Pronunciation feedback: give a usable next step

Tell people whether to tap or hold

Easy Speak explicitly asks for a long press on the microphone. Tahanan says Tap the microphone to start recording and offers Can’t speak right now. These are different interaction contracts even though both use a large mic icon. People should not need to discover the difference by losing the first half of their answer.

Keep ready, recording, processing and feedback states distinguishable. A press-and-hold interaction needs a usable alternative for people who cannot sustain the gesture. If releasing submits audio, say so; if a second tap stops capture, keep that instruction available. Do not change the gesture after a retry without explaining the new state.

Handle microphone denial, a disconnected headset and silence before assigning a score. Preserve the target phrase and lesson position while asking for recovery. If speaking is optional, let the learner continue through a truthful alternative rather than recording a fabricated pass. Test with a person who is in public and cannot speak at that moment.

Keep the target visible during a retry

Lucida’s Almost state keeps Día, its translation and a phonetic guide beside the microphone. Talkpal’s Sentence Mode keeps the full phrase, Replay, Explain and Skip in one task. These screens help retain the object of practice rather than replacing it with a celebratory or discouraging result page.

For a short word, keep spelling, translation and a useful pronunciation aid together without treating them as interchangeable. A phonetic aid is a support, not necessarily a standardized transcription. For a sentence, let the learner replay the target without accidentally recording again. Explain buttons should not clear the existing attempt or change which phrase is being assessed.

If feedback highlights a word, maintain its position in the original sentence. A learner should be able to hear the phrase in context and then practice the difficult portion. Distinguish a deliberate new exercise from another attempt at the same one. Progress counters should not advance merely because the microphone button was pressed.

Give model and learner audio distinct identities

BoldVoice’s ready state names the word there and uses Tap to Speak. Its feedback state offers Coach and You playback beneath guidance to listen and compare. That pair is more actionable than an isolated gauge, because it gives the learner a concrete way to inspect the difference being discussed.

Label audio sources consistently and avoid simultaneous playback that makes comparison impossible. If model and learner samples have different lengths, keep controls explicit about which one is playing. Retain the correct attempt when the person records again; do not let an old playback request load beneath the newest result.

If you keep an attempt history, associate the displayed score, recording and feedback with the same attempt identity. Showing a newer score while playing an older take can teach the wrong correction. Make deletion and retention of voice recordings understandable. Product analytics can record interaction outcomes without collecting the full spoken content unnecessarily.

Point to something the learner can change

Fluently highlights a sound comparison using You said and instead of. Loora offers word-level feedback under a separate Pronunciation tab, with individual playback and practice controls. Both go beyond a whole-sentence score, but the useful question remains whether a learner can understand and act on the specific instruction.

Start with the highest-priority correction rather than opening every phoneme detail at once. Explain the target in terms appropriate to the audience and let deeper technical notation remain optional. Use text, audio and visual highlighting as complementary cues; red text alone cannot carry the feedback. A whole word and an individual sound need different accessible labels.

Treat uncertainty honestly. If the service cannot evaluate an attempt because it was too quiet or too short, present that as a capture or assessment limitation. Do not turn missing evidence into a confident pronunciation judgment. Avoid implying that matching one reference accent is a complete measure of communicative competence.

Retry without trapping the learner

Speakly displays Room for improvement with two tries left and a Skip action. TalkMe’s feedback includes Almost there, a displayed score and Report a problem. These are useful reminders that retry rules and feedback quality need their own interface, not just a microphone placed beneath an error message.

If attempts are limited, explain what consumes an attempt and do not charge one for permission denial or an empty recording. Preserve a way to report incorrect feedback without forcing the learner to abandon the lesson. Separate a learning choice to move on from a technical recovery that resubmits the same audio.

General learning re-engagement patterns help with returning later, but the immediate task needs its own recovery. Test a noisy recording, a valid accented attempt, a delayed score and an interruption during playback. Measure whether learners understand the correction and can complete the next attempt; more retries can mean confusion rather than productive practice.

An attempt test that separates capture and assessment

Build a test lesson with one word and one short sentence. Start with microphone permission unavailable. The learner should see an access explanation and retain the exercise, not receive a zero pronunciation score. Then allow capture but provide silence or an unsupported audio format through a controlled fixture. Verify that the interface distinguishes these failures from a genuinely assessed attempt.

Submit one valid take, then start another while the first result is delayed. The newest recording, score and explanation must stay matched. Play Coach and You in quick succession and confirm that one source stops before the other begins. The learner should be able to identify the audio they are hearing without looking at a subtle color change.

Test the same flow with larger text and without a sustained press gesture. Long instructions should not push the only stop control off screen. Let someone use the can’t-speak route, then return later and inspect the progress state. It should truthfully reflect that the exercise was deferred, not assessed and passed.

Ask the learner to explain one correction in their own words and choose a specific next practice action. A person who only repeats louder because a percentage was low has not received useful guidance. Review that misunderstanding alongside capture errors and repeated attempts. Keep recorded voice retention separate from the minimum diagnostic events needed for the test.

Include a case where the learner disagrees with the assessment. The report action should retain enough attempt identity for diagnosis without publishing the recording or demanding unrelated personal information. Make clear whether reporting changes the score or only sends feedback. If the lesson offers a bypass, label that as moving on rather than correcting the assessment. Review the recovery with a person whose accent differs from the model voice, and avoid interpreting disagreement as learner failure. A trustworthy exercise can acknowledge uncertainty while still offering a specific practice action and preserving progress.

Review one complete speaking attempt

Use these checks on a working build, not only on the polished reference state.

• ✓Explain tap, hold, stop and submit gestures before capture.
• ✓Separate unavailable audio, silence and processing failure from assessed pronunciation.
• ✓Keep the target phrase and lesson position stable during help and retry.
• ✓Match each score, recording and explanation to the same attempt.
• ✓Label model and learner playback distinctly and prevent overlap.
• ✓Provide readable feedback beyond color and a usable alternative to sustained holding.
• ✓Test uncertain scoring, attempt limits, reporting and a can’t-speak route.

Related design decisions

The permission-screen examples help explain microphone access without confusing denial with a learning failure.

The error-state article provides a broader recovery framework for capture and processing problems.