Building an audio experience for the Bulletin’s journalism

Like other media outlets, the Bulletin is grappling with the rapid technical advancement and mainstream adoption of generative artificial intelligence (AI) tools and their various implications for how journalism is consumed by online audiences. The Bulletin quickly coalesced around a core principle: Generative AI will never be used to directly create news content or to create media that could challenge the editorial meaning of news content and/or mislead audiences.

So, how might we responsibly use AI-based solutions that serve our audience meaningfully and expand the reach of the Bulletin’s journalism? We know that our readers have different habits, preferences, and preferred media formats. For example, younger audiences tend to prefer video or audio, and any age group can have situational limitations or permanent or temporary disabilities where digital accessibility makes a real difference.

These considerations and surveying AI use in news media led us to design and implement a “text-to-speech” experience for the Bulletin’s journalism, which is now live on the website. Text-to-speech is when a computer reads written words out loud to the user, like having an automated narrator for Bulletin articles. The implementation journey from conception to launch has had its share of lessons learned.

The audio player at the beginning of this article is an example of the player that will appear on Bulletin articles. Use the red play button on the far left of the audio player to play or pause the audio. The forward/backward icons allow you to skip 10 seconds forward or back. The audio can be muted or unmuted by clicking on the speaker icon. The volume can also be adjusted using the volume bar to the right of the speaker icon. The remaining duration of the audio is displayed along with a progress indicator in the center.

Editors and staff were involved at multiple stages of creating the player. For example, we polled the team on which automated voice to select for the experience to ensure it aligns with the Bulletin’s professional, authoritative tone. They also helped catch issues related to occasional word skipping or unexpected changes in voice or tone. This feedback was incorporated into the prompt instructions for the AI to the extent possible and helped socialize the idea that automated voice generation will not always be perfect–a known issue for the OpenAI model in use. Members of the staff across different departments also worked together on the visual design of the audio player.

On the technical side, our core philosophy was to avoid getting locked in with any specific AI vendor so that the Bulletin can maintain maximum flexibility and independence. As a small nonprofit media outlet, there are also resource constraints within which any technology work must take place. Both have been satisfied by the free ClassifAI plugin that the Bulletin’s website vendor, Fueled, built for WordPress websites, and which now serves as the foundational middle layer for the text-to-speech experience. It connects to OpenAI’s API for the audio conversion (currently) using the GPT-4o mini-TTS model. Obstacles we hit and resolved with our website vendor included splitting Bulletin articles into smaller segments to work with OpenAI’s 4,096 character limit for text conversion, normalizing article content (for example, handling of bylines and footnotes, or skipping image captions and social embeds), and dealing with the–sometimes very long–execution time that the real-time audio conversion takes. This last issue would have had a detrimental effect on staff workflows that we resolved by queueing longer articles for scheduled batch processing every two hours instead of real-time conversion.

In this first iteration, the audio player is only available on standard free web articles. In the future, we plan to expand its use to multimedia and premium articles, which present additional technical complexities to overcome. Feedback from the Bulletin’s readers on the automated audio experience is encouraged and can be sent to [email protected].