Free French Speech-to-Text

French is official in 29 countries across five continents. The Organisation internationale de la Francophonie counted 321 million speakers in its 2022 report, of whom around 81 million speak it as a first language. Africa is where most of those 321 million now live: the same report puts 61.8 % of the world's francophones there rather than in Europe. French recordings therefore arrive from Paris newsrooms, Montreal call centres, Brussels committee rooms and Dakar radio studios alike, and no two of those sound the same.

French also puts a very particular problem in front of a speech-recognition engine: you cannot hear where one word stops and the next begins. Liaison pulls a silent final consonant across the gap and attaches it to the following vowel, so "c'est tout vert" and "c'est ouvert" are the same sequence of sounds, and "les os" and "les eaux" are indistinguishable. Where liaison is blocked — before an aspirated h — the effect reverses: "les héros" stays unlinked precisely so it does not turn into "les zéros". Meanwhile most final consonants are simply not pronounced at all. Petit, grand and parlez all end in a letter no microphone will ever pick up.

That makes French spelling largely a matter of grammar rather than hearing. The verb endings -er, -é and -ez are one single sound, so parler, parlé and parlez have to be told apart by what surrounds them. The plural ending -ent is silent, which means il parle and ils parlent are acoustically identical — the sample transcript below writes "qui les enveloppent" where Maupassant wrote "qui les enveloppe", and only the subject four words earlier decides which is right. Accents carry meaning that the ear cannot supply either: a and à, ou and où, sur and sûr differ in writing alone, and a cédille is what keeps français and garçon from being read with a hard c. The same transcript writes "ses voies" where the original has "ses voix" — a perfect homophone, resolvable only by context.

Numbers add one more layer, because above sixty French counts in twenties: soixante-quinze is literally "sixty-fifteen" and quatre-vingt-dix is "four-twenties-ten". Belgian and Swiss speakers say septante and nonante instead, which is why the same figure can arrive as two completely different words. Upload your audio or video and the engine writes out the spoken French as clean, searchable text with accents, elisions and punctuation in place. It runs in your browser, needs no installation, and your first minutes are free with no account required.

This is a real 45-second recording of a human reader, run through our AI engine exactly as you would run your own file. Nothing was corrected afterwards. Press play and read along.

Worth noticing: every elision and apostrophe lands correctly — s'en vont, l'existence, n'est, qu'un, l'imperceptible — and the full range of accents is there, from the é of sédentaires and étranges through the è of mystère and the ê of bêtes to the î of traînent. The literary inverted question Sont-ce is written as two hyphenated words rather than run together, and joncs keeps its two silent final letters. This passage is Maupassant, and word for word it is Maupassant's.

The slips are the interesting part, because all four are French-specific. The engine wrote ses voies where the original has ses voix — an exact homophone that only the sentence can settle. It wrote qui les enveloppent for qui les enveloppe, because the plural -ent is silent. It wrote robes de morte for robes de mortes, another inaudible plural s. And it ran feux follets together into feufollets. It also omits the space French typography puts before a question mark. This is careful, clearly articulated reading, close to a lecture or a dictated note; a noisy café interview with three people talking over each other will not come out this clean, and nothing automatic will. Recording quality remains the single biggest factor in the result.

Standard metropolitan French is what most broadcast, business and academic recordings use, and ConvertSpeech is tuned for it. One thing works in your favour everywhere: France, Belgium, Switzerland, Quebec and francophone Africa all share a single written standard, so a regional accent normally produces ordinary written French rather than a phonetic imitation of the accent.

Quebec French is the variety that diverges most in sound. Its clearest marker is affrication: before the vowels i and u, t and d become ts and dz, so tu is spoken roughly as "tsu" and dire as "dzire". High vowels also relax in closed syllables and long vowels tend to break into diphthongs, which together make the accent instantly recognisable. Vocabulary differs too — char for a car, magasiner for shopping, courriel for email, and a meal system in which déjeuner is breakfast, dîner is lunch and souper is dinner. Clearly spoken Quebec French transcribes well; broad, fast joual is difficult for any automatic system, and is worth a proofreading pass. If your recording is Canadian, our Canadian French page covers that variety directly.

Belgian and Swiss French are much closer to the metropolitan standard, and the differences that matter most are lexical. Both say septante for 70 and nonante for 90 instead of soixante-dix and quatre-vingt-dix. Switzerland goes further in the cantons of Vaud, Fribourg and Valais, where 80 is huitante — though Geneva, Neuchâtel and the Jura keep quatre-vingts, so the same country splits down the middle. Beyond numbers, expect Belgian GSM for a mobile phone, drache for a downpour and kot for a student room, and Swiss natel for a phone, cornet for a carrier bag and service! where France would say je vous en prie. These regionalisms are the words most worth scanning when you proofread.

African French now accounts for the largest share of speakers anywhere — the OIF's 2022 count places 61.8 % of the world's francophones in Africa. Senegalese, Ivorian, Cameroonian and Congolese French each have their own rhythm and their own borrowings, from Wolof, Lingala, Bambara and others, and urban speech such as Ivorian nouchi mixes French with local languages sentence by sentence. Formal and broadcast African French transcribes reliably; heavy code-switching between French and a local language mid-sentence is the hardest case, since only part of the audio is French at all.