Voice mode

Voice mode: your speaking coach and TCF examiner

You will practise speaking out loud with Claude's voice mode in the Claude app. One Claude Project holds all the rules, so each day you only type one line and start talking.

Four modes. Claude switches between them depending on the line you type:

• COACH (weeks 1–3): slow, simple French, explanations in English, gentle corrections, repeat-after-me, answers built step by step.
• DRILL: listening drills VL-01 to VL-04, each item played once, as in the exam.
• EXAMINER: a strict TCF Canada speaking test, in French only, run the way FEI's official demo describes it.
• TEF-EXAMINER: TEF Canada's rules, only if you also book TEF.

After every session Claude gives FEEDBACK: a level for each task, your main errors, better phrasings and one thing to drill tomorrow.

1. Set up once (Day 1, about 15 minutes)

1. Create the Project. In the Claude app or on claude.ai, create a new Project and name it TCF Examiner.
2. Paste the instructions. Open project-instructions.md, select everything from the first line to the last, copy it, and paste it into the Project's instructions box. Save. The item bank at the end of the file is part of the rules, so paste the whole thing.
3. Optional extra files. You can add templates/speaking-tcf.md (theme cards, question lists, idea banks) and, if you book TEF, templates/tef-hedge.md (the worked TEF ads) to the Project's files. Claude then has more material to draw on. The Project works without them.
4. First-day check (5 minutes, on your phone). Open the Project, start a new chat and type Jour 1 | COACH | S1-01. Switch on voice mode and say « Bonjour, je m'appelle… ». Then check three things:
   • Claude behaves like the coach. It speaks slowly and simply, explains in English and asks about you. If it doesn't, voice mode may not be using the Project's instructions on your device. In that case, start an ordinary chat, paste the whole of project-instructions.md as your first message, and send the session line after it. Do this every day.
   • Your French is written down as French. Look at the transcript of what you said. If it comes out as English guesses, look in the app's settings for a voice or language option. If there is none, dictate your turns with your phone keyboard's microphone set to French, and send them as text.
   • Claude's French is spoken clearly. Claude can't slow its own voice, but it can use shorter sentences. Ask for that if the voice is hard to follow.
5. Get a timer and paper. Keep a phone timer running for every timed task. A second device is easiest (another phone, a watch, a tablet or a kitchen timer), because then you never leave the Claude app during a task. If you use the same phone that runs voice mode, check on Day 1 that voice mode keeps going while the timer counts down. Have paper and a pen for your T2 notes, as in the exam.

2. Every session

1. Open the Project and start a new chat. Never continue yesterday's chat. Each chat is a fresh day: Claude remembers nothing from earlier chats, and the day number decides which themes and cards come up.
2. Type the session line from the calendar exactly as given, and send it. For example:
   • Jour 12 | COACH | S1-04 S1-05 | S2-01 | VL-01
   • Jour 30 | EXAMINER | full mock
   • Jour 30 | EXAMINER | T3 S3-02 If yesterday's feedback gave you a drill, add it at the end: | focus: passé composé with être.
3. Read Claude's short plan. It lists what will run and which timers you need. In EXAMINER and DRILL it does not reveal the card, the question or the items.
4. Switch on voice mode and say « On commence ».
5. Run the timer yourself (section 3).
6. At the end, say "feedback" to hear a short summary. Then leave voice mode and type FEEDBACK in capital letters to get the full written feedback.
7. Copy the LOG line at the bottom of the feedback into your log (section 8).

Type the session line before you switch on voice mode. A spoken session line comes out garbled.

3. The timer

Voice mode cannot time anything, so you run the timer and you call time. When it rings, say « temps » (or « stop ») at once, even in the middle of a sentence. In the real exam the examiner stops you when time is up.

• If you run out of ideas near the end of T3, sum up your position again (« Pour résumer… ») until the timer rings. That's what you'd do in the exam.
• If Claude ends T3 before your timer, note how much time was left and add it when you type FEEDBACK (FEEDBACK, T3 ended with 0:40 left).
• If you forget the timer, Claude says « Pensez à votre minuteur. » once.

4. What you can say during a session

Things worth knowing about voice mode:

• A pause ends your turn. Voice mode sends what you said when you pause. If you stop to think mid-sentence, Claude answers only « Oui… ». Just carry on.
• Claude's words also appear on screen. In DRILL, keep the screen face down until you have answered, or you will see the answer options and the audio as text. In T3, don't read the question on screen: in the exam it is only spoken, and you may ask for a repeat. In T2, looking at the card on screen is fine: in the exam you hold the card.
• Headphones with a microphone and a quiet room help the speech recogniser.

5. Honest limits

• Timing is yours. Claude cannot see a clock, so it never knows how long you have spoken. It estimates only from how many words you said.
• Feedback comes from a transcript, not from your voice. Claude reads what the speech recogniser wrote down. That has three consequences:
  • Pronunciation feedback is approximate. Claude can only spot a sound when it turned your word into a different word (tout written for tu). Your r, your vowels, your rhythm and your accent never reach it.
  • The recogniser can hide your mistakes. It may write the correct word when you mispronounced it, or the right article when you mumbled. Some errors you made will not be in the feedback.
  • Your hesitations, euhs and long pauses may vanish from the transcript, so fluency ratings are rough too.
• Claude's voice is one synthetic voice, with no Québec accent. For real native speech, use the audio sources in the listening kit: RFI's Journal en français facile (RFI-01) and Radio-Canada (RC-01).
• The practice examiner follows the official rules, and real examiners are people. In T3, FEI's documents describe a monologue: the only interruption they mention is when time runs out. 5 of 9 prep sites say an examiner may cut in with a question. No source says how often. So the practice examiner prompts only if you stop early, and on days ending in 0 or 5 (about 1 day in 5) it adds two short prompts, one of them a mild challenge.
• Levels are estimates. FEI does not publish how the six ratings (two raters, three tasks) become a mark out of 20. Use Claude's levels to see trends and choose what to drill.
• Claude starts every chat from zero. The focus: field and your log carry your progress from one day to the next.
• Voice sessions count towards your Claude usage limits. If you hit a limit, switch to written practice and come back later.

6. Record yourself once a week

The transcript can't judge your accent, so once a week (the calendar shows when) you check it by ear.

1. Record a full mock on a second device while you do it in voice mode, or record any T1 and T3 on your own.
2. Listen to 60 seconds of your T3. Note the words you restarted, the sounds you weren't sure of, and any long silence.
3. Take 3–4 sentences of native audio with a transcript: RFI-01, or Radio-Canada (RC-01) for a Canadian accent. Record yourself reading those same sentences.
4. Play the native version and yours one after the other. Compare three things:
   • the vowels u / ou (tu / tout) and the nasals (vent / vont / vin);
   • endings that should be silent (petit, ils parlent) and the links between words (vous‿avez);
   • the rhythm: French stresses the last syllable of each group of words.
5. Shadow the clip three times: speak along with it, about half a second behind, as in LR-P2 of the listening kit. Then record yourself again.
6. Pick one sound to fix and type it into your next session line: | focus: the u sound in tu, vu, rue.

7. Reading your feedback

The three criteria. These are the ones FEI's raters use for each task:

Levels and marks. Claude gives each task a CEFR level (the European scale that runs from A1, beginner, to C2, near-native; your target is B2) and the mark band it matches:

• Your target is 10/20, and you train for 12. Re-marking is suspended for sittings from 1 September 2026, so a 9 cannot be appealed.
• All three tasks count. One B2 task next to two B1 tasks is probably not enough for 10 (an inference: FEI doesn't publish its rule). Aim for B2 on T2 and T3, and at least a solid B1 on T1.
• For a full mock, Claude also gives an overall band. It is an estimate, because FEI's rule for combining the task levels is not published.

The rest of the feedback, and what to do with it:

• Facts from the transcript: how many questions you asked in T2 (aim for 8–12 or more), how many T1 questions you answered, and whether T3 needed prompts or ended early. These numbers are the easiest way to see progress.
• Top 5 errors. Add each corrected sentence (not just the word) to Anki or your notebook. An error marked "possible recognition error" may be the transcript's fault. Check it in your weekly recording.
• Three better phrasings. Use them as models for your own sentences next time. Never learn them as lines to recite.
• One thing to drill tomorrow. Put it on tomorrow's session line as | focus: …. Claude starts the next COACH session with it.
• TEF feedback uses TEF's own grid, four rows per section, each rated in the lower or upper half of a level. TEF needs 456 for NCLC 7, which is the upper half of B2.

8. Logging your scores

Keep one running log in a single place: a note called "TCF speaking log" on your phone, a spreadsheet, or a file voice-mode/log.md that you create. After each session, paste the LOG line Claude gives you. For example:

LOG | Jour 21 | EXAMINER full mock (S2-01, S3-01) | T1 B1 | T2 B1 (7 questions) | T3 A2 (2 prompts, ended early: yes) | band 6–9 | drill: questions with combien and quand
LOG | Jour 22 | VL-03 L2 | 4/5 at first hearing (+0 after replay) | misses: old plan | drill: listen for finalement and en fait

These two lines only show the format; your own levels will be whatever Claude finds.

Once a week, read the week's lines and:

1. Find your lowest task and give it extra time next week, for example by switching one speaking session in the calendar to that task.
2. Track the facts: T2 questions asked, T3 prompts needed, and whether T3 lasted the full 4 min 30. They move before the levels do.
3. Look for repeated errors. An error that appears in three different feedbacks becomes a focus: drill for several days.
4. Copy your speaking band from each full mock into your full-mock record. The course compares all four skills at every full mock, starting on Day 21 (see START-HERE.md).

9. The one rule that matters most

FEI treats reproducing memorised texts as fraud, and examiners can test a recited answer with follow-up questions. So this system teaches frames (the order of the parts) and short phrases that you fill with your own life and ideas. If Claude writes a full model answer, it is labelled Study model: adapt, never recite. Read it, close it, and say your own version. If you notice you're saying the same paragraph every time, change its details and its order.

Where the rules come from: the examiner's behaviour follows FEI's official speaking demo and six FEI and CIEP candidate manuals (CIEP is FEI's former name), summarised in exam-analysis.md §2.4 and §5 item 12, and in research/gap-5.md. The French lines are this course's wording, not FEI's script. The TEF rules come from LFA's examiner booklet (research/tef-format-scoring.md). The listening kit (templates/listening-reading-tcf.md §6) also gives a stand-alone prompt for each VL drill, for use in an ordinary chat. Inside this Project you only need to type the ID.