Wikispeech

This company has R+D knowledge and skills relating to language technology that WMSE does not have in-house. We therefore partnered with them in 2015 and have since been developing Wikispeech jointly.

Researchers at the Division of Speech, Music and Hearing (TMH) provide cutting engineering research expertise to the development of Wikispeech. They have been working in the project since 2015.

In general, KTH Speech, Music and Hearing looks to further develop the core speech technologies to be able to deal with conversational, naturally occurring real world speech, and to be useful in real-world situations and in real applications.

As KTH is in charge of the National Knowledge Centre for Speech Analysis they have a vested interest in the project as an end-user of the data produced.

As a disability association for people with reading and writing difficulties/dyslexia and their relatives, parents and other interested parties the team has a deep understanding of issues and needs and a strong network in the area.

As an interest organization representing one of the stakeholders they will provide valuable feedback on the tool and will help engage new volunteers to contribute to Wikispeech.

The association has a technical advisory role in the project and has actively been helping Wikimedia Sverige's developer team with architecture descensions and navigating the deployment process.

Developing the Wikispeech extension and supporting tools like the speech collection application.

In particular speech recordings.

To ensure it provides value to users Wikispeech needs to be continually evaluated from a user perspective. With the results from the user testing looped into into the development process.

The built in feedback functionality of Beta Feature will be used to collect larger sale feedback.

The co-developing organisation, KTH, is developing user tests specifically aimed at investigating how well a listener internalises the information of a read text. The results of this work will be used to improve the language models and voices used by Wikispeech.

The mission of the Wikimedia movement is to make free knowledge available to everyone. A hurdle in reaching that mission is that not everyone can read, at all or well, or prefer to read in order to learn. Being able to listen to Wikipedia articles is one way to overcome that hurdle and Wikispeech provides precisely that service – to the WMF as the developer of the MediaWiki-software and to the Wikipedia communities. Close to 300 million people of all ages have a visual impairment to some degree, and almost 40 million are completely blind. Many millions of people have reading difficulties because of dyslexia or are illiterate. We estimate that in total upwards to 25% of all potential Wikipedia users would prefer being able to listen to articles over reading them.*

Wikispeech will be built into MediaWiki. That means that any reader who wishes to take part of the information can do it via Wikispeech; no extra equipment or particularly powerful devices will be needed. The Wikimedia movement strives for knowledge equity. With Wikispeech, people across the world can listen to the information on the Wikimedia platforms, on equal terms

Unlike many other assistive technologies or text to speech services Wikispeech is also open source software and built on exclusively openly licensed language resources. It is also ultimately a community sourced solution – readers, listeners!, of Wikipedia can themselves contribute corrections to how an article is read, thereby improving the overall quality of all read articles. The community will also be able to contribute speech recordings, submit them for use by Wikispeech, and so gradually and iteratively extend the language coverage of Wikispeech and improve the quality of its voice synthesis. As such Wikispeech is perfectly aligned with the values and practices of the Wikimedia movement.

* References:
Wikipedia contributors. (2020, September 19). List of countries by literacy rate. In Wikipedia, The Free Encyclopedia. Retrieved 14:56, October 6, 2020, from https://en.wikipedia.org/w/index.php?title=List_of_countries_by_literacy_rate&oldid=979289293

Pascolini D, Mariotti SPM. Global estimates of visual impairment: 2010. British Journal Ophthalmology Online. First published December 1, 2011 as 10.1136/bjophthalmol-2011-300539.

Max Roser and Esteban Ortiz-Ospina (2016) – "Literacy". Published online at OurWorldInData.org. Retrieved from: 'https://ourworldindata.org/literacy' [Online Resource]

Buşan AM. "Learning styles of medical students – implications in education". Curr Health Sci J. 2014;40(2):104-110. doi:10.12865/CHSJ.40.02.04

The language resources crowdsourced through Wikispeech will be possible to use to enrich other Wikimedia projects, such as Wikidata's lexemes or Wiktionary.

The technology developed for Wikispeech and the speech collection application can be reused for both language preservation efforts or as a key piece for collecting oral citations.

As they manage the Wikimedia platforms we have a software provider relationship to WMF. They control the software integration process.

Even though the WMF delivers Wikispeech to the community via the Wikimedia platforms Wikimedia Sverige expects to (at least initially) be the organisation responsible for receiving any feedback and implementing it. Wikimedia Sverige will also to develop educational material/instructions for how new languages can be added and would expect (at least initially) to be involved in the addition of any such languages.

There are a number of open source projects with overlapping goals and somewhat similar approaches. One example is Mozilla Common Voice, with whom we have had an ongoing exchange with during the development of Wikispeech.

Many researchers, companies and organizations have a vested interest in developing speech solutions but are in need of language data, e.g. language resource and technology research centers, Google and Mozilla. Wikispeech can become an important source of data.

Our key resource is staff and partners with the competencies and knowledge required to build TTS-software for MediaWiki.

In order to be able to create new voices (including for new languages) the software must have access to large amounts of language resources, primarily speech recordings.

To be able to improve the quality of Wikispeech in a direct, and very wiki inspired way, edits can be made to the underlying lexicon. Users can contribute with missing words or improve the phonetic transcriptions.

Operating under a Wikimedia brand gives us credibility with funders and partners.

The hardworking community of volunteers improves the data allowing the Wikispeech solution to become better and better.

On Swedish Wikipedia, where the extension is available, a logged in user will initially have the option of activating Wikispeech through the Beta Feature mechanism. New beta features are communicated through newsletters and users have frequently chosen to opted-in to any new features.

Once promoted out of Beta Feature Wikispeech should ideally be available to non-logged in users as well.

The MediaWiki site, https://www.mediawiki.org/wiki/Extension:Wikispeech, is where the extension is made available to developers and system administrators.

The MediaWiki site, https://meta.wikimedia.org/wiki/Wikispeech, is where Wikispeech is presented to the Wikimedia community.

The recorded speech data will be made available to Språkbanken Tal, https://sprakbanken.speech.kth.se/, the national speech data centre of Sweden. Språkbanken Tal in turn makes their resources available on the equivalent European platforms such as CLARIN and the European Language Grid.

We should also evaluate and publish the Wikispeech language resources on popular research data platforms such as Zenodo and Figshare.

There are a number of web sites where datasets suitable for use in ML/AI-applications can be published, eg. Kaggle. We should evaluate them and publish our language resources on at least one (and then evaluate them again based on what use/reach we see).