A folder full of speaking recordings does not automatically create a feedback loop. Listening once usually leaves a vague impression: "too slow", "a bit awkward", "not very fluent". The next recording often contains the same patterns.
Transcribe just 30 seconds word for word, and those patterns become visible. Missing verbs, repeated false starts and unnecessary introductions stay on the page long enough for you to choose one repair.
Self-transcription is not ordinary dictation
In dictation, you reproduce someone else's model. In self-transcription, you write what you actually said, revise it and perform the message again.
Tony Lynch used this process with English for Academic Purposes students at the University of Edinburgh. Learners transcribed sections of recorded role-plays, edited what they could, received a reformulated version from the teacher and repeated the task. His 2007 report on two classes suggested that the class involved in self-transcribing maintained more of the targeted accuracy gains in later performances. It was a small classroom study, so use it as support for a practical routine rather than a universal guarantee.
Lynch's distinction between two kinds of problem is particularly useful:
- Slips: You know the form and can notice the mistake on review, such as a missing third-person *s*.
- Errors: Your current knowledge does not give you a reliable correction. You need a model, explanation or informed feedback.
Not every problem means "I should concentrate harder." Some require new input.
The 30-second, four-step routine
1. Record a normal answer
Speak for 60–90 seconds about a real question: "Why did you change jobs?" or "What problem does this app solve?" Do not slow down artificially or stop to repair every sentence.
2. Transcribe one informative 30-second section
Choose a section with enough content to analyse, not necessarily the worst one. Write the words you actually produced, including repetitions and unfinished starts. A slash can mark a filled pause; no special notation system is needed.
For example:
I work here... I have worked here three years, and my job is mainly... mainly help customers solve the problem.
Do not silently improve the sentence while transcribing. The transcript needs to represent the attempt, not the answer you wish you had given.
3. Separate repairs from questions
Change one problem you can identify confidently. For instance, *I work here three years* may become *I have worked here for three years*. Put a question mark beside expressions you cannot verify and ask with the full sentence, not just an isolated word.
Avoid replacing clear language merely to sound advanced. *Help customers solve problems* is a perfectly useful outcome; a longer alternative is not automatically better.
4. Answer the same question without reading
Close the transcript and record a new answer using the same information. Change words and order if you like. The only requirement is to apply the repair you selected.
Without speaking again, transcription trains editing. Without a later task, you may only become fluent with one edited script.
Keep a three-line speaking log
One recording can contain ten possible corrections. Trying to fix all of them usually makes the next attempt worse. Choose the issue that affects understanding, recurs often, or is useful in an upcoming conversation.
Write three lines:
- Original: What did I actually say?
- Repair: What is a clearer version?
- Transfer: In which new situation will I use it again?
A week later, use the same structure in a new context. If it only works when you read the edited transcript, it is not yet reliable spoken language.
When transcription is not the right tool
If you do not understand the source material, work on comprehension before demanding a long response. If you cannot hear the difference between two sounds, a written transcript will not solve that perception problem. And a natural-sounding sentence is not necessarily factually or professionally correct.
Speech-to-text can save typing, but compare its output with the recording. A transcription error is not evidence of bad pronunciation, and accurate recognition is not evidence of natural rhythm or wording.
Use MirrorSay for the repairs you cannot make alone
Take one useful but uncertain sentence from your transcript. Say the intended meaning to MirrorSay, compare the natural version and model audio, then repeat it. Return to the original question and answer again: notice the gap, receive feedback, produce a new attempt.
To hear sentence-level timing more clearly, read English Thought Groups. For the feedback-and-repeat cycle itself, continue with Corrective Feedback for Speaking.
References
- Lynch, T. *Student-responsible correction of spoken English*. University of Edinburgh. Classroom practice PDF
- Lynch, T. (2007). Learning from the transcripts of an oral communication task. *ELT Journal*, 61(4), 311–320. DOI
*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.*