Skip to content

Continuous multilabel vocal expression vs. softmax emotion classes

A softmax classifier divides one unit of score across mutually exclusive labels. A multilabel model can score several qualities highly in the same clip. Oruk Spectra-2 returns 15 emotion and 16 speaking-style scores on independent 0–1 scales.

Updated

Listen to the recording

Listen to “The airplane is almost full.” Could the reading sound both sad and worried? Decide before inspecting the scores below.

Acted announcement · CREMA-D

“The airplane is almost full.”

Acted audio from CREMA-D. The recording catalog identifies the source files. Database license and contents license.

Reading several scores at once

In the saved Resonance 1 analysis of this recording, worried scored 0.992, sad 0.946 and disappointed 0.915 (rounded). Their sum exceeds one. A softmax over those same competing labels could not produce that vector. The actor’s source label was sad. A single source label does not settle whether listeners could also hear worry.

Spectra-2 likewise exposes separate emotion and style fields. These are model scores for audible expression. A score of 0.9 does not establish that nine out of ten listeners would agree, or that the speaker feels that emotion. That interpretation would need calibration against representative human judgments.

Try the comparison

For an application that queues tense calls for review, label examples in context and allow more than one quality per clip. Select a threshold for each label against the cost of missed clips and false alarms. Keep the original scores so a human can inspect why a clip was selected.

See the saved Resonance 1 response to this recording

See the Spectra-2 score fields and audio contract

Explore the other paralinguistic signals.

Try it on your own audio.

The demo needs no account. The trial needs a card.