2019:Languages/Wikispeech - making Wikipedia accessible through speech technology/notes

SESSION OVERVIEW

Title

Wikispeech - making Wikipedia accessible through speech technology

Day & time

Friday 16 August, 13:30

Session link

https://wikimania.wikimedia.org/wiki/2019:A11y/Wikispeech_-_making_Wikipedia_accessible_through_speech_technology

Speaker

Sebastian Berlin (WMSE), André Costa (WMSE)

Notetakers

Jon Harald Søby

Attendees

~ 40 (room capacity was 30 but tables were moved out to give more space)

SESSION SUMMARY

• Problems:
  • 15–30 % of people are auditory learners (= learn better by listening)
  • 14 % illiteracy rate worldwide
  • These people are not served by our projects at the moment. How can we solve this?

• Existing solutions are proprietary solutions, which are normally only available in commercially profitable languages, and can be quite expensive

• Wikispeech's purpose is to make Wikipedia accessible for anyone that faces reading difficulties

• Wanted a server-based solution, so not everything is done on your device. Also, while Wikispeech is centered on the Wiki* projects, the components should be reusable by outside users as well

• Modular, to allow as many languages as possible

• First project period was 2016–2018

• Currently supports English, Arabic and Swedish

• Developer version available at wikispeech.wmflabs.org

• Showing video presentation about how it works

• Second project period 2019–2021:
  • Finalize the Wikispeech reader, take it into beta

• New related project: Wikispeech Speech Data Collector

• Without speech data, doing text-to-speech for a language is impossible

• Toolkit to crowdsource collecting speech data
  • Toolkit will collect,validate, annotate, and create phonetically balanced manuscripts

• Future uses:
  • More voices, new supported languages
  • Use for Wikidata & Wiktionary

• Possible future uses:
  • Language preservation
  • Oral citations
  • Third party tools

• Q: It should be possible to choose the gender and dialect of the voice that speaks.
  • A: The possibility to choose different voices exists, but we need more free voices to use

• Q: Are you consulting experts to ensure the correctness of the pronunciations (e.g. for languages like Hebrew and Arabic)?
  • A: Are in touch with an expert from a technical university in Stockholm

• Q: In some programs for Arabic/Hebrew, they use probability instead of grammar rules. What is used for Wikispeech?
  • A: Use probability, but users can correct it with annotations where it is wrong