What AI takes over in learning French
An AI tutor is a program that practises with you in writing or by voice, points out mistakes and adjusts the difficulty. In French it mainly takes over the practice a teacher has no time for, meaning daily speaking and repetition.
| Area | What AI practice covers | Where a teacher is still needed |
|---|---|---|
| Pronunciation | scoring nasal vowels, u and ou one by one | intonation in free conversation |
| Grammar | counting errors with gender, agreement and tenses | explaining why a rule applies |
| Writing | correcting accents as in 'école' and spelling | structuring a dissertation or synthèse |
| Tone | suggesting standard email formulas | choosing 'tu' or 'vous' at the right moment |
French uses the Latin alphabet, so you need no separate script feedback as you would for Chinese or Arabic. You do need accent correction, because 'ou' (or) and 'où' (where) are different words.
Error statistics by category show after a few weeks whether your mistakes with gender or the passé composé are becoming rarer. For many learners this is the most useful measurement an AI course provides.
Speech recognition for French with 9,752 training hours
Speech recognition turns spoken sentences into text, and only then can an AI check vocabulary and grammar. How well this works for a language depends strongly on the amount of training audio. The 2022 study on OpenAI's speech recognition model Whisper gives figures for this.
| Language | Training audio in Whisper | Word error rate in the MLS test |
|---|---|---|
| French | 9,752 hours | 7.3 percent |
| German | 13,344 hours | 5.5 percent |
| Spanish | 11,100 hours | 4.2 percent |
| English | 438,218 hours | 6.2 percent |
The word error rate is the share of words the system gets wrong. The values apply to the largest model in the Multilingual LibriSpeech (MLS) test, which consists of audiobooks read aloud. Among languages other than English, only Chinese, German, Spanish and Russian had more training audio than French.
The figures come from native speakers reading aloud. The study gives no values for learners with an English accent, and many app providers do not say which speech recognition they use. Test the recognition during the trial with your own voice and microphone.
3 limits of AI with French pronunciation and spelling
1 Forms that sound the same
'parle', 'parles' and 'parlent' sound the same, and so do 'parlé', 'parler' and 'parlez'. Speech recognition writes the form that fits the sentence most probably. It therefore cannot check whether you meant the right ending, so practise endings in writing.
2 Liaison and tone
Whether a liaison is required, optional or wrong depends on word type and style. Language models, AI systems that generate text, also suggest suitable formulas but judge poorly whether 'tu' or 'vous' and which tone fit a situation.
3 Regional pronunciation
The measured values come from standard pronunciation. The Whisper study gives no separate figures for the French of Quebec, Belgium or West Africa. Many pronunciation scores also compare you with native speakers, while the Common European Framework of Reference (CEFR) assesses pronunciation mainly by intelligibility.
- Endings that sound alike stay invisible in speaking exercises.
- Grammar explanations from language models sound confident and are sometimes wrong.
- A regional accent can lower a score although your pronunciation has not got worse.
Protecting your data in an AI French course
The EU's General Data Protection Regulation (GDPR) governs personal data, and the UK has an almost identical UK GDPR. An AI French course processes voice recordings, texts, error profiles and study times. If a recording is used to identify you uniquely, it counts as biometric data under Article 9 of the GDPR. Pronunciation scoring does not need that identification.
Ask four questions before you subscribe. Where are recordings processed, in the EU or elsewhere? Does the provider use them to train its AI, and can you switch that off? Can you have recordings deleted (Article 17) and take your learning data with you (Article 20)?
For children, Article 8 of the GDPR requires parental consent under 16 when a service relies on consent, and EU countries may lower that age to 13. In the USA, the Children's Online Privacy Protection Act (COPPA) requires verifiable parental consent for children under 13.
The EU AI Act has banned AI that infers learners' emotions in educational institutions since February 2025. Since August 2026 an AI system must make clear that you are talking to an AI. From 2 December 2027, AI that evaluates learning outcomes in educational institutions must meet the duties for high-risk systems.
AI practice and a French teacher in one weekly plan
A hybrid course combines AI practice with live lessons. The AI handles practice and measurement, and the teacher handles conversation, explanation and planning.
| Component | Who handles it | Time per week |
|---|---|---|
| Vocabulary with articles, tenses, agreement | adaptive learning system | 5 x 20 minutes |
| Repeating nasal vowels and intonation | AI conversation partner | 3 x 15 minutes |
| Free conversation, tu and vous, error patterns | live teacher | 1 x 60 minutes |
| Reviewing error statistics, new goals | teacher using AI data | 1 x 10 minutes |
If you are working towards the DELF or DALF, you also need a teacher who marks your writing and speaking against the grid of France Éducation international. An app does not mark against that grid.
During the trial, say the same sentence once carefully and once sloppily. A useful system scores the two attempts clearly differently and names the sounds that got worse.
Checking an AI French course in 5 steps during the trial
- Say 'tu' and 'tout', then 'vin' and 'vent'. The system should mark the sound that deviates.
- Read 'les amis' once with and once without liaison, and see whether the score notices the difference.
- Type 'ecole' without an accent and check whether the correction suggests 'école'.
- Ask for an explanation of the passé composé and the imparfait and compare it with a grammar book.
- Read the privacy policy for storage location, storage period and AI training.
Also check which features the provider offers for French, because some features do not exist in every course language.
Frequently asked questions about AI online French courses
How reliably does AI recognise spoken French?
AI recognises French read aloud in standard pronunciation with few errors. OpenAI's Whisper model got 7.3 percent of French words wrong in a test with audiobooks, compared with 6.2 percent for English. The study gives no values for learners with an accent, so test the recognition with your own voice.
Can an AI score French nasal vowels?
An AI with pronunciation scoring rates French nasal vowels such as those in 'vin' and 'bon' one by one and plays the target pronunciation. Systems with sound-level scoring mark the sound that deviates. A teacher judges intonation and liaison in free conversation more reliably.
Which French mistakes does AI speech recognition miss?
AI speech recognition mainly misses wrong French endings that sound the same, such as 'parle' and 'parlent' or 'parlé' and 'parler'. It writes the form that most probably fits the sentence. Practise endings in writing with feedback instead.
Does an AI French course replace DELF preparation?
An AI French course does not replace DELF preparation, but it adds daily practice to it. Juries mark the exam against the grid of France Éducation international, and an app does not mark against that grid. For writing and speaking you need feedback from a teacher.
At what age can children in the EU use an AI French course without parental consent?
Children in the EU can use an AI French course that relies on user consent without parental approval from 16, unless their country has lowered the age to between 13 and 15 under Article 8 of the GDPR. In the USA, COPPA requires verifiable parental consent for children under 13. Check the provider's children's account before signing up.
How do you combine an AI French course with a teacher?
You combine an AI French course with a teacher by letting the AI drill vocabulary, grammar and pronunciation daily while the teacher speaks with you live once a week. The teacher reviews your error statistics and sets new goals. Together that comes to just under four hours a week.
Sources
The information is based on these sources. Check fees and dates with the provider before you register.
- Radford et al.: Robust Speech Recognition via Large-Scale Weak Supervision (Whisper), 2022
- General Data Protection Regulation (EU) 2016/679 on EUR-Lex
- European Commission: AI Act, overview and timeline
- France Éducation international: DELF and DALF
- Council of Europe: Common European Framework of Reference for Languages (CEFR)