Being able to speak with the answer in front of you, then going silent when it disappears, does not necessarily mean you cannot learn. Your practice may simply never have moved beyond full support.
Model fading divides a speaking task into a complete model, partial completion and independent expression. As help decreases, you take over content decisions, sentence formulation and delivery.
How is it different from shadowing alone?
Shadowing keeps you close to a model's language and rhythm. Fading asks you to leave that model later and complete a related task yourself. They can work in sequence, but the final imitation is not proof of independent control.
Goh's review describes scaffolding as help that enables learners to accomplish a speaking task they could not yet do well alone, moving towards performance without that help. A full model is one support, not the destination. Too much support can preserve dependence.
Three stages around one task
Imagine telling a neighbour that a parcel was delivered to the wrong address.
Stage 1: full model. Listen to and understand a complete response: you received someone else's parcel, give the address and ask for confirmation. Repeat it once and mark functional language such as I think and Could you; do not copy the whole script.
Stage 2: leave a skeleton. Keep I think... / The address... / Could you...? and three keywords. Insert your own address, parcel situation and request, then record 30 seconds.
Stage 3: independent variation. Hide all English. Change the problem to a missing delivery or a damaged parcel. Explain the issue, evidence and action you want.
If the last stage is impossible, return to stage two. The answer is appropriate support, not pressure to invent more advanced sentences from nothing.
Remove one kind of help at a time
You might hide the text but keep keywords, remove audio but keep a task picture, or remove sentence frames but retain a scenario card. Do not remove content, vocabulary, structure and sound support simultaneously.
Check four types of support:
- Content: do you have facts, an opinion or an experience to communicate?
- Structure: do you know whether to request first or explain first?
- Language: do you have essential frames and keywords?
- Sound: have you heard key expressions in natural speech?
Restore only what is missing. Copying the full answer again is rarely the only option.
A 15-minute session
Spend four minutes understanding and repeating a model; four minutes producing two versions from keywords; four minutes responding to a new scenario without text; and three minutes reviewing one selected expression.
Speak at every stage. “I could complete it in my head” does not tell you whether the support is fading appropriately. A recording does.
Independence does not mean perfection
Independent speakers can pause, check a word or ask for clarification. The goal is not to recite a stored model while changing your voice; it is to reorganise language in response to new information.
If stage two changes only a name, and stage three retains the same people, time and event, the task has not varied enough. Change a condition that affects a real language choice.
Where MirrorSay fits
At stage one, say your intended meaning to MirrorSay, compare a natural formulation and repeat it with the audio. At stage two, keep only your own keywords. At stage three, hide the reference and complete a new task.
For choosing among structures, continue with interleaved practice. To check independence after a delay, use task repetition. Dictogloss can add reconstruction practice, but you do not need every technique at once.
References
- Goh, C. C. M. (2017). Research into practice: Scaffolding learning processes to improve speaking performance. *Language Teaching*, 50(2), 247–260. DOI · Institutional manuscript
- Goh, C. C. M., & Burns, A. (2012). *Teaching Speaking: A Holistic Approach*. Cambridge University Press. Publisher book page
*MirrorSay is a free AI speaking-practice app: say a sentence, see a natural way to express it, then shadow the model. No sign-up, and speech is processed on-device whenever possible.*