Case: Speechbert
Generating hours of audio within seconds for digital publishing giant ZEIT ONLINE.
ZEIT ONLINE is one of Germany's leading news websites, known for pushing the boundaries of digital publishing. In 2020, we teamed up with their editorial product development department to help design and build a range of novel audio experiences.
We asked ourselves: What if we could turn any article published on zeit.de into high-quality audio within seconds of the article going online?
Audio content is an effective and accessible way for media outlets to connect with their users. It allows users to listen at their own convenience and consume content without being tied to a screen. ZEIT ONLINE needed a fast and cost-efficient way to turn their vast portfolio of articles into convenient, high-quality audio experiences. This meant pivoting away from traditional human-read audio production. Together, we decided to explore text-to-speech technology and the role it could play in automating the article-to-audio conversion process.
Google WaveNet is a machine learning-powered system transforming text into natural-sounding speech in 220+ voices across 40+ languages. We chose Google WaveNet to power Speechbert and started with a simple, headless text-to-speech pipeline. We then tailored it to ZEIT ONLINE’s needs and integrated into their existing structures. The industry-leading technology allows Speechbert to vocalise, process and deliver hours of audio – all without the ZEIT team needing to lift a finger or record their voice to make it happen.
With the integrated audio player widget, users can listen to the Speechbert generated audio right on the article page. We supported the ZEIT ONLINE tech team with Speechbert improvements, one of them being audio quality. While Google WaveNet by default generates heavily compressed audio, Speechbert delivers high-quality MP3 files that ZEIT ONLINE users love listening to.
The first version of Speechbert was an invisible service running in the background without much user interaction. We designed the new Speechbert admin interface to be a reliable, easy-to-use tool for ZEIT ONLINE’s PO editors and any non-technical teams. Currently, it centres around three main areas of functionality: audio, replacements and feeds.
Even Google doesn't get every word right. Users love Speechbert’s audio because it addresses a significant text-to-speech challenge: pronunciation. Just like typos and spelling errors erode trust in a printed newspaper, wrong pronunciation will turn listeners away. The Speechbert replacement library allows ZEIT ONLINE to consistently improve and override WaveNet’s default pronunciation, targeting difficult words or characters and teaching it how vocalise them correctly.
Over the years Speechbert has become a large repository of all-things-audio at ZEIT. To make this content available to the outside world, we developed a dynamic playlist editor. The editor makes it possible for the ZEIT editorial team to create audio feeds that contain a mix of text-to-speech and regular, human-read audio like podcasts or advertisements. This variety is indexed via Speechbert, enabling a range of innovative storytelling formats.
Speechbert came to life thanks to the integrated team set-up and collaborative process. We were able to try out new technologies and tools – continually testing and iterating to create a system tailored to the ZEIT team’s needs.
ZEIT editors now use Speechbert daily, with over 250,000 text-to-speech conversions since launch. We’ve received heaps of positive feedback and feature requests from users and readers. In our ongoing collaboration with ZEIT ONLINE, we’re working to improve and develop Speechbert even further.
Team
- Jonathan Muth
- Mitch Ryan
- Tommy Svensson
- Christina Breuer
- Maja Persson
- Swetlana Dmitriev
- Harry Keller
- Sophie Nickolai
- Ida Dimic
Deliverables
- Strategic vision
- Product canvas
- Web application
- Data + activity pipelines
“As a non-sighted person, I have been using computers with synthetic speech output since 1987. Considering my experience in this area, I can only compliment you: the quality of your speech output is very good.”
ZEIT ONLINE reader