What should not be read aloud: the typographic noise a synthetic voice must ignore

An article is not only made of sentences. Around the text there is a layer of marks that serve the eye and mean nothing to the ear: the "Photo: Julie Martin / AFP" line under the image, the "Related reading" that points to another piece, the rule that separates two parts, the small "4 min read" at the top, and, more discreet still, the HTML entities that copy-paste leaves scattered in the source. On screen, all of this stands apart at a glance: a caption is small, a link is underlined, a rule is a line. Out loud, none of it is visible. A voice that blindly reads whatever it is handed will announce "photo credit colon" in the middle of a report, and the listener will wonder what they just heard.

The flaw only shows up when you listen

This is the kind of defect that survives every visual proofread. You reread your article, it is perfect. You do not think about the photo caption, because your brain filed it under "not the text" without even telling you. The voice engine has no such instinct. To it, "Share on Facebook" is a sentence like any other, and " " is a string of characters to pronounce. We measured what happens when unprepared text is fed through as is: a forgotten entity like   gets read as "n b s p", an … turns into gibberish, and a photo credit line barges in between two paragraphs of reporting. Each accident, on its own, lasts a second. Strung together across a whole site, they are the difference between audio people listen to and audio people switch off.

What serious preparation removes, and why it is anchored to the line start

The rule we apply is easy to state: strip what is written for the eye, keep what the writer meant to say. In practice, reading preparation removes credit lines ("Photo:", "Credit:", "Source:", "Caption:"), cross-references ("Related", "See also", "More on this"), calls to action ("Share", "Subscribe", "Newsletter"), the "Advertisement" label, the "Published on..." timestamp and the "X min read". The delicate part, the one that separates a useful rule from one that damages the text, is anchoring. "Photo:" at the start of a line is a credit and gets removed; "a photo: the one from 1998" in the middle of a sentence is not, and must stay. That is why our patterns are anchored to the start of the line and followed by a separator: a greedy rule eats prose, and that is exactly what we want to avoid. The approach is the same as charter mode, where every transformation is readable and tested one by one rather than handed to a model that might shift a word in passing.

HTML entities, the copy-paste trap

There is one special case worth naming, because it is invisible in the editor and very real in the source: HTML entities. Text imported from a CMS or pasted from a web page often arrives riddled with  , &, é, … and other codes the browser renders correctly but the file contains literally. Letting them through means reading code aloud. We decode them before synthesis: numeric entities are computed (é becomes "é", é too), which covers the whole accented alphabet without a table to maintain, and the common named ones are translated (  to a space, & to "and", … to an ellipsis). The listener never hears that there was any HTML in the text, and that is exactly right. The same logic applies to markdown markup: a subheading keeps its text but loses its hash, a bullet keeps its content but loses its dash.

A question to put to any voice tool

When you evaluate a reading solution, voice quality hits your ears in three seconds. Handling of typographic noise, by contrast, hides: you have to hand it a real article, with its photo caption, its "Related reading" and its entities, and listen all the way through. That is precisely what the French text to speech page works on: what happens between the raw text you publish and what comes out of the speaker. A tool that reads "photo credit" or "n b s p" has not done this work, and no voice, however beautiful, will make up for it. Silence about what should not be spoken is a form of quality, less spectacular than timbre, but it is the one that decides whether a listener stays.

Read next

• An audio player in the page or in an iframe: what that choice decides
• Image captions and photo credits: what French text to speech does with them
• Footnotes and note markers: what French text to speech does with them