acoustic fingerprint pilot

Labnote 016: acoustic fingerprint pilot🔗

Question🔗

Do lightweight acoustic fingerprints recover perceived emphasis and expressive delivery in the Prosody 015 corpus?

Frozen pilot🔗

• Reuse the eight private, blindly annotated Prosody 015 clips without synthesis or editing.
• Use faster-whisper small.en only to obtain word timestamps.
• Independently measure duration, voiced fraction, internal pauses, RMS-energy variation,

F0 center/spread/range/slope, three-part energy and F0 contours, and word-local energy, duration, F0, and preceding pause.

• Define each word's prominence as the unweighted mean of its available within-utterance

standardized word features.

• Count perceived-focus recovery when the highest-prominence aligned word appears in one

of the annotator's comma-separated focus phrases.

• Reduce the exploratory speech-act ontology to neutral-only versus expressive-any, then

compare between-label and within-label Euclidean distances after feature standardization. Evaluate the signed separation using the exact permutation distribution preserving the observed class counts; directional success requires positive separation and p <= 0.10.

The tiny corpus cannot train a classifier or validate affect recognition. This pilot asks only whether the representation contains enough of the listener-observed structure to justify collecting more candidates.

Result🔗

All eight clips aligned to the expected transcript exactly. The untrained, equal-weight word-prominence score recovered one of the annotator's perceived-focus phrases in five of eight clips.

Three clips carried at least one non-neutral speech-act label and five were neutral-only. The preregistered standardized fingerprint distance placed expressive readings farther from neutral readings than same-label readings by 2.328 distance units. That separation was the strongest of all 56 assignments preserving the observed 3/5 class split, giving an exact one-sided permutation value of 1/56 = 0.0179 and passing the directional gate.

An implementation audit before interpretation found that a missing voiced contour bin could be encoded as an extreme sentinel. Amendment 001 defined contours over the active voiced span and replaced missing pitch with the within-clip median; the initial output was discarded, while the frozen labels, feature families, statistic, and threshold were unchanged.

This is positive evidence for the representation, not for a deployable recognizer. The same listener labeled all eight clips, only two transcripts and one synthetic speaker are represented, and the expressive class contains three examples. The next justified step is to generate and label a larger candidate pool across more utterances, then freeze a held-out retrieval test before fitting or selecting feature weights.