CANON
Every person sounds different. Some people talk fast. Some talk slow. Some speak loudly. Others are naturally quiet. Some pause frequently between thoughts. Others barely breathe between sentences. Some voices sit deep, others sit high. These differences aren't noise. They're the foundation of accurate perception.
CANON is the personal vocal baseline calibration system within the OPM pipeline. Its purpose is simple but essential: learn what this specific person sounds like at rest so that the rest of the system can detect when something changes.
Without personal calibration, a perception system has to rely on population averages. "The average person speaks at 140 words per minute, so someone speaking at 100 must be disengaged." That conclusion is wrong for anyone whose natural pace is 100. "Most people speak at a moderate volume, so someone who's quiet must be nervous." That conclusion is wrong for anyone who's always been a quiet speaker.
This is a problem that behavioral science has grappled with for decades. Research on individual differences in vocal communication consistently shows that population-level norms are poor predictors of individual behavior. Paul Watzlawick's axiom that "one cannot not communicate" only becomes useful when you know what this person's communication sounds like at rest. CANON builds that knowledge.
A guided audio version of this page, adapted for clarity and flow rather than read word for word.
Calibration starts from first contact and deepens automatically.
Messages, calls, and sessions all feed the same baseline.
CANON never interprets against population-average norms.
ORACLE, LUCID, and TRACE all become more specific with CANON active.
Progressive trust, not one-shot enrollment
CANON operates as a progressive calibration system with five distinct layers, each building on the one before it. The calibration deepens over time, automatically, requiring nothing from the person being calibrated except natural interaction. Data collection is cumulative across all interactions, regardless of how those interactions are distributed across messages, sessions, or days.
Layer 1: First Contact (0 to 60 seconds)
60 seconds of cumulative vocal dataThe first layer activates from the very first seconds of interaction. As soon as vocal signal extraction begins (through CYGNUS, CYGNUS Lite, or CYGNUS ECHO), CANON starts building a preliminary profile.
What gets captured: Resting vocal range (fundamental pitch, typical speech rate, default volume level) and natural pause frequency.
This first layer is rough. Sixty seconds isn't enough to distinguish a person's natural voice from how they sound right now. Someone might be nervous in a first interaction, or unusually energetic, or uncharacteristically flat. But it provides a starting point: an initial sketch that ORACLE can reference immediately. Even a rough baseline is better than no baseline.
Layer 2: Conversational Range (60 seconds to 30 minutes)
30 minutes of cumulative vocal dataAs interaction continues and the person moves through different conversational states (asking questions, listening, responding, thinking, hesitating), CANON expands the profile from a resting baseline to a dynamic range.
What gets captured: How wide their pitch range is when they're engaged versus when they're reflecting. How their speech rate fluctuates between active discussion and careful thought. How their volume shifts when they're emphasizing a point versus trailing off. How their pause patterns change under different conversational conditions.
Layer 2 establishes the bandwidth of normal for this person's voice. It isn't about what their resting state sounds like. It's about the full range of vocal states they naturally move through during conversation. Research on vocal accommodation (the way people unconsciously adjust their speech patterns to their conversational partner) shows that the first 30 minutes of interaction reveal the core of a person's dynamic vocal range.
Layer 3: Reliable Baseline (30 to 60 minutes)
60 minutes of cumulative vocal dataAfter an hour of accumulated interaction time, CANON reaches a level of confidence where its baseline becomes statistically reliable. At this point, deviations from the baseline carry genuine meaning.
What gets captured: Stable statistical profiles of all vocal features. Characteristic vocal response patterns (how this person's voice typically shifts when they're asked a question, when they're thinking, when they're changing topics). Consistent vocal signatures (whether they tend toward rising or falling intonation, whether they accelerate when engaged or when nervous).
Layer 3 is where CANON's data starts feeding meaningfully into ORACLE's pattern detection. A finding of 'vocal pitch 20% above personal baseline' requires a reliable personal baseline to be meaningful. Before Layer 3, ORACLE still operates but relies more heavily on general behavioral principles. After Layer 3, ORACLE can make genuinely personalized detections.
Layer 4: Established Profile (60 to 180 minutes)
180 minutes of cumulative vocal dataWith three hours of interaction history, CANON's vocal profile becomes robust. The system has observed enough conversational diversity to distinguish between vocal traits (stable characteristics of this person's voice) and vocal states (temporary conditions).
What gets captured: Long-term stable vocal characteristics, reliable response signatures (how this person's voice consistently behaves when they're engaged, uncertain, or processing), and vocal quality patterns that stabilize over time: the system knows this person's characteristic breathiness, roughness, or clarity.
Layer 5: Deep Calibration (180 to 600 minutes)
600 minutes (10 hours) of cumulative vocal dataAt ten hours of accumulated data, CANON reaches its deepest calibration level. The system knows this person's vocal signature with high precision across the full range of conversational contexts it's observed.
What gets captured: Micro-level vocal patterns that only emerge over extended observation. Subtle contextual variations (slightly different pitch patterns in different types of conversations, if the data spans different contexts). Nuanced cross-session consistency metrics that allow TRACE to detect even slight longitudinal vocal trends with confidence.
Layer 5 enables the most precise vocal analysis possible. ORACLE can detect deviations that would be invisible against a rougher baseline. LUCID can interpret vocal findings against a rich personal context. TRACE can identify vocal trends against a stable, high-confidence reference point.
Every vocal feature becomes personal
Pitch Profile
Each person has a characteristic pitch range. Some voices naturally sit at 110Hz, others at 220Hz. What matters isn't the absolute number but the personal range and how the person moves within it. CANON maps this range and tracks how it's used across different conversational contexts.
Volume Profile
Volume is one of the most individually variable features in human speech. Some people consistently speak at 60% of the average conversational volume. Others consistently speak at 140%. Population-based systems would interpret the quiet speaker as withdrawn and the loud speaker as aggressive. CANON knows both are just being themselves.
Speech Rate Profile
Natural speech rate varies dramatically between individuals. A person whose comfortable pace is 100 words per minute isn't disengaged. A person whose pace is 180 isn't anxious. They're each speaking at their natural rate. Deviations from that personal rate are meaningful. The absolute number isn't.
Pause Profile
Some people are natural pausers. They think before they speak, leave space between ideas, and are comfortable with silence. Others fill every gap instantly. A 2-second pause from a natural pauser is nothing. A 2-second pause from someone who never pauses is significant. CANON knows the difference.
Vocal Quality Profile
Breathiness, roughness, strain, clarity, tremor. These acoustic properties are surprisingly stable for individuals and highly variable between individuals. A voice that's always slightly breathy isn't showing a sign of distress. A voice that's normally clear and suddenly becomes breathy might be. CANON tracks these qualities as part of the personal vocal signature.
Rhythm and Cadence Profile
The temporal pattern of speech. Whether this person naturally speaks in long flowing sentences or short punchy phrases. Whether their rhythm is regular or varies. CANON captures the characteristic cadence so that departures from it can be detected.
Calibration deepens regardless of format
CANON's calibration doesn't reset between sessions or between messages. It accumulates.
The baseline is built from total vocal interaction time across all interactions. A person who's had five 30-minute sessions gives CANON 150 minutes of calibration data. A person who's sent sixty 30-second voice messages gives CANON 30 minutes. The calibration deepens regardless of how the time is distributed.
Initial sketch
Dynamic range established
Statistically reliable, personalized detection begins
Traits distinguished from states
High-precision vocal signature
The difference between guessing and knowing
The difference between population-average analysis and personally calibrated analysis is the difference between guessing and knowing.
Consider a concrete example. A person enters a session. Their speech rate is 95 words per minute. Their pitch variability is low. Their volume is quiet.
A system using population averages would flag all of these as below-normal. The interpretation might suggest disengagement, low energy, or withdrawal.
But CANON knows this person. Across 15 previous sessions, their average speech rate has been 98 words per minute. Their pitch variability has always been in the low range. Their volume has consistently been below average. Today's numbers aren't deviations. They're this person's normal.
Now consider that the same person, in the same session, shows a speech rate of 130 words per minute during a specific segment. Against population norms, 130 is perfectly average. But against this person's calibrated baseline, it's a 33% increase. That's significant. Something happened in that segment that shifted their vocal behavior well outside their personal norm. CANON ensures this deviation gets detected. Without calibration, it would be invisible.
This is the core insight behind decades of individual difference research in communication studies: absolute measures are far less informative than relative measures. Albert Mehrabian's early work on nonverbal communication is often simplified, but the more rigorous finding is that nonverbal signals become meaningful primarily in relation to a person's own baseline. CANON operationalizes this principle for the vocal channel.
Value compounds wherever interaction repeats
Education
Students interact with AI tutoring systems across an entire semester. CANON learns each student's natural vocal style within the first few sessions. The student who speaks softly isn't flagged as disengaged. The student who speaks in short bursts with long pauses isn't flagged as confused. Over months, CANON enables increasingly precise detection: when the system flags a vocal shift, it's genuinely unusual for this student.
Sales and Client-Facing Roles
Sales professionals who use perception-enhanced communication tools benefit from CANON calibration on their prospects and clients. A prospect who's naturally soft-spoken and suddenly becomes louder and faster during a product demo is showing a genuine engagement signal. A prospect who's naturally animated and shows the same vocal energy isn't showing anything unusual. CANON makes this distinction possible.
Public Speaking and Coaching
Speakers and presenters who work with perception-based coaching systems develop rich CANON profiles over dozens of practice sessions. The coach (human or AI) can then detect micro-shifts in vocal delivery that represent genuine departures from the speaker's established style. 'Your pitch variability dropped 25% compared to your personal baseline in the closing segment' is actionable coaching feedback. 'Your pitch variability was below the population average' isn't.
Clinical and Therapeutic Settings
In therapeutic contexts where the same client meets with a system or practitioner over weeks or months, CANON's vocal calibration provides a reference that enhances clinical observation. Gradual shifts in vocal quality, changes in pause architecture, or evolving speech rate patterns become visible against a well-calibrated personal baseline.
Research
Longitudinal research studies that track vocal patterns over time require exactly the kind of personal baseline that CANON provides. Researchers working with repeated-measures designs need to separate individual vocal differences from experimental effects. CANON does this automatically.
Personal context for every downstream layer
CANON sits alongside the main pipeline layers rather than within the sequential flow. It receives vocal input from any active CYGNUS configuration and enhances every downstream layer:
Signal Extraction + CANON
CYGNUS, CYGNUS Lite, or CYGNUS ECHO provides the raw vocal signals. CANON tells the system what's typical for this person's voice so that 'raw' can be understood in personal context.
ORACLE + CANON
ORACLE's rules evaluate deviations. With CANON's baseline, those deviations are measured against the individual's vocal norms rather than a generic standard. A 'significant pitch increase' is significant relative to this person's typical pitch.
LUCID + CANON
LUCID's interpretations gain precision when grounded in personal vocal calibration. 'This person's speech rate increased significantly compared to their typical range' is far more specific than 'this person's speech rate increased.'
TRACE + CANON
TRACE maintains the long-term behavioral profile. CANON provides the calibrated vocal baselines that TRACE uses for trend detection and deviation alerts. Together, they create a system where every vocal analysis is personal and every trend is measured against the individual's own trajectory.
Attributed data, never raw media retention
CANON data is Attributed Data. The personal vocal baselines, calibration layers, and cumulative metrics are linked to an individual and constitute personal data under GDPR.
What CANON stores: Statistical vocal profiles (means, ranges, distributions of vocal features), cumulative interaction time, calibration layer status, and cross-session vocal norms. All stored as structured numerical data.
What CANON doesn't store: Raw audio, per-frame signal data, or any original media. CANON builds its profiles from aggregated session-level vocal statistics, not from raw inputs.
Retention and deletion: CANON data is subject to the same retention policies and deletion rights as all Attributed Data. The deploying institution configures retention periods, and individuals can request deletion at any time. Deleting CANON data resets calibration to Layer 1. Future sessions begin as if no prior interaction had occurred.
Consent and transparency: The existence of CANON calibration and the type of data it stores are documented in our Privacy Policy and communicated to users through the deploying institution's own consent processes.
Practical boundaries, reset behavior, and standalone use
No. CANON calibrates automatically from the first interaction. There's no separate enrollment, no onboarding session, no special procedure. The person simply speaks naturally, and CANON builds the vocal baseline from what it hears.
Calibration architecture in one frame
CANON is the personal vocal baseline calibration system within the OPM pipeline. For information about how CANON enhances pattern detection, see . For information about how CANON feeds longitudinal tracking, see . For information about how perception data is used across our products, see our .