether these are things that are kind of common you know, anybody bringing a new app-focused business to market or maybe specific to entering the Muslim market with Tertullo?
Yeah. So interestingly enough, a lot of the problems that we have, like they're basically your typical startup problems, like, you know, product market, you know, getting in front of your customers, talking to users and all that stuff.
But given the fact that there isn't a lot of like research of all of some market in general, It's very easy for you to find how many cups of Starbucks does this demographic buy in a day. but it's not easy to ask the same question.
Like how spiritually, you know, are you like how religious, you know, do you, do you visit like church or do you visit the mosque every day?
So these kinds of like demographic information, it's like stuff we had to like figure out on our own, just like understanding who the target audience is and who we should go after.
So the data wasn't there, so we had to even extrapolate on our own.
But I think what was very interesting is there's a lot of nuance associated with the Wilson market distribution, depending on where you're at.
Depending on where you're at, most of them behave very differently.
Sure. So those out in the West, you know, they're typically... more open-minded and you know they they like to you know explore they're they're open to like having you know different kinds of discussions and explaining things versus when we went to like the gcc or like in egypt it was uh like people obviously love the app they love the way it works but obviously they had their own like concerns like Oh, you know, do you have like a religious ruling that this is allowed or did you make sure that, you know, the court.
Quran is in this writing or is in this format.
And then each region has sometimes its own renderings or its own transcriptions of the Quran.
So people would ask us, Can you add this kind of transcription or can you add that kind of reading and so on and so forth?
While the Quran, as a text, is standardized, the way it's written, the pages it's on, and the way people recite it, It varies from region to region.
And on top of that, obviously, the purchasing power, the the mentality behind paying for religious applications is completely different.
So there's a lot of like communication and, insights, you know, I've had to like discover and solve for that are, they usually don't find in different, you know, applications or markets.
How big is the Tarteel team right now? We've been four people for like a good two years.
It's only last month that we just hired two engineers.
And so... Oh, congratulations. Well, yeah, just one global engineer and one ML engineer.
We're extending, you know, an offer to like a marketing manager.
So basically someone who can work more on like the marketing, the product and the strategy side as well.
So yeah, so we're just at six basically.
Three of us are in Canada. Two of us are in Boston.
And one is in the... Well, he was in the UK and now he's in Dubai, so...
So without asking you to reveal anything, you know, proprietary and stealthy, what are the plans for the company going forward?
So short term is basically solving some of the few technical challenges associated with speech recognition for So like some of the, like I mentioned, the Tijweed rules, the diacritics rules and whatnot.
So there's still some unsolved challenges there.
But I think overall, we're in a good position right now to be at the forefront, basically, of revolutionizing Islamic technology.
We like to call it the new Islamic golden age.
We talk about back then that a lot of especially like during the Ottoman Empire and even before, a lot of like technological innovation was done by the Arabs, by the Greeks and, you know, By the Europeans, there's basically a bunch of cross-collaboration contamination.
We like to call that the golden age. I feel like right now we're at a position where we can revive that from a religious perspective, where we can actually start to build the necessary technology to help Muslims not only build a stronger relationship with the Quran, but build a stronger relationship with their faith, break down, you know, misunderstandings and, you know, things that have come out of context, taking things like semantic search, taking things like image recognition for recognizing different Excuse me, different, like, styles of writing the poem.
There's different, like, you know, ways of writing the transcripts at Harrogate.
So... There's a lot of technical innovation that can happen.
Like one of the things that we love to talk about also is like, how do we take Web3 and blockchain and these technologies to the Islamic, you know, landscape?
And, you know, an example is like IPFS. It's like, We have all these digital assets for the Provence.
How do we make it accessible to everyone so that it's not censored by... any institutes or governments, but also how do we, you know, make sure that things are standardized, make sure that we know that we're getting the authentic source of Iran, because some, you know, small mistakes can be very, you know, very, very big, you know, very big things, you know, it comes from like a spiritual or religious perspective.
And so, How do we make sure we can verify the authenticity of digital assets?
How can we make sure that it's available for everyone?
How can we utilize semantic search? So there's a lot of open questions out there and a lot of engineering that needs to be done in terms of like bringing state of the art technology for Slavic applications and Hopefully, I would like to think that we can also solve for wider problems, not just necessarily for Arabic, but like other local resource languages, you know, other dialects. as well as like other religions.
Like I'm pretty sure, you know, the Christian community, the Jewish community have their own set of tools, have their own set of needs.
And it would be great if, A lot of the state-of-the-art technologies built were like financial institutions or social media.
But at the end of the day, there's a spiritual element to being a human and How do we bring that, how do we bring that, the use case in there?
Absolutely. I didn't bring it up mostly because I didn't want to interrupt you, but also I... wanted to be sensitive about not sticking my foot in my mouth, comparing religions and faiths and such.
But when you were talking about the app experience and all of the tones and inflections that go into and the vowels and things like that with reciting the Quran, it reminded me of my own experience. and studying half Torah for my bar mitzvah when I was younger.
I'm not particularly religious, spiritual, I'd say, but not religious these days.
But a lot of that sounded so familiar to me, learning, you know, it's not, there are marks in the text.
Any Judaic scholars listening will be thinking like, I don't know the words and I don't, but you know, just things that dictate how you, how you recite it and it's very melodic and everything.
And so I studied with you know, a teacher after Hebrew school who worked one-on-one and I had tapes and I was basically... it was as much memorization as recitation to learn my particular part and to try to do the melodic singing of it.
So it was interesting just recalling those experiences and thinking like, Like, okay, so we may do with a cassette tape and a Walkman, but if I had real-time error correction guiding me, my gosh, you know, like how much faster and better. the experience might have been.
So it's very, very cool to hear about. And it's fun.
I had not come across the term religious technology for some reason until, you know, prepping for this conversation.
So, yeah. Very, very cool that you and your colleagues are focused on this.
I think it's great. I mean, it's very new.
No one has really thought about it. I think only recently, like, glorifying how I started to come up with content-based applications and platforms about I think there's a whole ecosystem just untapped.
There's a whole lot of value add that could be brought to people, you know, of different faiths.
And so I'd like to think that we can help bring that forward and, you know, bring it to life.
Ennis, this has been a delightful, really fascinating conversation to have and, um, Again, the combination of the technical stuff you guys are doing, but then sort of exploring what what sounds like new territory for at least a chunk of your users out there, I think is really great to see.
For folks who want to learn more about Tarteal, the app, the business, the technology and research behind it.
Where should they go online once we're signed off here from the podcast?
Yeah, I mean, we have a website, touchfield.ai.
We also have a blog post where, you know, we talk about some of the insights we've had, you know, like, previous user experiences or even technical blog posts on our AI and ML.
So our blog.touchdl.ai. but i also recommend you know we we use like nvidia nemo and reba extensively so Reading up on, like, the papers where they had to, like, sit you down and confirm your models.
And even the stuff, like, you know, Hugging Face puts out and Minute UI in terms of, like, we've developed.
I think are great resources to kind of like learn about a technology and just to turn the company website is, is honestly the best place to go.
We also have a bunch of news articles and blog posts written on us, so you can also find those there.
Perfect. Tartile.ai. Well, again, Enes, thanks so much for taking the time to come on the program.
Best of luck to you and all of your teammates in the future.
Thank you. Thank you. Thank you.
