Vocal Analyzer Docs

← Back to app

Pitch (F0)

Fundamental frequency — the rate at which your vocal folds vibrate, measured in Hertz — is the strongest single acoustic cue to perceived gender in voice (Leung et al. 2018 report it accounts for ~42% of listener-judgement variance in their English-corpus study). Vocal Analyzer treats pitch as the primary cue with the strongest direct perceptual validation, and it is the only metric whose reference bands are drawn from converging calibrations across several studies (Gelfer & Bennett 2013; Davies et al. 2015; Holmberg et al. 2010).

Reference bands

The spectrum bar at the top of the dashboard overlays the following bands onto your median F0 (in Hz):

Band Low High
Masculine 90 155
Androgynous 155 165
Feminine 165 255

Caveat: Bands are reference overlays, not clinical cutoffs. They describe published central tendencies for English adult speech, not validated predictors of individual listener judgement.

Reported values

Vocal Analyzer reports several pitch statistics, each serving a different question:

Median F0

The central value of voiced F0 across the whole recording. This is what the spectrum bar plots. Less sensitive to occasional creak or falsetto frames than the mean, which is why it — not the mean — is the headline number.

F0 distribution

The 5th percentile, median, and 95th percentile of voiced F0, plus the interquartile range (IQR = p75 − p25). Together these summarise how tightly F0 clusters around the median. A narrow IQR with close p5/p95 values means a fairly monotone speaker; a wide IQR with distant percentiles means a very expressive speaker.

Pitch expressiveness

Standard deviation of F0 in semitones. Vocal Analyzer reports this in semitones rather than Hz because a semitone represents the same perceptual step whether you are speaking at 100 Hz or 250 Hz (Pépiot 2014). Typical conversational English speech sits around 2–4 ST; sustained vowels are usually under 1 ST; highly animated speech can exceed 5–6 ST.

The editable expressiveness overlay defaults to 3–6 ST. This is a deliberately broad continuity target for tracking practice over time, not a gendered or clinical reference range: language, task, affect, and speaking style can all move pitch variation substantially. All reference profiles use the same default, and Settings can reshape it for an individual goal.

Pitch range

Lowest and highest detected voiced F0 in the recording. Reported as raw min–max in Hz. Sensitive to occasional creak frames at the low end and to falsetto or noise artifacts at the high end — the F0 distribution tile (above) is a more stable view.

Pitch dynamism quotient (PDQ)

Standard deviation divided by the mean. A scale-invariant measure of variability (Leyns et al. 2024). Reported as a descriptive cue without reference bands.

Percentage voiced

Fraction of the recording that contained voiced speech. Long pauses or unvoiced consonants lower this number. Very low values (under ~40%) usually mean the recording has long silences or a lot of unvoiced sound — the median F0 is still valid if it crosses the minimum- duration threshold for a stable estimate.

Algorithm

Vocal Analyzer uses a two-pass, adaptive F0 extraction based on De Looze & Hirst (2008):

  1. The first pass uses a wide default range (the settings you see in the sidebar — pitch floor / ceiling). The median of that first pass is used to derive speaker-adaptive bounds.
  2. The second pass runs with those tightened bounds. Adaptive bounds give stabler F0 tracks and better reject creak and falsetto outliers than a fixed wide range would.

The effective_floor_hz and effective_ceiling_hz fields in the raw result record the bounds actually used for the second pass; these are surfaced in the Detailed Measurements card on the dashboard.

Under the hood the extraction is Praat's autocorrelation pitch tracker via parselmouth. See src/vocal_analyzer/engine/pitch.py in the repository for the implementation.

Caveats and measurement warnings

References