Methodology

How EchoDepth works.
The evidence base.

This page documents EchoDepth's technical methodology — how the platform analyses emotion, how it was trained, how bias is addressed, and how the ROI model is constructed.

Analysis Foundation

Facial Action Coding System (FACS)

EchoDepth's core analysis is built on the Facial Action Coding System (FACS), developed by Paul Ekman and Wallace Friesen. FACS is the scientific standard for objective facial expression measurement — it decomposes every visible facial movement into discrete, numbered Action Units (AUs).

EchoDepth tracks 44 Action Units per frame in real time. Because many AUs are involuntary — they cannot be consciously suppressed or manufactured — they provide an emotional signal that is independent of self-report and resistant to deliberate manipulation.

AU activations are scored on a continuous 0.0–1.0 intensity scale per frame. Temporal coherence is scored across the full session window, not individual frames — meaning EchoDepth measures emotional trajectories, not snapshots.

Output Model

Valence, Arousal, Dominance (VAD)

AU activations are mapped to three continuous emotional dimensions:

Valence (0.0–1.0) — Positive to negative emotional state. Low valence indicates distress, displeasure or anxiety.
Arousal (0.0–1.0) — Calm to high-activation. Elevated arousal may indicate stress, fear, urgency or excitement.
Dominance (0.0–1.0) — Submissive to in-control. Low dominance indicates vulnerability or powerlessness.

VAD scores are computed per frame and per session. Baseline calibration is performed from the opening window of each session, and all subsequent scores include delta-from-baseline values.

Training Data

6 countries. 14 cultures.
By design, not by accident.

Many emotion AI systems fail outside the demographic of their training data — producing systematically biased results for underrepresented groups. EchoDepth was deliberately built to avoid this.

Training data was collected across 6 countries and 14 cultural cohorts. Cultural expression variance is modelled rather than averaged — meaning the system accounts for documented differences in how emotions are expressed across cultures, rather than treating one culture's expression baseline as universal.

All training data was collected from spontaneous, naturalistic video — not posed expression datasets or laboratory conditions. This ensures the model performs on real-world interactions, not idealised inputs.

Bias Reduction

Active mitigation,
not passive hope.

Bias reduction in EchoDepth is structural, not aspirational:

  • Demographic representation — training data balanced across age, gender and ethnicity within each cultural cohort
  • Expression variance modelling — cultural differences in emotional display rules are encoded, not flattened
  • No posed datasets — all training on spontaneous expression captured in naturalistic settings
  • Active audit — ongoing bias testing across demographic segments
  • Deployment review — Cavefish will not deploy EchoDepth in contexts where demographic bias could produce discriminatory outcomes

Privacy Architecture

No biometric data stored. No exceptions.

EchoDepth's privacy architecture is structural — not procedural. The system is designed so that biometric data storage is technically impossible, not merely prohibited by policy.

🔒
In-memory processing

Video frames are processed in memory and immediately discarded. No raw video or facial images are retained at any point.

📊
Anonymised outputs only

Only VAD scores and AU activation data are output. These are statistical summaries — they cannot be reverse-engineered to reconstruct a face.

🏭
Edge deployment available

For organisations where video data cannot leave the premises, EchoDepth supports on-premise edge deployment with full API feature parity.

ROI Model

How the 3.1× ROI
is calculated.

The ROI figures presented on this site are modelled projections based on the following assumptions. They are not guaranteed outcomes — actual results will vary by organisation, volume and implementation.

Important: These are internal model projections. They have not been independently audited. Actual ROI will depend on case volume, current complaint rates, operational efficiency and implementation quality.

Model assumptions

Baseline vulnerable cases/year50,000
Escalation reduction25%
Vulnerability detection improvement15%
Modelled annual savings£467,000
Proof of concept cost£50,000
Year one ROI3.1×

Escalation cost benchmarks derived from FCA complaints data and industry analysis. Model does not include secondary benefits (reduced FCA scrutiny, reputational protection, avoided fines).

External Sources

Statistics referenced on this site

All externally sourced statistics used on the EchoDepth for Financial Services website are listed below with their primary sources.

£176M — FCA fines issued in 2024 (up 230%)

Source: FCA — 2024 Fines. 230% increase confirmed by SteelEye Financial Services Fine Tracker 2024.

£479M — Customer redress paid by UK firms in 2024

Source: FCA Complaints Data. Calculated from H1 2024 (£243M) + H2 2024 (£236M).

3.1× ROI / £467,000 annual savings / £50,000 PoC cost

Source: Internal model. See ROI model assumptions above. Not independently audited.

44 Action Units / FACS framework

Source: Ekman, P. & Friesen, W.V. (1978). Facial Action Coding System: A Technique for the Measurement of Facial Movement. Consulting Psychologists Press.

6 countries / 14 cultural cohorts

Source: Internal training data specification. Detailed breakdown available under NDA on request. Contact us.

Questions about our methodology?

We're happy to discuss our technical approach, training data, bias reduction strategy, or ROI model in detail.

Book Discovery Call