AI speaking homework works when it extends your lesson: students rehearse the same task with an AI partner, get instant feedback, try again, and you review the results before the next class.
In short AI speaking homework works when it extends the lesson you have just taught: students rehearse the same language with an AI partner, read the feedback, try again, and you review the results before the next class.
- Set one goal, two or three success criteria and a short task (about 10–15 minutes) built from your own materials.
- Ask for at least two attempts. The second attempt is where students use the feedback.
- Review in about 15 minutes: completion, the two or three most common issues, and a few responses you check yourself.
- Keep it honest with spoken, personal, class-linked prompts, a short reflection and a quick check in class.
- Treat AI scores as guidance, check age and data rules (UNESCO's 2023 guidance sets an age limit of 13 for AI tools in class), and expect the novelty to wear off.
Why is speaking homework hard to set without AI?
Speaking homework is hard to set without AI because speaking needs a listener who responds, and a teacher cannot listen to every student at home. Without a partner, a spoken task becomes reading aloud to an empty room; with voice notes, it becomes hours of listening for the teacher.
Speaking also needs more time than class can give. Paul Nation's four strands framework (2007) says a well-designed language course gives roughly equal time to meaning-focused input, meaning-focused output, language-focused learning and fluency development. Speaking sits in the output and fluency strands, which are hard to fit into a lesson: in whole-class talk, 25 students sharing 45 minutes get less than two minutes each, even if the teacher says nothing.
Anxiety adds to the problem. A study of 573 Indonesian vocational high school students (Bashori and colleagues, 2022) describes speaking as the most anxiety-provoking of the four skills, and the students interviewed felt less anxious speaking to speech-recognition websites than to peers.
AI speaking homework is a speaking task students complete outside class with an AI conversation partner or speech-recognition tool that replies and gives feedback on each attempt. A review of 25 chatbot studies (Huang, Hew and Fryer, 2022) found three practical strengths: timeliness, ease of use and personalisation. Automatic speech recognition (ASR), the technology that turns speech into text and compares it with what was expected, helps pronunciation too: a 2024 meta-analysis of 15 studies found a medium overall effect (g = 0.69), largest when feedback was explicit.
How do you set AI speaking homework, step by step?
You set AI speaking homework in eight steps: set the goal, choose the task type, build it from your own materials, set criteria, assign it, let students practise with instant feedback, review and grade, and use the results in class.
- Set the goal. Choose one thing from this week's lesson that students should be able to do, such as "ask for and give directions using two landmarks".
- Choose the task type. Match the task to the goal: a role-play for functional language, a timed retell for fluency, an opinion task for discussion, a sentence drill for pronunciation.
- Build it from your own materials. Start from the textbook page, text, slides or video used in class, so the homework recycles the same language. Write the scenario, the AI's role and three to six target phrases, then try the task once yourself.
- Set criteria. Write two or three success criteria you can hear in an answer, such as "explains the problem" or "uses three target phrases". Our ESL speaking assessment rubric has criteria to borrow.
- Assign it. Give a due date, a time guide (10–15 minutes), the number of attempts and whether it counts for completion or a grade.
- Students practise with instant feedback. Students do the task, read the feedback and do it again, fixing at least one thing.
- Review and grade. Use the short routine below rather than listening to everything, and adjust grades with your own criteria if the homework counts.
- Use the results in class. Open the next lesson with the two or three most common issues, then have students repeat the task with a partner.
AvoLingo, an AI language-learning service by AvoLabs, covers steps 3 to 7 in its teacher dashboard. Teachers turn their own materials (an uploaded PDF or image, a typed topic, a YouTube link, or whole books in Bulk Builder) into speaking and writing homework across 8 assignment types, and edit and preview each task; nothing reaches students until the teacher assigns it. Students get instant AI feedback and pronunciation scores on words and sentences. Teachers review afterwards, grade with their own criteria and weights, and follow progress in Homework by Date and Classroom Insights. Teachers can start free.
Why ask for a second attempt?
A second attempt lets students act on the feedback while the task is fresh. In the 4/3/2 technique, learners tell the same story three times to different listeners in four, three and then two minutes; Arevart and Nation (1991) found substantial gains in speaking speed and fewer hesitations. In classroom research, feedback that pushes learners to correct themselves had larger effects than simply hearing the correct form (Lyster and Saito, 2010).
A 15-minute review routine
- Check completion and count it as participation.
- Scan the class's feedback for the two or three most common issues.
- Listen to or read two or three responses yourself, including one you are unsure about.
- If the homework counts, adjust grades with your own criteria.
- Plan a five-minute follow-up for the next lesson.
12 speaking homework ideas for ESL classes
These 12 speaking homework ideas for ESL and other language classes each extend a lesson rather than replace it. Levels use the CEFR scale (A1 for beginners up to C2) and are suggestions only.
| Idea | How it works | Suggested level | Main skill |
|---|---|---|---|
| Role-play rehearsal | Repeat the lesson's role-play (café, hotel check-in, job interview) with the AI as the other person, twice | A2–B2 | Interaction |
| 4/3/2 retell, adapted | Tell the same story three times with less time each round, reading the feedback in between | A2–C1 | Fluency |
| Retell the reading | Summarise the class text aloud, then answer the AI's follow-up questions | A2–B2 | Reading into speaking |
| Picture talk | Describe a picture from the unit while the AI asks about details | A1–B1 | Description |
| Opinion with reasons | Give a view with two reasons; the AI politely disagrees | B1–C1 | Discussion |
| Interview a character | The AI plays a character from the class book; students ask five questions | A2–B2 | Asking questions |
| Target-word challenge | Use five words from the lesson naturally in conversation | A2–B2 | Vocabulary in use |
| Pronunciation drill | Say the key sentences, check the scores, repeat the weakest two | A1–B2 | Pronunciation |
| Presentation rehearsal | Rehearse a two-minute talk, then answer the AI's audience questions | B1–C1 | Extended speaking |
| Exam-format practice | Rehearse one exam task format, such as a long turn, against your class criteria | B1–C1 | Exam task skills |
| Teach the AI | Explain today's grammar point to the AI as if teaching it | B1–C1 | Accuracy |
| One-minute reflection | Say what you learned this week and one thing to improve | A2–C1 | Reflection |
Mix fluency tasks (retells, reflections) with accuracy tasks (drills, target words) across a term, so homework keeps Nation's balance. Exam-format practice is rehearsal, not a score prediction. In AvoLingo, several ideas match Conversation modes such as Image Talk, Debate and Teach the AI, and the pronunciation drill matches Sentence Learning, which scores each sentence for accuracy, fluency and completeness.
A task card template you can copy
A task card is a short brief that tells students, and the AI, what the speaking task is, what a good answer includes and what to bring back to class. Copy the template and replace the example, a B1 shopping unit, with your own lesson.
| Field | What to write | Example |
|---|---|---|
| Lesson goal | What students can do afterwards | Return a faulty item and ask for a refund or exchange |
| Scenario and AI role | The situation, and who the AI plays | Your new headphones stopped working; the AI is the shop assistant |
| Target language | Three to six phrases from the lesson | I'd like to return… / It stopped working after… / Could I exchange it? |
| Success criteria | Two or three things you can hear | Explains the problem; asks politely for a solution; uses three target phrases |
| Level | CEFR band | B1 |
| Time | Total, including both attempts | 10–15 minutes |
| Attempts | How many, and what changes | Two: read the feedback, then fix one thing |
| Reflection | One sentence at the end | "The AI corrected… so the second time I…" |
| In-class follow-up | What students bring back | One correction and one new phrase |
If your tool lets you write instructions for the AI, paste the scenario, role and target language straight in. For role-plays built around a profession, see our guide to ESP course design.
How do you keep AI speaking homework honest?
You keep AI speaking homework honest by setting tasks an AI cannot simply do for the student: answers spoken in the student's own voice, about their own life or this week's lesson, followed by a reflection and a short check in class. Speaking is harder to outsource than writing, but a student can still ask a chatbot for a script and read it aloud.
- Require spoken production. Ask for voice rather than typed answers where your tool allows it, and set a time limit so reading a long script is impractical.
- Make prompts personal and local. "Describe your route to school" or "use three phrases from Tuesday's lesson" is harder to hand to a chatbot than "talk about travel".
- Use unscripted follow-ups. When the AI asks questions, disagrees or plays a character, students have to respond, not recite.
- Ask for a reflection. End with one thing the AI corrected and what the student changed on the second attempt.
- Check in class. Ask students to tell a partner one correction and one new phrase, or to repeat part of the task live.
- Lower the stakes. Count completion on most practice tasks and grade fewer, clearly assessed ones, so there is less reason to cut corners.
What about student data and age limits?
Before setting AI speaking homework, check your school's policy and local law, because speaking tasks can record students' voices and store their answers. UNESCO's Guidance for generative AI in education and research (2023) sets an age limit of 13 for using AI tools in the classroom, according to UNESCO's announcement of 7 September 2023.
National rules differ, so check local rules for your country and type of school. Three examples:
- United States: the COPPA Rule sets requirements for online services that collect personal information from children under 13.
- European Union: GDPR Article 8 lets children consent to online services themselves from 16, and member states may lower that to 13. Schools may use a different legal basis, so ask your data protection officer.
- South Korea: Article 22-2 of the Personal Information Protection Act (in Korean, added in 2023) requires a legal guardian's consent when consent is needed to process the personal information of a child under 14.
When you compare AI tools for language teachers, ask each provider:
- What is recorded (audio, transcripts, scores), where is it stored and for how long?
- Who can see it: the teacher, the school, parents?
- Is student data used to train AI models?
- Can a student's data be deleted on request?
The US Department of Education's 2023 report on AI in teaching and learning makes "humans in the loop" its first recommendation: teachers keep the final say on what AI output means. If your school needs to review how AvoLingo handles student data, contact us.
What are the limits of AI speaking practice?
AI speaking practice for students has clear limits: speech recognition makes mistakes, works better for some voices than others, helps individual sounds more than intonation, and loses its novelty. Plan for these from the start.
- The novelty effect. The 2022 chatbot review lists the novelty effect among its main challenges: interest can fall once a tool is familiar. Vary task types, keep tasks short and tie each one to the lesson.
- Recognition errors. ASR can give wrong feedback; Neri, Cucchiarini and Strik (2008) warned this could lead to "learning breakdowns". Treat scores as guidance and check a sample yourself.
- Uneven accuracy across speakers. A 2020 PNAS study of five commercial speech-recognition systems found an average word error rate of 0.35 for Black speakers in the US against 0.19 for white speakers. It did not test language learners, but it is a good reason not to grade on a recognition score alone.
- Sounds more than rhythm. The 2024 meta-analysis found a large effect on individual sounds but a small one on stress, rhythm and intonation, and short treatments did no better than practice without ASR. Plan pronunciation work over weeks and teach intonation in class.
- Practice alone is weaker. The same analysis found a large effect when students practised with peers alongside ASR and a small one when they practised alone, which is why step 8 brings the task back into pair work. Effects were clearest for adults and intermediate learners, so watch how younger students and beginners respond.
- Not a real conversation. An AI partner is rehearsal for talking with classmates, teachers and other people, not a replacement for it.
FAQ
How long should AI speaking homework take?
About 10–15 minutes, including a second attempt, fits most homework slots. This is a practical suggestion, not a research figure; short, regular tasks are easier to keep up than one long task.
Is AI pronunciation feedback accurate?
ASR-based feedback helps pronunciation, especially individual sounds, but it can misrecognise speech and is less reliable for stress, intonation and some accents. Use scores as guidance and listen to a sample yourself.
Can students cheat on AI speaking homework?
Yes, for example by reading aloud a script a chatbot wrote. Personal prompts, unscripted follow-up questions, a time limit and a short in-class check make it much harder.
Do I still need to give feedback myself?
Yes. The AI gives instant feedback on each attempt, but you decide what matters, adjust grades with your own criteria and deal with common issues in the next lesson.
Is AI speaking practice suitable for young learners?
Check your school's policy and local law first. UNESCO's 2023 guidance sets an age limit of 13 for AI tools in the classroom, and many countries require a parent's or guardian's consent for younger children's data.
Sources
- Nation, P. (2007). The four strands. Innovation in Language Learning and Teaching, 1(1), 2–13.
- Arevart, S. and Nation, P. (1991). Fluency improvement in a second language. RELC Journal, 22(1), 84–94.
- Lyster, R. and Saito, K. (2010). Oral feedback in classroom SLA: a meta-analysis. Studies in Second Language Acquisition, 32(2), 265–302.
- Huang, W., Hew, K. F. and Fryer, L. K. (2022). Chatbots for language learning: are they really useful? A systematic review of chatbot-supported language learning. Journal of Computer Assisted Learning, 38(1), 237–257.
- Ngo, T.-N., Chen, H.-J. and Lai, K.-W. (2024). The effectiveness of automatic speech recognition in ESL/EFL pronunciation: a meta-analysis. ReCALL, 36(1), 4–21.
- Neri, A., Cucchiarini, C. and Strik, H. (2008). The effectiveness of computer-based speech corrective feedback for improving segmental quality in L2 Dutch. ReCALL, 20(2), 225–243.
- Bashori, M., van Hout, R., Strik, H. and Cucchiarini, C. (2022). Web-based language learning and speaking anxiety. Computer Assisted Language Learning, 35(5–6), 1058–1089.
- Koenecke, A. and colleagues (2020). Racial disparities in automated speech recognition. PNAS, 117(14), 7684–7689.
- UNESCO (2023). Guidance for generative AI in education and research and press release, 7 September 2023.
- US Department of Education, Office of Educational Technology (2023). Artificial intelligence and the future of teaching and learning.
- US Federal Trade Commission. Children's Online Privacy Protection Rule (COPPA).
- European Union (2016). General Data Protection Regulation (EU) 2016/679, Article 8.
- Republic of Korea. Personal Information Protection Act, Article 22-2 (in Korean).
Written by the AvoLingo team. We used AI to help research and draft this article; sources are linked in the text.
Back to all articles