Flemish text-to-speech

by | Nov 10, 2020 | 0 comments

Application Track:

Ready Made

Code:

REACH-2020-READYMADE-VRT_1

Domain:

Proposed by:

De Vlaamse Radio- en Televisieomroeporganisatie nv (VRT)

Entity Logo:

Summary of the entity:

VRT is the public broadcaster of the Flemish Community in Belgium. Its mission is to inform, inspire and unite and so reinforce Flemish society. As a service providing organization, the VRT wants to take up a special position in society.

Summary of the challenge:

The objective is to improve algorithms for text to speech for Flemish.

Description:

VRT daily produces a large amount of content, ranging from radio content to news articles to video content.  Furthermore, as a public broadcaster, VRT tries to inform and inspire as many people as possible with that content. One way to achieve this is to reuse it and provide it in another format than it was originally made, e.g. by reusing the audio of video content as podcasts. Another interesting case we are looking into is automatically providing audio content based on written content, for example news articles. This way, we can provide written content to people that are visually challenged, to people that love to hear the news instead of reading it, etc.

In this context, VRT is very interested in a solidly working text to speech algorithm for Flemish. In our experience, the algorithms for text to speech in English do not perform well here, so their performance should be improved. Additionally, Flemish is slightly different from standard Dutch, hence standard text-to-speech libraries will probably need some extra training.

VRT offers videos (mp4 format) of their evening news bulletin with time coded closed captioning in Flemish as well as a collection of Dutch news articles.

Data:

Expected outcomes:

At the end of the project, VRT expects the following results:

  • The trained, working models (the model files)
  • Results of the evaluation of the models: how were they evaluated, using what part of the data, what metrics were used to evaluate the performance…
  • The code to train and evaluate the models.
  • A pipeline (or at least an idea) to perform text-to-speech news articles after they are written by our journalists.
  • Spoken version of the transcripts that we provided with the videos.
  • Spoken version of the Dutch news articles that we provided.
  • A way for journalists to adjust the created speech when it is not to their liking.

How do we apply?

Read the Guidelines for Applicants

Doubts or questions? Read more about REACH on the About Us page,

have a look at our FAQ section or drop us an email at opencall@reach-incubator.eu.