Regresses continuous valence-arousal-dominance affect from each AudioFrame with a wav2vec2 dimensional model, emitting a {arousal: Double, dominance: Double, valence: Double} record, a raw [Double] triple, or [Analytics.Metric].
Typical backends
Score recorded speech for affect and log the dimensional values.
Score affect from live microphone audio and route on the value through a conditional branch.
Caveats
- State lifetimeThe model and the device are read once at startup, so changing either takes effect only after a container restart.
- CompatibilityAudio is resampled to 16 kHz internally regardless of the upstream sample rate, matching the model's training rate.
- AccuracyOutput is continuous regression over the valence-arousal-dominance space, not a discrete emotion label; the three values are each roughly bounded and have no calibrated probability interpretation.
- I/O contractEmpty audio yields a neutral midpoint reading rather than failing, so downstream consumers always receive a complete triple.
- LimitationA single affect estimate is produced for the whole frame, so very long segments average over their internal variation; segment by speech turn upstream when finer granularity matters.