Can French text-to-speech read poetry and plays? An honest answer

Sooner or later, someone running a literary magazine, a library website, a publisher's catalogue or a French class asks the same thing. A synthetic voice reads a news article well enough. Can it read Victor Hugo? Can it perform a scene from Racine?

The honest answer is: partly, and the parts it cannot do are different for poetry and for theatre. If you publish French literary text, this is what to expect before you press the button.

Why French verse is harder than English verse for a machine

English poetry is mostly about stress. Classical French poetry is about counting syllables, and the count changes how words are pronounced. That is the core difficulty.

The mute e. In everyday French, most unstressed e's are swallowed: "samedi" is two syllables in casual speech. In classical verse, an e followed by a consonant counts as a syllable and is voiced, at least lightly. An e before a vowel is elided; an e at the end of a line does not count. A trained reader applies these rules without thinking. A voice trained on everyday speech applies everyday habits. We wrote a full piece on the French mute e, and verse is its most demanding case.

Diaeresis. A poet can require two vowels normally said together to be split: "li-on" as two syllables, "pass-i-on" as three. Nothing in the spelling marks it. Only the line's syllable count tells you.

Line endings. In a poem, a line break is rhythm. When a sentence runs over into the next line, the reader chooses whether to mark the break or follow the syntax. That is interpretation, not rule.

What our tool does with a poem

We would rather state this precisely than let you discover it by ear.

WeDispatch's text preparation contains no verse-specific processing. It does not count syllables, apply diaeresis, handle the poetic mute e, or insert a pause at line endings. A poem is prepared exactly like prose: numbers, acronyms and abbreviations are turned into words, then the voice reads the text with its natural intonation.

In practice:

• syntax is respected: commas, full stops, questions and exclamations come through as in any text. For free verse or prose poetry, that is most of what matters, and the result is often perfectly listenable;
• metre is not: a twelve-syllable alexandrine can come out at eleven if a mute e is swallowed, and a diaeresis will be read as one syllable;
• a line ending is not punctuation: where it does not coincide with a comma or full stop, nothing guarantees you will hear it. What punctuation actually does to silence is covered in our piece on punctuation and pauses.

For a classical sonnet whose metre you want heard, a synthetic voice is not the right tool today. An actor is.

The fix that exists, and its side effect

The pronunciation lexicon lets you write a word the way it should be said, so you could force "li-on". But the lexicon applies to the whole account. Every "lion" in every article would then be split, including the sports report about Olympique Lyonnais. It is a reasonable move for an account dedicated to poetry, and a bad one on a general-interest site.

Plays: a voice per character, if the layout allows it

Theatre has a different problem: several people speak. One voice for every character makes a scene hard to follow, particularly for a blind or partially sighted listener who has no text to glance at.

WeDispatch can give a character its own voice. The account declares its characters (up to twelve) and the voice for each one. A line that starts with one of those names followed by a colon is then read in that character's voice. The wider mechanism, including quotes, is explained in our piece on reading with several voices.

The way the code recognises a line matters a great deal for classical drama, so here it is exactly:

• the name must be followed by a colon: "Hermione: Je ne t'ai point aimé, cruel ?". The convention of printed editions, the name in capitals alone on its line followed by a full stop ("HERMIONE."), is not recognised, and the whole scene stays in the main voice;
• the character's voice lasts to the end of the line, and no further. A verse speech written one line per verse will only have its first line in Hermione's voice; the rest falls back to the narrator. For a long speech, put the whole speech on one line, separating verses with a slash if you want them visible;
• the name itself is read aloud, in the main voice, as an attribution, so the listener knows who is speaking. Capitals and accents do not matter for matching;
• stage directions ("She leaves.") stay in the main voice, which is usually what you want.

Preparing a literary text for listening

1. Listen to one text before you voice fifty. The first listen tells you at once whether the voice suits the register. A calm, low voice serves an elegy better than an announcer's.
2. For drama, rewrite the speech layout: "Name: speech", one speech per line, and check that each character is declared with exactly the spelling used in the text.
3. Tell listeners what they are hearing. One introductory line ("read by a synthetic voice") stops a poetry lover from judging the work on a metre the tool does not claim to respect.

Our approach, set out on the French text-to-speech page, is to read French as it is written in articles, and to read it well. Metrical verse is outside that scope for now, and we would rather say so here than have you find out by listening.

Read next

• Security questionnaire for an audio tool: what to ask, and our answers, including the noes
• Breton, Basque, Corsican, Alsatian: why French text-to-speech trips on regional names
• A woman's or a man's voice for a quote: what French grammar reveals about who is speaking