KKEBI Architecture Overview: From Voice Check-in to Organizational Signal

The KKEBI architecture is designed so that a single short voice check-in passes through five stages: capture → feature extraction → personal baseline → aggregation → organization-level de-identified signal. The heart of this flow is that the part that helps the individual understand themselves and the part delivered to the organization are clearly separated, with de-identification sitting at that boundary.

This article does not cover which vendor or model KKEBI uses. Instead, at a conceptual level, it lays out the design principles and trade-offs that a system for voice-based mental care of workplace members should uphold.

---

01. The Full Flow at a Glance

Suppose a person leaves a short voice check-in in KKEBI. That voice flows through the following sequence.

1. Capture: record the voice check-in on the basis of consent.
2. Feature Extraction: extract acoustic and linguistic features from the raw audio that may relate to state.
3. Personal Baseline: compute change by comparing against that person’s usual pattern.
4. Aggregation: bundle into units of enough people that no individual can be identified.
5. Org-level Signal: deliver only de-identified group trends to the organization.

The first three stages are for the individual, and the last two are for the organization. The boundary between them is the most important point in this architecture.

02. Capture: It Starts on the Basis of Consent

The first stage is recording the voice check-in. The principle to uphold here is not to gather as much data as possible but to collect only the minimum needed for the purpose.

• The purpose and scope of collection must be understandable to the user.
• How much of the raw audio to retain, and to what extent, should be decided carefully from a privacy standpoint.
• Where possible, performing sensitive processing close to the user’s device reduces exposure risk.

The trade-off at this stage is clear. Retaining the raw audio for a long time enables more sophisticated analysis later, but privacy risk grows accordingly. So the capture stage already decides “what will not be kept.”

03. Feature Extraction: Not the Content but the Manner

This is the stage of extracting signals from the captured voice that may relate to state. What matters here is not just the content of speech but the manner of speaking. Prosody (pitch, rate, pauses), voice quality, and spectral features are representative examples. Why such acoustic features can be linked to state is covered in more detail in the Voice Biomarker Guide.

An important design decision is where the features are extracted. Extracting only features on the device and minimizing the raw audio favors privacy, but constrains compute resources and model updates. Processing on the server is flexible but requires managing the risk that comes with moving raw data. There is no single right answer; you balance according to sensitivity and purpose.

04. Personal Baseline: Change, Not Absolute Value

What KKEBI attends to is not “is this person’s voice higher than others’” but “how is it changing compared with this person’s usual state.”

People differ in their baseline pitch, speaking rate, and pausing habits. Applying a single absolute standard to everyone easily mistakes individual differences for changes in state. So we establish each person’s usual pattern as a baseline and, using data from many points in time, observe the long-term change that deviates from that baseline. This principle was also explained as the key to early detection in Early Detection of Employee Burnout.

The output up to this stage is entirely for that person themselves. It is used to understand and reflect on one’s own state, and here voice is merely a supporting signal interpreted alongside other signals such as language and check-ins.

05. Aggregation and Org-level Signal: The Individual Disappears

Between the personal stage and the organizational stage there is a boundary that cannot be crossed. What is delivered to the organization is not an individual’s result but only a group-level signal aggregated so that no individual can be identified.

• A signal is only created once it is bundled across enough people (a minimum aggregation threshold).
• Outputs in a form that could trace back to a specific individual are not provided to the organization.
• What the organization sees is not “who is struggling” but “which direction the group is moving.”

Thanks to this boundary, the organization can detect signals such as an overall burnout trend among members early, without infringing on individual privacy. How organization-level signals connect to outcomes such as turnover intention is covered in Burnout and Turnover Intention.

06. The Three Promises This Architecture Keeps

In summary, the KKEBI architecture stands on three principles.

• Privacy-by-design: de-identification and aggregation thresholds are not features added later but a design engraved into the boundaries of the data flow from the start.
• Personal-baseline-centered: we look at within-person change, not absolute values.
• Voice as a supporting signal: we do not judge from a single voice sample but interpret it alongside other signals.

How these principles actually lead to a product and philosophy can be seen in About SYMPLE.

---

References

1. Insel TR. Digital phenotyping: technology for a new science of behavior. JAMA. 2017;318(13):1215–1216.
2. Low DM, Bentley KH, Ghosh SS. Automated assessment of psychiatric disorders using speech: A systematic review. Laryngoscope Investigative Otolaryngology. 2020;5(1):96–116.
3. Cummins N, Scherer S, Krajewski J, et al. A review of depression and suicide risk assessment using speech analysis. Speech Communication. 2015;71:10–49.