Sebenza means "work" in isiXhosa and isiZulu, two popular languages in South Africa. Our mission is to create 1 million jobs in Africa. There is 28% unemployment in South Africa right now. Most of those people have a smartphone and they all have free time.
There is no isiXhosa voice-to-text model but many voice-to-text use cases e.g. automated medical voice doctor that would greatly benefit from voice-to-text in isiXhosa. If you get a human listen to an hour of audio to transcribe it, you have to trust them with the potentially private content in the audio and it will take them an hour. Instead, you can "digitally shred" the audio into several 5-second pieces and play them for separate people so that no one person ever hears more than a 5 second slice of audio without any context. Doing it in parallel means you can do it much quicker.
Each worker does a test on signup and we calculate their "skill". We cannot trust any single worker to be correct so we play each slice to several workers and use their skill to calculate a "skill-weighted-consensus" in their transcription to get an accurate result.
This skill is stored on a smart contract as a sort of verified public CV (to demonstrate literacy etc.). We pay workers in Dai stablecoin to a wallet we set up for them on registration.
What it does
How we built it
Django, python, ruby, solidity
Challenges we ran into
Ran out of time. Getting dai on test net was difficult.
Accomplishments that we're proud of
It actually works!
What we learned
What's next for Sebenzai
We have final round ycombinator interviews in May. We are working on every type of data labeling for machine learning, not just audio. We are on a mission to create a million jobs in Africa.