SpeakBalance

Privacy-conscious participation analytics experiment for online language classes.

SpeakBalance was a short technical feasibility prototype. It explored whether Zoom active-speaker metadata could estimate participation balance without recording or transcribing lesson content.

Project status: Technical feasibility prototype - concluded after identifying a platform-data limitation.

The Problem

Language teachers and teacher trainers often discuss Teacher Talk Time and Student Talk Time, but measuring that balance objectively during a live online lesson is difficult.

Recording or transcribing lessons can create privacy, consent, and data handling concerns. SpeakBalance explored whether a lighter metadata-only signal could support teacher reflection.

The Hypothesis

Could Zoom active-speaker metadata estimate participation balance without using lesson audio, video, transcripts, chat, or content analysis?

The intended data boundary was:

• participant display name
• participant metadata needed for in-session event identity
• active-speaker event time
• teacher/student role mapping
• estimated active-speaker duration

What I Built

• React/Vite frontend
• Zoom Apps SDK onActiveSpeakerChange integration
• local active-speaker capture flow
• event normalisation and duplicate-event handling
• speaking-segment generation
• inactivity-timeout estimation for unclear gaps
• teacher/student/unknown role mapping
• per-speaker and per-role participation summaries
• final session report view
• plain text report export
• sample-data demo pipeline for local review

What I Tested

The Zoom integration was tested in real Zoom desktop meetings, not only against mocked data.

The tests checked whether active-speaker events fired, whether they included usable participant metadata, how duplicate events behaved, how silence was represented, and whether speaker handovers could be detected.

What I Found

• onActiveSpeakerChange successfully identified active participants.
• Repeated and duplicate events required cleaning before analytics.
• Same-speaker events could be sparse during continuous speech.
• Zoom did not provide a sufficiently reliable stop/silence signal for exact stopwatch-style timing in this implementation.
• Timeout logic could estimate unclear gaps, but it could not measure silence exactly.
• Additional Zoom SDK event paths were investigated, but they did not provide a reliable replacement for exact start/stop timing.

The result is best described as estimated participation analytics, not exact talk-time measurement.

Decision

Development stopped at the feasibility stage.

The prototype proved that the app could integrate with Zoom and capture useful active-speaker metadata. It also showed that the available metadata was not granular or consistent enough for accurate speaking-duration measurement.

Building authentication, databases, backend storage, payments, deployment, or a larger product workflow on top of that assumption would have been premature. Stopping was a deliberate scope-control decision based on the technical finding.

Privacy By Design

The prototype does not record, store, upload, transcribe, or analyse lesson audio/video/content. It is designed around active-speaker metadata and local prototype processing.

SpeakBalance is privacy-conscious, not privacy-free: participant metadata still needs careful handling, consent, and clear disclosure in any future product.

Tech Stack

• React
• Vite
• JavaScript
• Zoom Apps SDK

Running Locally

Install dependencies:

npm install

Run the local app:

npm run dev

Run the sample analytics demos:

node src/analytics/demoCalculateDurations.js
node src/analytics/demoNormaliseZoomSpeakerEvents.js
node src/analytics/demoSummariseSpeakerSegments.js
node src/analytics/demoApplySpeakerRoles.js
node src/reporting/demoGenerateTextReport.js

Build the app:

npm run build

Generated text reports are written to reports/ and ignored by Git.

Documentation

• Case study
• Privacy by design
• Zoom active-speaker spike
• Calibration findings
• Technical decision