Twenty minutes of speaking practice a week that a twenty-eight-person class cannot provide
The situation
In a three-hour class with twenty-eight students each person speaks for maybe four minutes, and the students who most need practice are the ones who will not volunteer. Advanced students have conversation partners through the tutoring centre; Level 2 students have nothing between Tuesday and Thursday. What I assess is whether a student can sustain a turn, recover from a breakdown, and notice their own repair.
Steps
-
Write scenario cards tied to what you taught this week
Google Docs and your syllabus
Build four to six card-sized scenarios — registering for classes, disputing a charge, calling in sick — each naming the role, the goal, and two target structures from the current unit. Open-ended “just chat with the AI” produces pleasant, low-demand talk that pushes nobody.
What you only learn by doing it: Put the communicative goal on the card, not the grammar point. “Get the fee waived” produces real negotiation; “practise the past perfect” produces a grammar drill in voice form. Target structures go on the back, for the debrief.
-
Give students a role-lock instruction to paste first
ChatGPT voice mode or Gemini Live, free tiers
Supply a standard opening: stay in character, do not switch into teacher mode, do not correct my English during the conversation, ask follow-up questions, do not accept my first answer. Without this the model breaks role within three turns and starts explaining grammar.
What you only learn by doing it: Add “if I say something unclear, ask me to clarify instead of guessing.” Models are trained to be maximally accommodating, which is exactly wrong for language practice — an interlocutor who always understands you removes the repair work that is the actual skill.
-
Ask for feedback only at the end, and only on two things
The same session
After the role-play ends the student asks for feedback restricted to two categories chosen in advance. End-loading preserves fluency practice; restricting prevents the wall of undifferentiated corrections students then ignore.
What you only learn by doing it: Students will ask “was my grammar correct?” and get thirty items. Require them to name their two categories before they start, on the card.
-
Require a transcript excerpt with one breakdown marked
Screenshot or paste into a shared doc
Students submit a short excerpt and mark one moment where communication broke down — misheard, could not find a word, had to rephrase. Marking the breakdown is the metacognitive move that converts private practice into something you can teach from.
What you only learn by doing it: This is the tip that matters most. When speech recognition mangles a student's accent — and it will — many conclude their pronunciation is bad. It often is not; the recogniser is. Teach them to distinguish “the system misheard me” from “I said it unclearly,” or this workflow can do real damage to confidence.
-
Debrief in class from the breakdowns, not the transcripts
Whiteboard and your scenario cards
Spend fifteen minutes collecting breakdown moments and working the repair strategies as a group — circumlocution, asking for repetition, self-correction. This is the part the chatbot cannot do and where transfer actually happens.
What you only learn by doing it: Do not read transcripts aloud or grade them. The moment students believe you are assessing the conversation itself, they script it in advance and the practice value evaporates.
Where this breaks down
Speech recognition accuracy degrades measurably on accented speech. Participants in the published study reported recognition failures around 10%, and the broader literature shows systematically higher error rates for non-native phonology — an equity problem falling hardest on exactly the students this is meant to serve.
The models are thin on pragmatics, idiom and culturally specific politeness, which is a large share of what an ESL student actually needs. They should not be the only speaking partner.
Free-tier voice caps and data costs will silently exclude students. Offer an alternative carrying equal credit — tutoring centre conversation hour, a partner in class — and never make the AI route the only route.
There is no evidence yet that gains in chatbot conversation transfer to unstructured real-world talk. The researchers say so directly.
Provenance: partly documented. Grab (2025), “Integrated AI chatbot practice: A pathway to improved ESL speaking skills,” Social Sciences & Humanities Open — a 12-week study with 120 university EFL students showing large effect sizes, though with paid apps and a different population. The role-lock instruction and marked-breakdown debrief are our adaptation for a free-tier community college context.