Advancing Dutch speech synthesis for digital news

How might we create clear, natural-sounding audio versions of news articles?

Inform

When Dutch text-to-speech falls short

Microsoft Azure could already convert written text into lifelike speech, but its support for Dutch grammar and pronunciation was still limited. Words could be mispronounced and complex sentences did not always sound natural. News added another layer of complexity. Numbers, punctuation and sentence structure all influence how a text should be read aloud. The newspaper needed a way to improve the existing technology for Dutch news without building an entirely new speech engine.

Analyse

From Dutch grammar to natural speech

We built an additional language layer on top of Microsoft Azure’s text-to-speech technology. A growing lexicon recorded Dutch words the system frequently mispronounced, while phonetic transcriptions specified how particular words and complex sentences should be spoken. Each correction could then be reused for future articles. We also pre-processed news content before sending it to Azure, improving how the system handled elements such as large numbers and punctuation. A Speechmaker dashboard allowed the team to test different voices, speaking speeds and pitch settings.

Activate

Giving the digital newsletter a voice

After testing and refining the pronunciation and structure, we built a working demo and integrated it into the newspaper’s digital newsletter. Readers could listen to news content directly in the newsletter. The result showed how an existing speech service could be adapted to Dutch journalism without building the underlying technology from scratch. It gave the newspaper a new audio format for sharing its digital content.