Correcting mistakes in spoken Chinese seems straightforward: a learner speaks, the teacher hears a wrong tone, a missing particle, or Russian word order and immediately supplies the right version. The logic appears sound. Yet the same mistake returns a few minutes later.
It is tempting to blame inattention. But the learner may have understood the correction perfectly well without turning it into a usable skill. During a conversation, they are deciding what to say, finding words, assembling a Chinese structure, pronouncing syllables, and watching the other person's response. Feedback adds another task: hold on to the original sentence, understand the correction, compare the two versions, and recover the thread. Precise feedback starts competing with the very performance it is meant to improve.
A correction is not useful merely because the learner heard the right answer. It is useful when they can recover the form themselves and use it in a different sentence.
Why repetition does not correct a speaking mistake
Suppose the teacher says: 我昨天买了这本书 (wǒ zuótiān mǎi le zhè běn shū — "I bought this book yesterday"). The learner repeats it accurately. At that point, most of the problem has already been solved for them: the words have been chosen, the order is fixed, and the position of 了 (le — perfective aspect marker in this example) is visible. All that remains is to hold a short sound sequence in memory.
There is no ready-made solution in the next utterance. The learner must decide whether a structure with 了 (le — perfective aspect marker in this example) fits, place it correctly, and keep the rest of the sentence intact. A flawless echo therefore tells us more about short-term recall than about a changed speaking skill.
Frequent prompting creates another trap. The learner becomes used to offering tentative fragments and waiting for confirmation. Their speech looks more accurate with the teacher present, but part of the control has effectively been outsourced. Remove the person who signals the error and the old form returns.
What we know: after a prompt, the learner can produce the correct version.
What we still do not know: whether they will notice the same problem without a prompt and self-correct before the listener reacts.
When to correct immediately and when to wait
An immediate interruption makes sense when the error changes the message or the exercise is explicitly focused on that form. Suppose the task is to practise the question 你明天有空吗？ (nǐ míngtiān yǒu kòng ma — "Are you free tomorrow?"). Omitting 吗 (ma — question particle) affects the target of the exercise. A brief correction does not derail the task; it brings the learner back to it.
Free narration is a different mode. Every interruption forces the learner to return to something they have already said and then reconstruct what was meant to come next. It is usually better to let them finish a complete thought, note one or two recurring problems, and discuss those afterwards. The conversation keeps its direction, and feedback does not become a constant background noise.
Sometimes the learner does not need the finished answer at all. The teacher can repeat the problem area with questioning intonation, gesture toward a position, or say, "Check whether the action is completed." A neutral cue leaves some of the search to the learner. It takes more effort, but it also reveals whether the correct form is available for independent retrieval.
Immediate correction, delayed feedback, and a cue without the answer solve different problems. Choose between them according to the cause of the error, not the teacher's habit.
Diagnosing mistakes in spoken Chinese
"Correct yourself" is only a meaningful instruction if the learner can hear the difference. If two tones still sound identical to them, self-correction becomes guesswork. They first need to compare contrasting recordings, slow down the model, and check perception. The article on segmenting the stream of spoken Chinese looks more closely at the gap between a familiar word and its sound form.
The pattern can be different: the rule is known but cannot be applied quickly enough in speech. In a quiet setting, the learner can explain the construction; in dialogue, they revert to familiar word order. This is no longer a gap in knowledge. The learner needs an easier speaking context and a brief cue that can gradually be withdrawn.
There is a useful intermediate test: can the learner recover the form after a neutral prompt? If so, outside help can be reduced over time. If not, it is better to return to a clear model and focused practice than to repeat the same request more loudly.
Finally, the correction must survive a change of example. After working through 我昨天买了这本书 (wǒ zuótiān mǎi le zhè běn shū — "I bought this book yesterday"), ask the learner to describe a different completed purchase with another object and time. Repeating the old sentence tests memory for that sentence. Producing a new one tests whether the learner is beginning to control the construction.
The broad verdict "the same mistake again" can therefore hide very different cases: the contrast is not perceived; the rule is not known; the rule is known but unavailable under pressure; or only the familiar example can be repaired. The same kind of correction will not work for all four.
A feedback protocol: one target at a time
Selective correction begins before the conversation. Choose one target for a short stretch of practice: a question particle, the position of a time expression, or one tonal contrast. Other mistakes are corrected only when the meaning breaks down. This is not permission to ignore accuracy. It is a way to give limited attention a clear job.
The learner then needs a real communicative task: arrange a time, retell an event, or compare two options. While they speak, the teacher does not repair every sentence in real time but notes a recurring problem. Once the learner completes the thought, the teacher brings back one short fragment rather than launching into a lecture on the whole grammar system.
Start with the smallest useful cue: "Check the question," "Where does the time go?" or "Compare the two tones." If the learner finds the solution, they say the whole sentence again rather than replacing a single syllable. The correct form is reconnected with the meaning and rhythm of the utterance.
Now change the conditions. Repeating 你明天有空吗？ (nǐ míngtiān yǒu kòng ma — "Are you free tomorrow?") is not enough. Ask about a different day or address a different person. The principle stays the same, but the memorized sequence no longer supplies the answer.
A few minutes later, remove the prompt altogether. Did the learner notice the deviation before anyone reacted? If so, control is beginning to move inward. If the mistake returns, go back to diagnosis: perhaps the form is not perceived, cannot be retrieved, or the task itself is still consuming all available attention.
The aim is not to correct as many mistakes as possible. It is to choose an amount the learner can notice, recover, and test in a new context.
Accuracy and fluency in Chinese speaking practice
It helps to separate work on accuracy from work on fluency explicitly. In accuracy mode, the target is known in advance, so brief interruptions are expected and do not break the learner's plan. In fluency mode, completing the thought matters more; feedback comes after a meaningful stretch of speech.
Accuracy mode: one narrow target, a quick cue, self-correction, and several new attempts.
Fluency mode: a communicative task, very few interruptions, and a short review of one recurring problem after the response.
Demanding both modes at once means asking someone to speak freely while consciously checking every element. The form is first stabilized in slow, controlled work, then moved into a short dialogue, and only then tested in freer speech. The article on a construction-based approach to Chinese grammar explores a related boundary between knowing a rule and having a usable construction.
How Tomyo can support the process
Tomyo does not diagnose speaking mistakes or replace a teacher or attentive conversation partner. It is useful at a different point in the process: when the learner needs a stable model. Practical situations include phrases, short dialogues, pinyin, translations, and audio, and the lines can be heard and repeated one at a time. This makes it easier to prepare the form before a conversation, when attention will already be occupied by meaning.
Cards with examples and audio, short study sessions, and spaced review maintain regular contact with the material. Recognizing a card, however, does not mean a word is available in a spontaneous utterance. It is preparation, not the final test.
Daily texts and adapted news provide fresh contexts: a familiar construction can be seen and heard, then used in an oral retelling. The app will not identify the cause of an error. The learner or teacher still has to choose the target and test transfer.
Where the protocol stops working
Selective correction does not mean postponing an error that changes the message. If the meeting time is wrong, the listener understands the opposite, or a request becomes unclear, stop and clarify immediately. Communication matters more than preserving a neat lesson protocol.
The approach also fails without a reliable model. If the learner cannot hear a tonal contrast, "watch your tone" adds nothing. They need separate perception work, comparison of recordings, and controlled production. An unfamiliar construction first needs an explanation and transparent examples. An automated error may have to be rebuilt in slower speech. The article on why Chinese tones do not automatically become natural speech after textbook practice examines that problem in more detail.
Not every pause is a reason to intervene. In natural speech, people search for wording, revise what they meant to say, and adjust their stance. If every uneven moment is treated as an error, learners quickly retreat to simple sentences because those feel safer.
Good feedback gradually makes itself unnecessary: the outside cue shrinks, while independent noticing, repair, and transfer remain.
The useful measure is not the number of corrections in a lesson or the perfection of a repeated sentence. Look at the sequence instead: the learner perceived the deviation, recovered the form, said the whole sentence, and applied the same principle in a new utterance. If one link is missing, more corrections usually add load rather than learning. The first task is to identify which action is not yet available.
