Realtime Paralinguistic Analysis Engine
The same audio yields the same value
CASP ENGINE does not convert speech into text. It extracts indicators from the voice signal and aligns them as an amount of deviation against a personal baseline. Even when the reference values are updated, past analysis results remain reproducible.
Analysis pipeline
Five stages operate in order. Each stage's output feeds the next, and every stage's value remains in the source record.
-
1. Raw Signal
The session audio is segmented into 10–15 second windows on pause boundaries. No text conversion.
-
2. Feature
About twenty paralinguistic indicators are extracted: speech rate, pause ratio, pitch standard deviation, energy variance, and others.
-
3. Transition
Not a single point but the change between one segment and the next. Shifts in rate, pitch, and volume around a pause are captured here.
-
4. Score
Indicators and transition values are combined into a raw score. This is a source value that has undergone no correction.
-
5. Calibration
The raw score is aligned to a personal baseline and group distribution into an interpretable coordinate system. Individual change and position within the group can be read together.
Realtime — separating live processing from reference updates
If live analysis wavered every time reference values changed, the engine could not be trusted. CASP keeps the two flows apart.
Realtime · per session
Incoming audio is immediately converted into indicators and a raw score. This path does not query reference data, so it is unaffected by reference-value updates.
Batch · reference sync
Reference data does not intervene directly in live analysis; it is synchronised in batches. Even if the calibration basis changes, sessions in progress are unaffected.
Reproducibility
To use an analysis result as evidence, the same value must appear months later. If past results shifted as reference values grew more refined, those results could not serve as evidence. CASP preserves both the source value and the calibrated value.
Raw-score preservation
The uncorrected source value is stored as is. It does not change when reference values are updated.
Calibrated-score preservation
The corrected value is stored separately, together with a record of which basis it was corrected against.
Snapshot calibration
The state of the reference data at a given moment is kept as a snapshot. Past analyses can be re-checked against the basis of their time.
Two references
The basis for producing indicators and the basis for interpreting the result play different roles; merging them misflags an individual's traits as anomalies.
Anchor Point
Fixes the endpoints of the reference axis: what to set as 0 and what as 100 when computing an amount of change.
Normal Range
The group distribution by region, age, and sex. It lets an individual's indicators be read as a relative position within the group.