This is an excerpt from a letter to a friend, who runs a non-profit that helps refugees in camps throughout Sub-Saharan Africa.

“I believe if a technology initiative was humanitarian-metrics first approach, we could both increase the quality of life on earth for the poorest of humanity, and could create a customer base of the future.

Here are some basics about how spoken language translation apps are made – and why there are so few for African languages.

Text and spoken word translation is done through machine learning – a technology that is a little like a magical black box –  even the greatest computer scientists aren’t 100% for how it works. It is like the machine can somehow think on its own, in a language humans can’t understand. More on that later.

To create a machine learning “model” for a language, one in which it can recognize spoken works in Kinyarwanda. For example, you need thousands and thousands and thousands of hours of audio in Kinyarwanda that have been transcribed into text and meticulously added to the “model”.

And then what happens from there is quite complicated and beyond the scope of this writing, which is about humanitarian ideas. 

For our purposes, what is important to know is that for each of the 27 languages we’ve picked, for about half, there are wonderful “NLP – natural language processing” developers all over the world who are working on this. There are initiatives, funding sources, grants, and a GitHub group called Masakhane, which has a 2,000 person Slack channel that is very active. 

For most machine learning language model makers, one of their biggest challenges is getting those thousands of hours of carefully transcribed audio in each language – especially rare, low-resource languages like Dinka. 

So this is my thinking at this point:

In the past (and present), the creation of translation apps was a for-profit endeavor, as much of technological advancement is for-profit, so they think about profits in every project they do. 

When ML developers look for those 10,000 hours of transcripted audio data, they are trying to find free sources, like audio books or newscasts. Again, they are focused on profit. 

My project – my hope – is that we will think instead about how to lift poor people out of poverty, and getting this audio and transcription – that is a byproduct. 

The vision is this: G*ce, who speaks Luganda, and lives in a refugee camp, could work for 10 hours a week, for $5/hr (USD) and uses her cell phone to record and transcribe about 5 hours of Luganda language every week. There are consistent pay increases. Underperforming workers receive remedial training and assigned different types of digital work. She uses a phone, works remote, and works on her own schedule. The recordings are fun to make and transcribe – it’s a creative act and helps build strong ML training lessons about the vocabulary and words used by poor people. The background noises of people in camps.

The actual recordings can be many things – people telling stories, having conversations, talking about their day, reading wikipedia entries, asylum testimonials. The more voices, the better. The majority of voices are women. They can also do WhatsApp voice notes. 

Once ***e gets good at this work (we’ll pay for her time training), and creates high quality data that is approved by the ML developers, then she can teach a friend in the camp how to do this work as well. (one month). We pay for the phone and the internet. It’s like a pyramid scheme where people actually benefit. We especially reach out to single mothers and LGBTQIA. 

Now in this vision 2 people are making $50 USD per week and the Luganda project developers are getting 10 hours of beautiful data each week, then 20 hours, then 50 hours a week – edited, cleaned, AI assisted. And then, one person at a time, we slowly scale. Until we get, say, 5,000 hours of transcribed audio. Hundreds of people, quietly pulled out of poverty — at least enough to be safe, fed, happier – more on that later. This is all something that can scale very quickly, with the addition of a $100,000 donation.

For concerns about data privacy, see my post:
Privacy and Data Collection: How to Not Be Total Bastards

Today, Google would say, “No! $50,000 for data in just 1 African language!” Why not pay them 25 cents an hour?”

But this is how it is a humanitarian-first mission: again, the KPI (key performance indicator) is the metrics of lifting people out of poverty. The metrics of hunger reduction, reduced stress, feeling of security and safety from harm. Additional funding supports the camp, and supports infrastructure projects, like building cell towers. “