Annual Report 2021/Story: Wikispeech is activated!
In this Story we are focusing on how we, after 5 years of development, finally were able to launch Wikispeech on Swedish Wikipedia. By doing so we have increased accessibility of the content for people who, for different reasons, have a hard time reading. It was not a straight path at all and the initiative was close to fall apart at the very end.
In 2015 we started conceptualizing a future software solution to make Wikipedia accessible to the millions upon millions of people for whom it is hard to access written text, such as people with visual impairment or dyslexia, or those who simply prefer listening to reading. But how could such a tool handle the quickly changing texts on Wikipedia covering all kinds of topics? How could we build a tool that was possible to improve by the volunteer community so that it would be really useful and possible to scale to different languages over time?
What we ended up with was Wikispeech, a text-to-speech solution built on a computer voice that is created with machine learning (when people talk about Artificial Intelligence (AI) they are often thinking of machine learning). However, back in 2015 the focus on machine learning was not as clear-cut a choice as it is today with the rapid improvements that we have seen.
After researching the topic for some time we realized that the international Wikimedia movement, jam-packed as it is with tech and linguistic savvy volunteers, would not only be able to help improve the Wikispeech solution, but that our movement could help solve a major problem with unethical AI – that is the lack of data available from people other than white English-speaking men which prevents AI solutions from working equally well for everybody. Collecting data at scale from a massive amount of people is hard but needs to be tackled.
So we started the development of Wikispeech with the clear intention to identify paths for the volunteer community to contribute with this data. The first step, which has now been achieved, was to develop a fully functional text-to-speech solution so that people can listen to Wikipedia. When they hear mistakes they will be inclined to provide solutions and by doing so helping to improve the available data. In June 2021 the solution was presented to the Swedish Wikipedia for early feedback and in the coming weeks the bugs identified were solved. On the 22 July 2021 Wikispeech was activated as a gadget on Swedish Wikipedia. Testing continued in 2021 to measure performance, solve reported bugs and add more requested features. In 2022 we will organize an official launch event and start scaling to more language versions of Wikipedia.
We are also continuing developing the tools needed to contribute with speech data so that AI over time will work for everybody, regardless of your gender, if you have a pronounced dialect or an accent, a speech impairment, the darkness of your voice etc.
In time we hope that Wikispeech will be added as an official extension of MediaWiki, as was our original intention. The code has already gone through a security review by Wikimedia Foundation so we are hopeful that this will be achieved in time. If you would like to test Wikispeech in Swedish you can visit sv.wikipedia.org, log into your Wikipedia account and in the user settings activate the gadget. Then you just pick a page you want to listen to and press play!