# Practise a new language with AI: speak daily, get corrected, track progress

Set a level and a goal, give the assistant one standing instruction, talk for 15 minutes a day, review your repeated errors weekly, and check with a person once a month.

Workflows · v1.0 · draft · https://hayzlabs.com/workflows/practise-a-language-with-ai/

In short:
- Give the assistant your level, your goal and one standing instruction: one correction at a time, a hint first, and a running list of your errors.
- Talk for 15 minutes a day, put your three most repeated errors and new words into flashcards each week, and check with a person or a sample test each month.
- You need an assistant you can talk or type to, a notes file and a flashcard app. Voice mode helps but is optional.

**Use it when** you want daily speaking or writing practice in a language you are learning, with corrections, at a time that suits you.

**Not when** you need a translation you will rely on: see [Translating with AI](https://hayzlabs.com/guides/translating-with-ai/). To get a draft in a second language corrected, see [Writing in a second language, or with dyslexia](https://hayzlabs.com/guides/second-language-and-dyslexia/).

**You need** an assistant you can type or talk to, a notes file for your error log, a flashcard app with spaced review, and 15 minutes a day. The steps, times and wording are our suggestion. The general rules are in [Learning with AI without losing the learning](https://hayzlabs.com/guides/learning-without-losing-the-learning/).

## Steps

Set up once:

| # | Step | Who | Time | Check |
|---|---|---|---|---|
| 1 | Pick your level and one goal. Read the Council of Europe's [global scale](https://www.coe.int/en/web/common-european-framework-reference-languages/table-1-cefr-3.3-common-reference-levels-global-scale) for levels A1 to C2 and copy the line that sounds like you now. Write a goal you could test, such as "order food without English". | Person | 10 min | You have one copied line and one goal sentence. **If it fails:** choose the lower level. It is easy to move up. |
| 2 | Paste the standing instruction below into a new chat, or save it as a custom instruction, a setting the assistant reads in every chat ([Tell it who you are, once](https://hayzlabs.com/start/tell-it-who-you-are/)). | Person | 5 min | The assistant replies in the language at your level and asks a first question. |

Then repeat:

| # | Step | Who | Time | Check |
|---|---|---|---|---|
| 3 | Daily: a 15-minute conversation on one everyday topic. Use voice where you have it, or type. Try each correction again out loud. | Person and AI | 15 min | You said each correction again. **If it fails:** if it corrects everything, repeat "one correction at a time". |
| 4 | Daily: at the end, ask "update my error list". Paste it into your error log. | Person and AI | 2 min | Today's date has at least one row. |
| 5 | Weekly: read the log. Pick the three errors that came back most. Put them and the week's new words into your flashcard app as short sentences. | Person | 20 min | Three errors are marked as this week's focus. |
| 6 | Monthly: check with a person (a tutor, a class, an exchange partner) or an official exam board's sample test for your level. | Person | 45 min | Someone other than the assistant heard or read your language. **If it fails:** record a two-minute voice note to compare next month. |

> [!OUR RULE]
> Ask for a hint before the answer. When you make a mistake, the assistant should first point at it ("check the verb") and let you fix it. Only if you cannot should it say the correct sentence. One correction per turn, so the conversation keeps going.

The hint-first order comes from a meta-analysis of 15 classroom studies with 827 learners. Spoken correction had "significant and durable effects", and they "were larger for prompts than recasts" ([Lyster & Saito, 2010](https://doi.org/10.1017/S0272263109990520), meta-analysis). A prompt makes you fix the error; a recast just says the correct version.

## The standing instruction

```text title="Language partner instructions · v1.0"
You are my conversation partner for <TARGET_LANGUAGE>. My first language is <MY_LANGUAGE>.
My level: <LEVEL>. In my own words: "<LEVEL_LINE>".
My goal: <GOAL>.

Rules:
1. Speak only <TARGET_LANGUAGE>, at my level. Short sentences, everyday words.
2. Ask me one question at a time on <TOPIC>, and wait for my answer.
3. When I make a mistake, correct only the most important one per turn.
   First give a short hint so I can fix it myself. If I cannot, say the
   correct sentence and ask me to repeat it.
4. Keep a list of my errors. When I say "update my error list", show a table:
   my sentence, the correct sentence, the rule in one line, times seen.
5. Switch to <MY_LANGUAGE> only when I write "<SWITCH_WORD>".
6. If you are not sure a correction is right, say so.
```

Replace the following:
- `<TARGET_LANGUAGE>`: the language you are learning, such as `Spanish`.
- `<MY_LANGUAGE>`: your fluent language, such as `English`.
- `<LEVEL>`: your level from step 1, such as `B1`.
- `<LEVEL_LINE>`: the line you copied, such as `Can deal with most situations likely to arise whilst travelling in an area where the language is spoken`.
- `<GOAL>`: your goal from step 1.
- `<TOPIC>`: today's topic, such as `my weekend` or `booking a hotel room`.
- `<SWITCH_WORD>`: a word you will not say by accident, such as `HELP`.

## The error log

Keep one file. Add rows daily and mark the focus weekly.

```text title="Error log · v1.0"
date,my_sentence,correct_sentence,rule,times_seen,focus_week,in_deck
<DAY>,"<MY_SENTENCE>","<CORRECT_SENTENCE>","<RULE>",<TIMES>,<FOCUS>,<IN_DECK>
```

Replace the following:
- `<DAY>`: the date, such as `2026-10-12`.
- `<MY_SENTENCE>` and `<CORRECT_SENTENCE>`: what you said, and the fix.
- `<RULE>`: a few words, such as `past tense of ir`.
- `<TIMES>`: how often it has come up.
- `<FOCUS>` and `<IN_DECK>`: `Y` or blank.

> [!CHECK]
> Before a correction goes into your deck, check it in a dictionary, a grammar book or with a person. A wrong card gets learned as well as a right one.

Of ten common study techniques, self-testing and spacing practice out had the strongest evidence ([Dunlosky et al., 2013](https://www.psychologicalscience.org/news/releases/which-study-strategies-make-the-grade.html), review). That is step 5.

## Voice mode, as of 2026-10-10

| Assistant | Voice conversation | Plans | Where |
|---|---|---|---|
| ChatGPT | Yes | Every plan, including Free. Free has limited access and "Limits may change". | Phone apps and ChatGPT.com |
| Claude | Yes, in beta. Languages other than English are in beta too. | All plans. Counts toward your usage limits. | Phone apps, desktop app and web |
| Gemini | Gemini Live | Google's help page names no paid plan. | Gemini phone app on Android, iPhone or iPad. Not in the web app. |

Sources: [OpenAI](https://help.openai.com/en/articles/8400625-voice-mode-faq), [Anthropic](https://support.claude.com/en/articles/11101966-use-voice-mode), [Google](https://support.google.com/gemini/answer/15274899). Set the voice language in settings (in ChatGPT, Settings > Voice > Language): OpenAI says it "can help ChatGPT understand your speech more accurately". Plans change often.

**Privacy.** Your voice and your chats may be stored. OpenAI says audio from its newer voice conversations is stored with the chat and "Clips are retained for 30 days". Anthropic says Claude voice transcripts are saved to your chat history. Leave out names and addresses ([What not to paste](https://hayzlabs.com/start/what-not-to-paste/)) and check your settings ([Privacy settings pass](https://hayzlabs.com/setup/privacy-settings-pass/)).

## How to tell it worked

- Your error log shows the same errors less often from one month to the next.
- Words from your deck turn up in your talks unprompted.
- The person or sample test in step 6 agrees with the level you set, or you move up a line on the global scale.
- It has failed if the assistant corrects almost nothing while a person still finds many errors. Add what the person found to the log.

## What the research says

- Practising with conversation systems helps on average: a medium effect across 17 studies (d = 0.58), with support for systems that set a clear task and give corrective feedback ([Bibauw et al., 2022](https://www.lltjournal.org/item-detail/1035), meta-analysis).
- For speaking, 16 studies showed a medium effect (g = 0.61): large for vocabulary and fluency, medium for pronunciation, not significant for grammatical accuracy. Nearly all had fewer than 60 learners ([Hou & Min, 2026](https://doi.org/10.1017/S0958344025100268), meta-analysis). That is why the weekly review targets your grammar errors.
- Most of these systems predate today's assistants. A 2026 conference abstract pooling 67 studies found the technology type was not a significant moderator; recent chatbots powered by LLMs (large language models, the technology behind today's assistants) showed the largest estimate (g = 0.64, against 0.61 overall) ([Wang et al., 2026](https://hdl.handle.net/2078.5/278464), conference abstract, not peer-reviewed).

> [!LIMIT]
> Accuracy is uneven across languages. In Anthropic's own tests, Claude Haiku 4.5 scored 96.4% of its English result in Spanish, 78.3% in Swahili and 52.7% in Yoruba, on a translated knowledge test ([Anthropic](https://platform.claude.com/docs/en/build-with-claude/multilingual-support)). It is not a grammar test, but in a less common language expect more confident mistakes. Speech recognition is uneven too: the model card (published fact sheet) for OpenAI's Whisper, a speech recognition model, notes "disparate performance on different accents and dialects" ([model card](https://github.com/openai/whisper/blob/main/model-card.md)). A transcript that reads correctly does not prove your pronunciation was right, which is one reason step 6 involves a person. No study we found tests this routine as a whole.

## Sources

1. Council of Europe. CEFR global scale, Table 1 (CEFR 3.3). Official page. [Link](https://www.coe.int/en/web/common-european-framework-reference-languages/table-1-cefr-3.3-common-reference-levels-global-scale)
2. Lyster & Saito, 2010. Oral feedback in classroom SLA. Meta-analysis, 15 studies, 827 learners. [Link](https://doi.org/10.1017/S0272263109990520)
3. Bibauw, Van den Noortgate, François & Desmet, 2022. Dialogue systems for language learning. Meta-analysis, 17 studies. [Link](https://www.lltjournal.org/item-detail/1035)
4. Hou & Min, 2026. Dialogue-based systems for second language speaking. Meta-analysis, 16 studies. [Link](https://doi.org/10.1017/S0958344025100268)
5. Wang, Bibauw, Noreillie & Desmet, 2026. Meta-analysis of conversational chatbots on L2 learning. Conference abstract, EUROCALL 2026. [Link](https://hdl.handle.net/2078.5/278464)
6. Dunlosky et al., 2013. Review of ten study techniques. [Link](https://www.psychologicalscience.org/news/releases/which-study-strategies-make-the-grade.html)
7. OpenAI, Voice mode FAQ. Vendor help page, read 2026-10-10. [Link](https://help.openai.com/en/articles/8400625-voice-mode-faq)
8. Anthropic, Use voice mode. Vendor help page, read 2026-10-10. [Link](https://support.claude.com/en/articles/11101966-use-voice-mode)
9. Google, Talk naturally with Gemini Live. Vendor help page, read 2026-10-10. [Link](https://support.google.com/gemini/answer/15274899)
10. Anthropic, Multilingual support. Vendor documentation, read 2026-10-10. [Link](https://platform.claude.com/docs/en/build-with-claude/multilingual-support)
11. OpenAI, Whisper model card. Vendor documentation. [Link](https://github.com/openai/whisper/blob/main/model-card.md)