French text to speech: the silences that let a long article breathe
Inside a sentence, punctuation sets the rhythm. Between the paragraphs of a long text, what lets the reading breathe is not forced with artificial pauses: it is measured, and sometimes it is left alone.
Two questions often get mixed up, and they do not have the same answer. The first: inside a sentence, where does the voice mark its stops? A comma, a period, a semicolon and a colon do not carry the same length, and badly punctuated text gives itself away at once. We handled that question on its own, in punctuation and how TTS breathes. The second is quieter, and yet it decides whether a listener lasts twenty minutes: between two paragraphs, between two sections, what separates the ideas to the ear? That is the subject here.
Why the silence between blocks matters more than it seems
On the page, moving from one paragraph to the next is free. The eye sees the white space, skips the line, moves on. To the ear, that white space does not exist on its own: if nothing marks it, two distinct ideas run together, and the listener gets a continuous stream with no sense of when one part ends and the next begins. That is the first cause of drop-off on a long format, before the voice ever starts to feel monotonous. The listener does not think "this voice is tiring," they think "I have lost track of where I am," and they stop.
The temptation, once you notice this, is to force it: slip pauses in everywhere, stretch the silences, put clean hard stops between sections to be sure they are heard. That is exactly what you should not do, and we can say why with a number rather than a hunch.
What we measured, and what it taught us
A stance is only worth anything if it rests on a measurement. So we probed how our engine actually behaves on a plain colon, the sign that most clearly hinges two halves of a sentence. The result changed between two voice versions, and that is the whole lesson.
On an earlier version, the engine observed no stop at the colon at all: the measured silence was zero seconds, the voice ran on, the statement tumbled out. On the version we serve today, the same engine stops for about half a second (0.5 s) on its own, with nothing asked of it. The breathing we wanted to add, it already produces.
And when you try to force a pause anyway, with a technical tag of 300 milliseconds, the result is counterintuitive: the engine adds only six milliseconds of actual silence, but it shifts the entire preceding sentence by about six tenths of a second. In other words, the tag does not create a silence, it re-times the whole reading upstream. Nobody would have caught it except by listening, timestamp by timestamp. Forcing a pause, on an engine that already breathes, means damaging the rhythm you thought you were fixing.
Our rule: measure, then hold back when the engine is enough
From this comes a simple, slightly austere rule: we only place a pause where a probe has measured that one was missing, and only for the measured length. An engine we have not listened to gets no forced pause, because betting on an unmeasured silence risks it stretching across a sentence, or worse, ending up read out loud if the engine ever stopped interpreting it. The current engine, which already stops on its own, therefore gets nothing. Its natural prosody does the job better than an instruction placed by guesswork.
Two guardrails frame that choice, and both can be checked in the code. First, a pause is never applied to a colon stuck to what follows it: a web address (https://), a time (14:30), a ratio (1:2) keep their colon intact, so no silence lands in the middle of a token. Second, if you strip any tags from the prepared text, you get the original text back character for character. A pause therefore cannot deform what will be spoken: it is added, it rewrites nothing. It is the same promise as on the French text to speech pillar page: what is shown as "what will be read" is exactly what will be read.
What is on you, the writer
Since we do not force the breathing, it comes mostly from your text. Three moves serve it, and none of them is technical.
Short paragraphs first: a ten-line block on the page becomes a tunnel to the ear. Break where the idea changes, and the engine will mark the shift. Real subheadings next: a section announced by a heading is heard as a section, because the structure of the page carries into the reading. Finally, one idea per sentence in the dense passages: stacked asides, nested clauses, everything you read by glancing back, listens badly, because the ear does not glance back.
Past a certain length, no silence setting replaces splitting the text: a very long format is better served as episodes or chapters than as a single file. We cover that threshold and what it changes in listening to a long text without dropping off. The breathing of an article, in the end, is decided more in the writing than in the settings: the engine knows how to fall silent in the right place, as long as the text tells it where the right places are.
Give your articles a voice with WeDispatch
This blog is itself voiced by WeDispatch. Curious how it sounds on your content?
Book a demo