🌍 AI in Africa - Voice & language tools

In the third of the “AI in Africa” spotlight episodes, we welcome Kathleen Siminyu, who is building Kiswahili voice tools at Mozilla. We had a great discussion with Kathleen about creating more diverse voice and language datasets, involving local language communities in NLP work, and expanding grassroots ML/AI efforts across Africa. Sponsors: Fastly https://fastly.com/ – Our bandwidth partner. Fastly powers fast, secure, and scalable digital experiences. Move beyond your content delivery network to their powerful edge cloud platform. Learn more at fastly.com https://www.fastly.com/?utm_source=changelog&utm_medium=podcast&utm_campaign=changelog-sponsorshipChangelog++ https://changelog.com/++ – You love our content and you want to take it to the next level by showing your support. We’ll take you closer to the metal with no ads, extended episodes, outtakes, bonus content, a deep discount in our merch store (soon), and more to come. Let’s do this https://changelog.com/++! Featuring: Kathleen Siminyu – LinkedIn https://www.linkedin.com/in/kathleen-siminyu, X https://x.com/siminyu_katChris Benson – Website https://chrisbenson.com, GitHub https://github.com/chrisbenson, LinkedIn https://www.linkedin.com/in/chrisbenson, X https://x.com/chrisbensonDaniel Whitenack – Website https://www.datadan.io/, GitHub https://github.com/dwhitena, X https://x.com/dwhitenaJoyce Nabende – X https://x.com/j_nabendeShow Notes: “Introducing myself” from Kathleen https://foundation.mozilla.org/en/blog/kujitambulisha-kathleen-siminyu-introducing-myself-kathleen-siminyu/Masakhane https://www.masakhane.io/Common Voice https://commonvoice.mozilla.org/enOpen for Good Alliance https://www.openforgood.info/#StartUpcoming Events: Register for upcoming webinars here https://practicalai.fm/webinars!

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  We have many young people who this may be their first interaction with any AI topic or their first interaction with NLP. And we pretty much have a great pathway for them to advance academic careers. Many have gone on to get acceptances to masters and PhD programs, which are amazing. Basically, there's a really great pathway for people who want to advance research careers. But I'd like us to build a similar pathway that is just as strong for individuals who want to productize and build companies.

• 0:35Speaker 01

  Big thanks to our partners, Leno, Fastly, and Launch Darkly. We love Leno, they keep it fast and simple. Check them out at leno.com/Changelog. Our bandwidth is provided by Fastly. Learn more at fastly.com and get your feature flags powered by Launch Darkly. Get a demo at launchdarkly.com.

• 0:52Speaker 02

  Hey, Jared here. One of the things we can count on in the software industry is change. The state of the art changes so fast, in fact, that keeping up can feel like a whole other job on top of your actual job. That's why we created Changelog Weekly. It's our totally free newsletter that we drop in your inbox each and every Sunday. We link to the latest news, the best articles, and the most interesting projects that you should be aware of. We also add a little commentary from us saying why something's important, pointing you to other instances of a trend, or just making a dorky joke to keep it lively. So if you haven't yet, I recommend subscribing to Changelog Weekly and help us help you keep up with the latest.

• 1:35Speaker 02

  Head to changelog.com/weekly and sign up today. Again, it's totally free and we never spam you. Yuck. One last time, that's changelog.com/weekly.

• 1:58Speaker 02

  Welcome to Practical AI, a weekly podcast that makes artificial intelligence practical, productive, and accessible to everyone. This is where conversations around AI, machine learning, and data science happen. Join the community and Slack with us around various topics of the show at change.org/community and follow us on Twitter. We are at Practical AI.

• 2:24Daniel Whitenack?

  Welcome to another episode of Practical AI. This is Daniel Whitenack. I am a data scientist with SIL International. And I'm joined, as always, by my co-host, Chris Benson, who is a tech strategist at Lockheed Martin. How are you doing, Chris? I'm doing very well today, Daniel. How are you doing? Doing wonderful, because today is another episode in our series of sort of spotlight podcast episodes, AI in Africa. This is a joint initiative between the International Development Research Center in Canada, Practical AI, and the Change Log, and GIZ, Fair Forward Project, all of which

• 3:04Daniel Whitenack?

  are sort of involved in one way or another with the Open for Good Alliance. And as part of these episodes, it's been really wonderful to have with us sort of guest co-host, Joyce Nbede, from the Macarara AI Lab at Macarara University. Welcome, Joyce.

• 3:22Speaker 04

  Thank you, Daniel. Hi, Chris. Hi. Thank you.

• 3:25Daniel Whitenack?

  Hey there.

• 3:27Speaker 04

  It's just another opportunity for us to discuss on this podcast, and I'm excited to be here.

• 3:32Daniel Whitenack?

  Yes, it's always wonderful to have you with us, Joyce. And why don't you, I'll sort of pass it over to you to introduce what we'll be talking about today.

• 3:42Speaker 04

  Right, thanks Danielle. So over the past podcasts, we've been trying to look at different studies on practical AI, particularly on the African continent. And we are very excited for this episode because we are going to focus on AI, but looking at community building in Africa and especially looking at the very exciting field of natural language processing. So for this podcast today, we are very happy to have Kathleen Siminyu with us. Kathleen is currently Akiswahili Machine Learning Fellow with Mozilla. But before Mozilla, there's a lot of work that Kathleen has been, are involved in. So Kathleen, you're very welcome to this podcast.

• 4:22Speaker 00

  Thank you for having me. Thank you very much.

• 4:25Speaker 04

  Alright, so I think we're ready to dive in. So we'll have, you know, just over to you Kathleen, tell us, you know, just an introduction about NLP for African languages. What work have you done in NLP? How has it been like for the last couple of years, you know, in the space of NLP for African languages?

• 4:45Speaker 00

  Okay, thanks again Joyce. So I'll start with a bit of an introduction to myself and how I got into NLP. My background is in math and computer science, and with that I went into data science. I worked briefly in industry for a company in the telecommunications space that headquartered in Nairobi in Kenya. And one of our products was SMS. And through working at this company, I came to realize that support for African languages when it came to NLP and just tech tools or digital tools, if we could say so, was pretty lacking in comparison to say your English.

• 5:28Speaker 00

  And this got me pretty interested in NLP, and I wanted to build particularly resources for African languages. So yeah, this is where my interest began. And when I ventured into academic research at this point, my next role was with IDRC or other AI4D. So I was regional coordinator of AI4D for a couple of years. And this work involved mobilizing funding for communities, let me see. IDRC had had funding which is going into AI research in Africa. And I helped a lot with that in terms of identifying communities that Could and would make use of of this funding.

• 6:09Speaker 00

  So I'll backtrack a bit and say that during my my role in industry, I did a lot of community work. I began a lot of community building, first with the Nairobi Women in Machine Learning and Machine Learning and Data Science. So this was sort of my first foray in the field as I was a data scientist as well in industry. And then along with transitioning into academic research, I encountered Masakane. And they were doing exactly what I wanted to do in terms of focusing on resources for African languages. So my organizing also transitioned into that. And in the middle, I I got, I attended several of deep learning endeavours.

• 6:50Speaker 00

  So the endeavour is movements to strengthen African machine learning. One of the events they have is an annual summer school. And I attended one, the first one in in South Africa in Stellenbosch. After which we got to talking, they realised that there was an M L A I community in Nairobi, courtesy of the meetup I was organising. And so we organised the second one together in preparation for the third one, which we then came to host in in Nairobi in twenty nineteen. So I've spoken a lot, but then let me say our capes by saying that I I have worked a lot with M L and A I communities in Africa. And that brought me to the intersection of my interest in N O P and the

• 7:35Speaker 00

  existence of Masakane with, you know, a focus on African languages. And I'll say, here we are, it's probably where I spent, I spend most of my organising time at the moment. I'm very excited by the fact that we have a wealth of diversity from the African continent in terms of people who are working on languages and it's languages that they care about. So also a wealth of diversity in terms of African languages included. I'll plug in the fact that as recently as last week, we finally have Masakane registered as an entity. So we are officially the Masakane Research Foundation. Up till this point, we've been

• 8:18Speaker 00

  Really just organizing as a, as a group of people and that's been great. But yeah, I guess twenty twenty two is going to start, it's going to be the beginning of us starting to see how we can formally organize and beyond running on volunteer capacity, what it looks like when we organize and we can put funding to some of our planning. I will stop there. I kind of forgot what the question was, so I'll throw it back to you guys.

• 8:43Daniel Whitenack?

  That's so wonderful. I really appreciate all that context, Kathleen. I wonder for those that maybe, thinking of those out there that might be listening to this podcast that have never, ever heard of Masakane and the things that they're doing. Could you just kind of give a picture of, you know, who's involved in Masakane? How do they interact and what are the sorts of activities that they're, that they're doing?

• 9:08Speaker 00

  Okay, so the Masakane was founded in twenty nineteen. And this actually took place at the Deep Learning DABBA that was hosted in Nairobi, Kenya. And it was founded by Jade Abbott and Laura Martina. So basically they did a paper for machine translation involving a bunch of South African languages and created benchmarks. And part of the work done in this paper, they came to present at the Deep Learning DABBA with the idea that if they create a notebook, which is very easily replicable, then people can take it up, take it upon themselves to train similar models for their languages. Therefore, likely creating fast benchmarks because very little research exists for African languages.

• 9:54Speaker 00

  And that is how it began, as a focus on machine translation. The initial work that they did was based on the JW three hundred data set. So JW is Jefovas Witness and they're an organization that through their work, evangelizing in Africa for many years. Have translated the Bible to many local languages in Africa. So the notebook leveraging this data set made it super easy for literally anyone to start up the notebook, put in the language code for the language you would like to work with and in an hour or two have a benchmark trained. And it gained a lot of momentum in that way.

• 10:34Speaker 00

  I I like to think that even though many of these languages have not been the point of focus in research in the past or in terms of product development for markets that can, you know, pay for products, they haven't been a focus from that perspective, but this gained a lot of momentum because of the setting, right? Because the planning indaba is bringing loads of young people from across the African continent who, given the fact that we'd had two previous indabas, had basic skills, right? So some efforts in capacity building had been done. And it it was sort of a sweet spot in terms of many of us had attended previous planning indabas and we have the basic skills and we now have a desire to specialize in something.

• 11:19Speaker 00

  And here come Jade and Laura and they have a notebook and they're telling us, hey, this is how you can start in machine translation. So it gained a lot of traction. Personally, I trained models for Kenyan languages and I think that's how many people rationalize in terms of what languages do I care about? What languages do the communities I'm adjacent to speak? And then went on to, you know, do that. If you could find the data set for your language and you went ahead and trained a machine translation model. We wrote a really great paper, two, I think. One I'd like to highlight is participatory research for low resource ma- machine translation, a case study in African languages.

• 12:00Speaker 00

  And this one in particular because it it describes the ecosystem for machine translation to be successful and why there's many gaps in an African context. Right, so one large problem or one problem is access to data. And we talk about the fact that typically content creators create data, right? But in an African context, these content creators may not have access to keyboards or they may not have access to digital dictionaries, which hinders their development of data to start with. So there's a whole ecosystem description and I'll leave it at that for now.

• 12:40Speaker 00

  But then we've seen great success, first in machine translation with being able to do very multilingual and inclusive work. And then progressing from there, we realized that individuals who were participating and contributing to Masakane were not only interested in machine translation, but other tasks in NLP as well. So at this point, we sort of took a step back and generalized and now have membership or participation from people building in NER, in speech, and basically just a wide scope of tasks.

• 13:19Chris Benson?

  Got a quick question for you. I was interested in the data set that you mentioned, the JW data set, being a Bible translation. When you look at a lot of different Bible translations and kind of the language may not be completely what you typically would talk in today. Did that present any kind of challenge or was that generalized enough to where that that didn't affect it or anything? I was just curious about that being the basis and and what challenges that might present that were unique to the data set.

• 13:45Speaker 00

  So that's a super interesting question. We encountered a phenomenon which I don't know if we called it or it is known as the biblification of system. So one example I remember is, actually, I don't remember what word, but in machine translation system, a certain word kept on being changed to canon because canon shows up many times in the Bible, but you don't not referencing canon often in. Conversation.

• 14:12Speaker 06

  Right.

• 14:13Speaker 00

  So it has been a challenge to rely only on Bible data. And it's to answer your question, it does not or has not generalized well. Although actually, interestingly, after a year or two of us using the GW three hundred data set, the organization actually pulled it from the internet. So it was scraped from the website and the organization claimed that they were not aware of this data set existing and had not given permission for it. So, I mean, we've been trying to ask them for many months to see if they would make it open anyway, because it has spurred a lot of our research and we're still waiting to see how that turns out. But then, yeah, just to bring in the fact that access to data is a huge problem.

• 14:57Speaker 00

  And in many cases, we face, you know, IT and copyright issues, even with data that is accessible. So at the moment, a lot of activities unfortunately have to start from a point of data set creation, which I wish was not the case, but we also have found that it's an easy way to start including and upskilling individuals, because someone can start out on our projects today labeling data or creating data or evaluating a model, but then can progressively then pick up other more valuable skills and, you know, grow their skills. So I'll say it's a blessing in a cast, but mostly a cast, because as LPL researchers, I wish we had the luxury of, you know, just accessing data for whatever task it is we

• 15:42Speaker 00

  wanted to work on.

• 16:01Speaker 02

  The change log is deep discussions in and around the world of software, and it's been going for over a decade. We interview hackers like Chris Anderson from 3D Robotics.

• 16:11Speaker 07

  At the time, drones were like predators and global hawks and military industrial. They were classified and super, you know, ten billion dollar things. And we had just built a drone with Lego pieces around the dining room table programmed by a nine year old. And it's like, okay, that should not be possible. You know, it's not, when, when a nine year old can do something that is classified, that literally export control it as ammunition with Lego, with toy pieces, it was something important in this world has changed.

• 16:40Speaker 02

  Leaders like Devin Zogel from GitHub.

• 16:43Speaker 08

  In the like 10 to 15 year range or 20 year range, what I would really like is for if you have like three 12 year olds hanging out and one of them is like, I want to be a firefighter. Another one's like, I want to be a lawyer. I want one of them to say that I want to be an open source developer.

• 16:57Speaker 02

  And innovators like Amel Hussein.

• 17:00Speaker 09

  I've yet to kind of see applications at scale that don't use multiple languages, that don't you have just arcane stories behind why this weirdo thing exists. You know, like, all right, when you open this file, you're going to have to turn around three times and tap your nose once.

• 17:17Speaker 02

  There be dragons.

• 17:18Speaker 09

  Just the most hilarious stories, you know, but applications are living, breathing. They have craft that's normal. So I want to normalize weirdness because that's just how applications evolve over time.

• 17:32Speaker 02

  Welcome to the change log. Please listen to an episode from our catalog that interests you and subscribe today. We'd love to have you with us.

• 18:06Daniel Whitenack?

  So Kathleen, I'm actually in the Masakane Slack group. And I just took a quick look at the involvement. I see in the Masakane Slack group that there's, at least at the time I'm looking, there's 1,304 people in the Slack group, which is pretty amazing that this community has grown in the ways that it has. And I'm wondering if you have any, you know, looking back on how the community was formed. And the activities that you did and how you welcome people and those things. Did you have any insights into kind of why you think it kind of grew, has grown the way that it, you know, that it has?

• 18:48Daniel Whitenack?

  And maybe looking forward, kind of characteristics of the, now that it's an entity and a foundation, characteristics of the community that you would like to, you know, ensure that they are characteristics of the community, you know, moving forward? Any thoughts?

• 19:05Speaker 00

  Yeah. So in terms of it growing, I'll say it was unprecedented, even for me. I will attribute it first of all to Jade's never-ending energy. I swear, Masakane would not be what it is without Jade. I sometimes even struggle with just keeping up with all the conversation that goes on on chat. It can be very overwhelming. I think one of the reasons why it has grown again is just the fact that these languages are very underrepresented in digital platforms, but that is not reflective of the communities that use these languages. They say there's 7,000 languages on earth and maybe 2,000 languages in Africa. And many of those are living with thriving language communities that use them every

• 19:50Speaker 00

  day. They are just not used on digital platforms. And unfortunately, because of, I think, social linguistic factors, they're not used in sort of formal spaces. At least in Kenya, I know for a fact that parents will, are more likely to encourage their children to become better at English than say their mother tongue. Because with English, you can walk into any office and have a conversation and potentially get a job versus your mother tongue, which is only useful at home or, you know, in whatever local context you are founded. But then that doesn't take away from the fact that these languages are used and at least from a sentimental point of wanting them to be preserved and sort of captured on digital platforms.

• 20:33Speaker 00

  I think that's one huge reason why Masakane has been so relatable to students and budding researchers across the continent. I'll also say we have a lot of room for absolute beginners. Again, it's really easy for someone to walk in today. Young people have loads of energy, right? And many will come and they'll say, hey, I want to do something. And you know what they can add to literally every single project? The fact that they're multilingual, right? So if we have an N R project and you walk in today and you say, hey, I want to do something and I speak Kisahili in addition to English. Then we can say, hey, look, there's this project where you can go and here's a Kisahili text and you can start labeling. And that's a way to start involving someone.

• 21:17Speaker 00

  And they'll be interested in, okay, what else can I do? What's this person doing? What's that person doing? So I think just the fact that we, literally, anybody who is multilingual can participate. And then we've been progressively working towards pathways for capacity building. So we have several study groups that now run every three to four months. There's one that's particularly for beginner N L P. So any beginner who's involved in any, you know, something in the community can also plug into that and start gaining more skill. Last week I was on the call and Julia offered to do one for machine translation.

• 21:55Speaker 00

  And I'll say, let, let that be a segue into another factor because we've, we've gained quite some recognition globally. And the wider N L P community has also been super supportive. So in terms of more experienced researchers wanting to know how they can be of use and then proceeding to actually support us, right? So Julia has a PhD, she works at Google Translate and she shows up probably for ninety percent of Masakane activities. If you ever want to debug, I remember my experiences with training those initial models for Kenyan languages. She was literally always available on Slack. So I'll also shout out the fact that we received a lot of support.

• 22:39Speaker 00

  A lot of mentorship support and that continues to happen. And I hope it continues to happen for an extremely long time. So the second part of the question, now that we are a registered entity and are growing somewhat exponentially. Characteristics that I would like to see sustained. First, I think is is a very distributed nature of leadership. So, Masa Khan is like a, it's like a storm. It's not a coordinated storm, it's just something that started happening and now it's got a life of its own. I love the fact that it's not reliant on any one person. So there's never a day where person X is sick and okay, now we can't hold the meeting because person X was supposed to chair it.

• 23:22Speaker 00

  That's never happened because literally everybody is empowered to walk in and chair the meeting because we have a template, it's recorded, everybody else can catch up, right? And we see that a lot also with the leadership in projects. So, we never sat down and said, hey, we want to make N E R a focus. But somebody in our community did and they not only ideated, but then went ahead to organize regular meetings and recruit people and come and give updates after our weekly meeting and that turned into a paper that was accepted to E M and L P. So, distributed leadership, I would like to see this continue. Something we are a little worried about is the fact that

• 24:07Speaker 00

  having an entity, well, it's a great thing first of all, that having an entity means that we can receive funding. And we have been able to receive funding in the past through organizations that we have collaborated with. But having our own registered entity presents an opportunity, yes, to fund our research. So far, it's been mostly run on volunteer efforts and the results have been great. And so our one concern is that now when there's funding, a decision has to be made, right? About what gets funded versus what doesn't get funded. And we're worried about how that's killed. Will it stifle the distributed leadership, especially on projects? Or will it far its fore?

• 24:50Speaker 00

  So, I'd like to see us maintain that despite funding and and hit us to turn into an organization where the characteristics change because. Now money is available. Something else I'd like to see us grow into is just further supporting pathways for individuals who are interested in productizing their work. So again, I'll go back to the fact that we have many young people who this may be their first interaction with any A I topic or their first interaction with N L P. And we pretty much have a great pathway for them to advance academic careers. Because they're taking part on projects, which means they're writing papers.

• 25:32Speaker 00

  Some of these papers are getting accepted at leading conferences. They're being mentored by leading researchers in the field. Many have gone on to get acceptances to masters and P H D programs, which are amazing. Internship opportunities. Basically, there's a really great pathway for people who want to advance research careers. But I'd like us to see us build a similar pathway that is just as strong for individuals who want to productize and build companies. So that's something I'm I'm hoping we can grow into. And then one of my absolute favorite things about Masakane is that we have a lot of female leadership. So as a woman in tech, this is something I look at.

• 26:14Speaker 00

  As someone who has organized communities for women, this is something that's close to my heart. And I absolutely love the fact that we have pretty good female representation. I attribute it to the fact that this is a movement that was started by women. And I think it's very powerful signaling. I'll I'll tell you, whenever I'm or I have been in a position to be assessing a company that I want to join, I'll probably go to their website and see if there's any women on the tech. And if there's none, I start to ask myself, do I want to be the first one? But if there is one, then that's one person I can write to and say, hey, can we have a chat about, you know, what's working for company X is like. So I think it's very powerful signal that this was started by women.

• 26:56Speaker 00

  And I love it that there is a a little sisterhood or perhaps a not so little sisterhood that is part of Masakane. And I would love to see that continue to be the case.

• 27:07Daniel Whitenack?

  Joyce, I'm curious if as a researcher in the university setting, working on related problems to what the Masakane community is doing, how has this sort of like groundswell of community building in Masakane influenced your research team and maybe the things you're able to do and the way you're able to engage? And then maybe you have a follow-up question for Kathleen with regard to some of the things on your mind as you think about your research group and engaging in the community building.

• 27:43Speaker 04

  Yeah, thanks. Just listening to Kathleen is really very enlightening to where the background.

• 27:49Daniel Whitenack?

  Super inspiring.

• 27:50Speaker 04

  Yeah, and the background and the roots of Masakane that I'd never knew about. So that's very interesting. And I think for us, how it's benefited the lab is unlike Masakane that is wide and does many things anyway, because it's a large community. We started out in the NLP space, you know, building out in speech and now in machine translation, like she mentioned, for the language that we care about, which is Luganda, the main language in Uganda. And Masakane has really come in to support the researchers in the lab. As you said, if there are people who need help, maybe they are running their models and they are stuck. I've seen many of the people in my research lab go to Masakane, put in a query and get responses. And this has enabled them to move much faster with their development,

• 28:32Speaker 04

  with their model development, unlike if it was just a closed community for us. Unlike if you went and wrote an email to one of the researchers that maybe developed a model or did something. But here it's on Slack, it's within the community and they get faster responses. Or maybe you find a situation where someone has encountered the same issue, maybe with the models, it's easier for them to even respond. And I like that it's open, right? They could just put in a question there and then you get an answer. So that sort of community building is something that's very critical that the lab is also learning to adopt and leverage on as well. But also listening to Kathleen speak right from the beginning, I think, Kathleen, you said community, our community building, community is very important to you. And I know that with the new role that you have with Mozilla

• 29:17Speaker 04

  that there's also a lot of community building involved in there. So can you maybe tell us more about the current work that you're doing around community building with Mozilla? And Kiswahili, in particular, your language that you're passionate about?

• 29:32Speaker 00

  Yeah, okay, so my current role is machine learning fellow at Mozilla and I'm working particularly on Common Voice and particularly in Kifahili. So I'm supporting work to build a Kifahili dataset on Common Voice. And this is starting literally from the collection of sentences. But I should step back and say that it's starting from community building again, like you highlight. So we want Kifahili speakers to care about this work. We are working to communicate to them that the existence of this dataset is something that is of of benefit to them because it's intended to be a digital public good.

• 30:14Speaker 00

  And we're working to build ties with organizations that are already working in Kifahili. So I realize that we have the tech capacity, but then we're walking into a space where people have been working for years to build language communities and I'd like for us to be sensitive to that and we'd like to be sensitive to that. So reaching out to the language boards, the universities that have linguistics, literature departments and such other communities, just to get input from them. And then this work begins at, of course, collecting text. So Common Voice is a project for building speech recognition.

• 30:54Speaker 00

  And we start with text because we can't have audios without, you know, their line of text. So building relationships with these organizations basically is also in our best interest because then we can find avenues to get text from them if they're existing or build programs that can enable us to create text in the course of our work. And then we've we've learnt so far from existing, from linguists especially, that there's a lot of diversity in Kifahili speakers. And this is something that, well, I innately know as a Kifahili speaker because listening to someone speak, there's a lot of nuance. I can potentially tell what part of the country they come from, if they're Kenyan, I can potentially tell.

• 31:37Speaker 00

  If they're not Kenyan or are Tanzanian or from or are from the D R C. And then there's an additional level of nuance that is apparent among people for whom Kifahili is their mother tongue. And this is a distinction that I'll say even I wasn't aware of because many of us in in Kenya and in Tanzania learn Kifahili. Because it's a national language in the country, but it is actually someone's mother tongue. And like many African languages, there are related dialects. So we're also learning a lot of the nuance, the fact that for people whom Kifahili is their mother tongue. If they listen to someone speak, they know, okay, you're originally African dialect X or you're from this particular part of the coast.

• 32:20Speaker 00

  And they sort of label all the rest of us as af-country Kifahili speakers, af-country implying away from the coast. The East African coast, which is the home of of Kifahili. So learning about all these nuances and recognizing that it would be great if we could capture this diversity in the data set. That's also been an interesting journey, working with the linguists. The way this work, well, at least the engagement with the linguists, that's currently shaping up as subsets of the data set. That's a representative of the various dialects and variants of the language. So there's also nuance, I'll say, depending on what other language, because to try and explain further. In Nairobi, for example, Nairobians rely heavily on English.

• 33:05Speaker 00

  And it becomes apparent in our Kifahili speech because we talk Kifahili. And it sounds like it's a translation as opposed to it sounding like speech that is naturally in Kifahili. Versus someone from the D R C, now I imagine they would mix the Kifahili with French and, you know, that as a level of distinction. And to people who are native Kifahili speakers, to them, we are all, you know, just af-country Kifahili speakers. So there's layers and it's super fascinating. But again, at this point, we're really thinking about it in terms of subsets of the entire data set. Which can be used to first fine tune to various contexts.

• 33:45Speaker 00

  So that if you're, if you're building an application for the coast, then we have data sets which make it easy for you to find to that context. Or if you're building for Nairobians, or if you're building for people in inland Tanzania, then there's data sets that you can use to fine tune. But then beyond that, to evaluate performance, because at the end of the day, the, I tend to think that Kiswahili could do to other African languages, what western languages have done to all African languages. Which is to be the focus of research and funding and development and at the expense of others, right?

• 34:24Speaker 00

  So now Kiswahili is being spoken in parts of southern Africa, and I think that's amazing that we're now pushing for, you know, potentially one language to be spoken across the African continent. But then I sort of worry that that may come at the expense of smaller languages or other languages in general, because now parents may start to think, hey, you should learn Kiswahili for, you know, upward or career mobility. But then that may make the mother tongue come second place or be absolutely forgotten. But going back again to the work with the linguists, we also want evaluation data sets, which can mean that we ensure that there's at least some minimum performance for all diverse speakers of Kiswahili.

• 35:06Speaker 00

  There's also a very strong gender thread in all our work. We realize that women tend to be, well, I'll start with the fact that speech recognition systems generally perform worse on women. And this is, this can probably be attributed to an imbalance in whatever the original data set is, right? So fewer women contribute to these data sets, and that could be for a multitude of factors. So we're being very intentional about creating spaces for women to contribute to the data set. Because we realize that they face maybe unique, but beyond that, we would like for them to also be part of developing voice distribution.

• 35:50Daniel Whitenack?

  Well Kathleen, I'm really excited to hear about the work that you're doing in terms of building these data sets and the way that you're thinking about having things like a strong gender threat in all your work and why that's important. I know myself, I've been challenged in this conversation to think about ways in which myself as a practitioner can be involved in sort of creating more diverse data sets, getting involved in these types of communities. I was wondering if you could maybe close us out by talking about how, from your perspective, if practitioners are listening to this podcast and they have a desire to maybe contribute

• 36:30Daniel Whitenack?

  to work related to this, either building sort of language diversity into text or speech or other data sets, or maybe there's like people in local language communities that are listening to this and wanting to get involved, wanting to build up data sets that could promote this sort of technology with their language. What are ways that these two groups can kind of get involved in this work? What recommendations would you make to them? How would you recommend that they sort of connect with people doing this work and get involved so that the community can grow?

• 37:05Speaker 00

  Thank you for this question. So first to the researchers who could potentially contribute to this work. I'll highlight the model I've seen with Masakani because I think that works great. It could be through mentorship. I think there's definitely ways to find groups of junior researchers who belong to these language communities or these underrepresented groups that could benefit from working or collaborating with you. So that's one thing through mentorship. The second is, in the event that researchers are working or are keen to work on the release of languages. It's possible that they'll have funding to create this data set.

• 37:45Speaker 00

  So I challenge them to, beyond looking for native speakers and having them create the data set which you then, you know, take away and go work in your lab in isolation. I challenge them to use that as an opportunity again to mentor. So it may be that you're not only looking for a native language speaker, but rather looking for a native language speaker who is also a junior researcher interested in this field. And so they can start out contributing by creating the data set. But then you can also then create avenues for them to contribute to what comes next in terms of the analysis and the actual work that goes after the data that has been created. And I think that's another model that would be great and, you know, give much more value addition.

• 38:29Speaker 00

  Then turning back to language communities, I'm going to, again at this point, highlight Common Voice. Because I think it's it's pretty amazing that Common Voice as a platform means that you as language contributor or a language community don't have to start thinking about this work from a standpoint of, okay, what tools do I need to collect the data, where am I gonna store the data, what's the infrastructure going to be like, how do I access the data and all this dynamic? I think it's it's as simple as if you have access to the internet and can access the Common Voice portal, then you can start creating, in this case, a speech recognition data set. And it's as simple as that.

• 39:11Speaker 00

  And if you look a little, I I'm willing to posit that there's opportunities for you to identify other such free resources or platforms where you can start creating data sets. Beyond that, I like to think of language communities, particularly language communities that are not originally Afrika. So look at Kithahili, for example, it's a huge language, there's loads of speakers, there's loads of interest from people off of the continent in having Kithahili resources exist. But I personally see it as an entry point to other local languages that are spoken in the places where Kithahili is spoken. Right? So if a speech data set exists for Kithahili, it's possible to then take that text and translate it into

• 39:55Speaker 00

  another local language that you speak, if you're a Kithahili speaker. And in that way, it could be a machine translation data set that comes out of it. It could go on to become a speech recognition data set. But then I'd also challenge these communities to then think about. Licensing, right? And licensing is a, is a pretty interesting topic in, in the field of NLP. Because on one end, we have the very well-resourced researchers who don't ask for permission and just scrape everything off of the web. And then there's the language communities and the junior researchers who work very hard to create this data set. And then I'm not necessarily the one who gets fast publication or most interesting publication because of constraints like skills or resources.

• 40:37Speaker 00

  So I would challenge them to think about licensing that can center their needs. And it, it may be that you say it's useful or it can be used for academic purposes, but non-commercial ones. Or it may be that you say you don't want the data used in any manner until individuals from your community can be the ones building the solution. Or it may be that you say, hey, you know, this is a language that we speak and there's a hundred thousand of us and nobody cared about this language before we started building the data set. So we actually reserve the right to control all the solutions that are built from this data set. Because at the end of the day, you are the one that will be directly affected by whatever those solutions are.

• 41:19Speaker 00

  So basically, I think as we bring more languages online, we should empower those communities to start thinking about how they can center their needs. Because they, they don't need to just create resources and make them available for the most resource or the most skills to swoop in and build tools which they then package and resell to the language communities that work hard to, to do it. In bringing these two together, I don't know, maybe the, the things I've highlighted already in my response, but I'd like to leave it there, I think.

• 41:54Daniel Whitenack?

  I definitely think that those are great jumping in points. For our listeners who are interested, we will definitely include links to the Masakana community and the Common Voice platform in our show notes. So please take a look at those and get involved and think about how you can start thinking about contributing in various ways or mentoring or whatever it might be. But uh, thank you so much for joining us, Kathleen. It's been a real pleasure to talk with you. I really appreciate you bringing your perspective and the hard work that you're putting in on these, on these problems. So thank you so much.

• 42:27Speaker 00

  Thank you for having me, Daniel, and everyone else at the Practical AI Podcast.

• 42:35Speaker 02

  All right, that's Practical AI for this week. Thanks for listening. If this is your first time with us, subscribe now at PracticalAI.fm, or simply search for Practical AI in your favorite podcast app. We're in there. And if you're a longtime listener, do us a solid by recommending the show to a friend. Word of mouth is still the number one way people find new podcasts they love. Special thanks to our partners for supporting our work, Fastly, Launch Darkly, and Linode. We appreciate it. And to the mysterious Breakmaster Cylinder for cranking out new beats for us all the time. That's all for now. We'll talk to you again next week.