Shengjian

The problem: understanding now, finding it later

In a bilingual lesson, listening, comparing languages, and deciding what to write compete for attention. Captions help with immediate understanding, but reviewing a difficult idea later also means finding the question, the original audio, and the slide being discussed. I treated that transition from following a lesson to revisiting it as the core product problem.

The starting point was my own learning workflow and a second pre-created account, rather than interviews with a wider learner population. I owned product scoping, interaction design, browser audio capture, AI-service integration, storage and recovery, and deployment.

Design around the learning task

The workflow now spans preparation, live listening, and later review. These tasks describe the implemented product, not findings from a formal task study.

A 28-second product demo

Scope: a working Web product for a small, known audience

I chose a browser-based product for personal Mac and iPad use. Recording-only and live bilingual modes share a recording library. The current workflow adds optional course selection, recognition-term reuse, classroom photos, question and important-point markers, and manual review status alongside paragraph replay and study notes.

Course names organize a person’s recordings. Term reuse is explicit and uses the most recent completed live recording for that course; it is not a shared glossary or a guarantee that specialist words will be recognized correctly.

Capture still needs a foreground page and an awake screen. The product warns about that condition and responds to an interrupted microphone. Public signup, payment, and native-app distribution remain deferred.

Live: keep attention on the conversation

Once recording begins, inactive language and mode settings collapse while pause, finish, learning markers, and the camera remain reachable. This reserves more of the screen for the material being heard.

Original and translated text are paired in the same passage. Focus mode and font controls adjust reading without changing the recording state. Elapsed session time remains separate from saved audio duration, because a pause contributes to the former.

Keep the slide with the moment

A classroom photo keeps its position in the recorded audio. It appears alongside the relevant bilingual passage, or as a separate image card when there is no matching text. The learner can enlarge it, replay that moment, or adjust its placement.

Photos save through their own device queue, so an image-upload failure does not stop microphone capture. Audio can be saved while a photo is still pending; the interface keeps that distinction visible and waits for accepted photos before completing the cloud archive.

Replay: return to the difficult passage

Paragraph timestamps remain a direct route back to the original audio. A learner can replay the sentence they want to check, change speed, and use short skips without searching through an entire lesson.

Following the current paragraph helps during playback, while keyboard focus entering the text pauses following for deliberate reading. Photos also have their own replay action, and enlarging an image does not trigger paragraph playback.

Mark now, return after class

The live workspace offers two quick marks: a question and an important point. Each stores an audio position rather than a paragraph ID, so later transcript layout changes do not move the original moment.

After class, a marker can replay from ten seconds earlier and show nearby bilingual text plus the nearest preceding classroom photo. The learner explicitly marks it as reviewed, can filter pending or reviewed items, and can change it back. Reviewed records an action; it does not certify understanding or transcript accuracy.

Review: check notes against the source

Study notes group an overview, topic explanations, details, source examples, explicit tasks, and points needing confirmation. Empty tasks or uncertainties are omitted, and the prompt asks the service to stay within the transcript.

Generation is a deliberate request. Results are cached and replacement requires explicit regeneration; a failed or incomplete request preserves the previous notes. Changing a course or review status does not trigger an AI request. Photos are not sent to the notes service.

The screenshot shows manually prepared notes for the same fictional lesson. Model output still requires checking against the bilingual transcript and original audio.

Organize lessons for the next review

Course labels and filters make the recording library useful across multiple lessons. Course filtering combines with name, date, and language search, while a pending-review filter helps a learner find recordings with questions still to revisit.

Course assignments and manual review status persist with the recording and its archive. The saved-audio state remains separate from review progress and cloud-backup status; completing one does not silently complete the others.

Documented iteration: issue, change, check

These examples come from the September 29 interface review and isolated browser checks. They are concrete implementation revisions, not before-and-after participant performance measures.

A completion message must say what actually completed

Ending microphone capture, finishing the transcript, and verifying the cloud archive remain independent events. Photos introduce another pending-save condition, so saved audio can coexist with an image still waiting to upload.

With durable device storage available, short upload failures can retry without immediately ending capture. The app stops when the pending queue approaches its capacity or saving can no longer continue safely, and keeps a recovery path. A recognition disconnect is reported separately; the missing interval is not automatically translated.

Evidence: what was checked, and how

Real use, user-reported acceptance, and isolated failure drills answer different questions. I keep those evidence types separate rather than combining them into a single reliability or learning-effectiveness claim.

What I learned: make each promise explicit

The difficult design work was deciding what each state promises. A visible translation, saved audio, an uploaded photo, a reviewed question, and a verified cloud copy are different outcomes. Keeping those distinctions visible makes the classroom-to-review workflow easier to reason about.

The next evaluation should ask new invited learners to prepare a lesson, mark a difficult moment, capture a slide, replay it, and return to their pending questions. Task completion, errors, and whether questions are actually revisited would test the workflow. This is proposed work, not a completed participant study.