Emoji read aloud: what a French text-to-speech engine should do with them

An emoji does not read like a word, yet it carries one. Behind 😀 is an official name, "grinning face with big eyes", that the operating system knows and that a screen reader announces on purpose for a blind user. But an article is not a screen reader, and an emoji dropped at the end of a heading or inside a quote is almost never meant to be spoken. The trouble for a speech engine is that it has to decide: read the name, skip the character, or leave a pause. All three happen, and two out of three are wrong in the context of an article meant to be listened to.

The three possible behaviours, and the one you hear

Faced with an emoji, an unprepared engine does one of three things, and you have heard all of them somewhere.

| What is written | Faulty behaviour | What you expect |
|---|---|---|
| Thanks everyone 🙏 | "Thanks everyone, folded hands" | "Thanks everyone" |
| That was great 😂😂😂 | "face with tears of joy" repeated three times | "That was great" |
| New 🚀 | "New, rocket" | "New" |
| Well done 👏 team | "Well done, clapping hands, team" | "Well done team" |
| Price: 15 € 🔥 | "fifteen euros, fire" | "fifteen euros" |
| ✅ Delivered | "check mark, delivered" | "Delivered" |

Reading the emoji name is the most common failure of generic voices, and the most disorienting to the ear: "folded hands" in the middle of a sentence of thanks breaks the meaning instead of supporting it. The second failure is quieter: some engines do drop the character but keep the pause it occupied, which produces an odd silence at the end of a sentence. The right behaviour, in an article, is almost always the third: strip the decorative emoji before synthesis, without leaving a gap. That is the underlying logic of the French text to speech page, where the essential part happens before any sound comes out: normalise the text, decide what gets read and what does not.

Why this is a rules problem, not a voice problem

You might think a good voice settles the question. It does not, because the choice does not depend on timbre but on a decision made upstream: is this emoji a decorative sign or a piece of information? The machine does not know, and the Unicode name is no help, since it always exists. An engine that predicts sound character by character sees a code point like any other, and either it has a name table and recites it, or it does not and stumbles on it. In both cases there is no understanding of the emoji's role in the sentence.

The real difficulty shows up in three families. Emoji fired off in a burst ("😂😂😂") should be treated as a single gesture, not read three times. Emoji joined by a zero-width character, like some flags or professions, are one visible symbol but several code points: a naive engine reads the pieces. And skin-tone or gender emoji add invisible modifiers that, handled badly, get spoken separately. None of these is fixed by a prettier voice. They are fixed by normalisation that recognises the emoji as such and decides its fate, exactly as for the other forms of typographic noise that should not be read aloud.

The case where the emoji carries meaning

There is an exception, and it has to be named so as not to mislead. Sometimes the emoji is the message. A review that shows only "⭐⭐⭐⭐⭐", an instruction where "✅" and "❌" tell allowed from forbidden, a list where each point starts with a pictogram standing in for a word: there, skipping the emoji erases information. These cases are rare in a written article, where the author writes their sentences, but they exist in short content and lists. So the right principle is not "never read an emoji", it is "do not read a decorative emoji, and when in doubt prefer the clean omission over the chatter". One "smiling face" too many is more jarring on listening than a silent smile.

What WeDispatch does about it

Decorative emoji, the ones that punctuate a heading, a sentence or a sign-off, are removed by the normalisation applied to every article, without leaving a silence in their place. A burst of identical emoji does not turn into a repetition, and composed emoji are not taken apart into separately spoken pieces. The principle is simple: in an article meant to be heard, the emoji is a visual sign, and rendering it literally in speech hurts the text more than it helps.

We prefer to put it this way rather than promise a perfect understanding of each pictogram's meaning. The machine does not understand your intent, it applies a sensible rule: the decorative disappears, the text stays. If a pictogram genuinely carries information in a piece of content, the good practice is to spell it out as well, for the eye and for the ear.

The two-minute test

Take a real heading or paragraph from your site that contains an emoji or two, not a demo text. Have the tool you are evaluating read it and listen to those passages only: do you hear "smiling face", an abnormal silence, or a sentence that flows without a snag? That is more telling than a whole-article listen. And if you want a full method for choosing between two voices, it is in how to evaluate a neural voice. A well-handled emoji is not heard, and that is exactly the point.

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