Sibilants

Sibilants are the /s/- and /z/-like high-frequency fricative sounds. Their acoustic shape — specifically where the high-frequency energy concentrates and how broadly it spreads — has been discussed in the voice-training literature as a perceptually gendered cue. Vocal Analyzer reports a set of sibilant-spectrum summaries when the heuristic detector finds enough fricative-like energy in the recording.

Default caveat: With the default energy detector, confidence on sibilant tiles is capped at moderate because it cannot distinguish /s/ from similar high-frequency fricatives (/ʃ/, noise bursts, some affricates) reliably. Switching to the optional word-gated method (see below) uses transcript words, or aligned /s/ and /z/ phone intervals when available, to raise that cap to high when the timing layer and spectral detector agree.

Reported values

When sibilant segments are detected, four tiles appear in the dashboard's Voice Detail card:

Center of gravity

The spectral center of gravity (CoG) in Hz. Higher values indicate energy concentrated toward higher frequencies; lower values indicate a more "muffled" fricative spectrum. Typical /s/ CoG in clean recordings of English speech sits above 5 kHz.

Spread

Spectral spread around the CoG, in Hz. A broader spread corresponds to a less concentrated fricative spectrum — the fricative has energy across a wider frequency range.

Skewness

A dimensionless measure of asymmetry in the sibilant spectrum. Near zero means approximately symmetric around the CoG; positive skew means energy trails toward higher frequencies, negative toward lower.

Kurtosis

A dimensionless measure of how peaked or flat the fricative spectrum is. Higher values mean a sharper, more concentrated peak.

Detector details

All four moments use a fixed 1–8 kHz band. Batch sibilant analysis resamples higher-rate input to 16 kHz with the same anti-alias filter used by preprocessing, including when --no-preprocess is selected. Thus the same recording uses the same filter and FFT grid on the web and CLI paths. Historical results computed over 1–11 kHz use a different measurement method.

The sibilant detector works on high-frequency energy above roughly 4.5 kHz, looking for segments with a characteristic spectral shape consistent with /s/-like frication. Because the detector does not use transcription to localise phones, it will:

• Miss sibilants recorded at low sample rates where the 4.5 kHz floor sits close to the Nyquist ceiling.
• Detect non-sibilant high-frequency content (tongue clicks, breath noise bursts, some plosive releases) and mix it into the aggregate.
• Vary heavily with microphone quality and room acoustics — the detector's idea of a "sibilant-like" spectrum is sensitive to the recording chain.

Why no gender bands

Published acoustic sex differences for sibilant CoG exist, but are small relative to the speaker-to-speaker and recording-chain variability Vocal Analyzer sees in practice. Given the moderate- confidence framing of the detector, a band overlay would imply more precision than the measurement actually has. Sibilants are reported as descriptive cues only.

Choosing a sibilant method

The sidebar Sibilant analysis setting accepts three values:

• Energy heuristic (default) — runs the high-frequency detector over the whole recording and reports the four tiles when enough sibilant-like segments are found. Confidence caps at moderate.
• Word-gated — runs the same spectral detector but only inside transcript words whose orthography suggests /s/ or /z/. Requires transcription, which only the installed app has (its web UI runs Whisper automatically; the CLI needs the transcribe extra and a --sibilant-method word_gated flag). If aligned phone intervals are present, aligned /s/ and /z/ phones take precedence over transcript-word windows. When both the timing layer and spectral detector agree on a segment, the tile confidence unlocks to high with a word_gated or phone_gated reason. If timing is unavailable, the sibilant tiles show transcript_unavailable; if the available timing layer contains no sibilant candidates, they show either no_sibilant_words_in_transcript or no_sibilant_phones_in_alignment.
• Off — skips the detector entirely. Useful for recordings where the high-frequency spectrum is dominated by noise or music and the detector produces misleading values.

Word-gated trade-offs

The orthographic filter is intentionally narrow: it accepts words containing s or z and rejects anything containing sh (conservatively covering English /ʃ/, including cases like "mishap" where the s and h happen to sit across a syllable boundary). This leaves a few honest edge cases:

• False positive: words like "measure" (/ʒ/) match on s and are scanned. The spectral detector still has to agree for the segment to contribute, which bounds the damage but does not eliminate it.
• False negative: soft c ("city" → /s/) and silent letters ("island") are skipped. The whole-recording energy method remains the right choice if you care about every fricative.

If either trade-off matters for your use case, stick with the default energy method.

Aligned phone windows avoid the orthographic trade-offs above when a phone-level layer is available, but they inherit the aligner's own dictionary, model, transcript, and language limits. They are timing evidence, not pronunciation scoring.

Caveats

• Sample rate. Sibilants require at least 16 kHz input to cover the full analysis band. Lower-rate input is not upsampled.
• Short recordings. A handful of detected segments is not enough to produce a stable CoG. The segment count is reported as part of the tile's tooltip.
• Lossy compression (especially MP3 at low bitrates) can remove the high-frequency content the detector depends on.

References

• Fuchs, S., & Toda, M. (2010). Do differences in male versus female /s/ reflect biological or sociophonetic factors? In Fuchs et al. (eds.), Turbulent Sounds: An Interdisciplinary Guide. De Gruyter Mouton.
• Munson, B., McDonald, E. C., DeBoe, N. L., & White, A. R. (2006). The acoustic and perceptual bases of judgments of women and men's sexual orientation from read speech. Journal of Phonetics 34(2), 202–240.