The Kinyarwanda DeepSpeech Hackathon was held in February saw individuals from the Rwanda technology ecosystem come together and ideate upon the Kinyarwanda DeepSpeech API made available by Digital Umuganda for Kinyarwanda Speech Recognition. Efforts to include Kinyarwanda on Common Voice began in 2020 and now the Kinyarwanda dataset is one of the largest available on the platform.
I view this work as a predecessor to ours in building for Kiswahili on Common Voice and there is a lot we can learn from their experience and hope to improve upon. Here’s three…
1. Government Support will Amplify Efforts
It was refreshing, and frankly shocking, to witness just how intertwined industry and government are in Rwanda and how much access these groups have to each other. We learnt that government support of the project led to obtaining access to over 3 million sentences in Kinyarwanda, and this has been key to the success of the project.
Additionally, we had a government representative attend the pitch session at the hackathon and participate as a judge. There was a lot of constructive feedback, suggestions to improve the applications proposed, potential partners the teams could consider pursuing along with offers to connect the teams to them, some of these partnerships being with government institutions.
They say that Kigali is small and everyone knows each other or knows someone who knows the person you need to be connected to. It is wonderful to see this spirit being put to use in support of developers and startups.
2. Community Diversity is Key
The launch of the Kinyarwanda Deep Speech API is a significant milestone. It is however not the end of the road. The Digital Umuganda team acknowledges that there was still a lot of work that needs to be done to improve the error rate of the published model. Part of this effort needs to be intentionality in diversifying the dataset in terms of contributors.
The 2000+ hours of Kinyarwanda data made available on the Common Voice platform has been contributed by a community of about 1000 people and this may be one of the reasons for the high error rate, that the model does not generalise as well as expected, given the amount of data made available.
I do wonder whether there would be a marked difference on performance between an individual who was a top contributor on the dataset and another who did not contribute at all. Speculatively, perhaps there is.
This project was kicked-off at the onset of the pandemic when physical meet-ups were impossible and subsequently, it was difficult to raise interest in the project among people who were yet to encounter it. We are working to gather contributions from a wider spectrum of people towards the Kiswahili dataset.
3. Start-ups can and should own the Innovation Opportunities
The story of Digital Umuganda is both a positive example and a cautionary tale. It is exemplary that a start-up has owned and driven the activities around data collection, model training and finally making available these APIs for general public use. They have also played a key role in the country in the time of the pandemic through Mbaza, a chatbot made accessible via USSD, that gives critical information related to the Covid-19 pandemic. They have generated revenue in the millions of dollars for the telcos hosting the service. However, most of this revenue goes directly to the telco…and here’s the cautionary part. This is a symptom of the support that startups need, beyond technical ability. The know-how to identify a market and figure out a revenue model that can sustain them and one day get them to profitability.
All in all, it was great to see a collaborative spirit among the participants, facilitators and the judges, of wanting to build with the teams and encouraging them to work on their ideas even beyond the hackathon. To learn more about the hackathon attendance and winners, check out this post.