Skip to contents

Resample each token's f0 contour to N equidistant points

Source: R/resample.R
resample_f0_equal.Rd

For each token, replaces the native pitch-frame grid with n points equally spaced between the token's first and last frame time, so every token contributes the same number of samples at the same proportional positions (with n = 21, one point every 5% of the token's duration; with n = 11, every 10%). Intended for monosyllabic (single-contour) tokens; for multisyllabic data, TextGrid landmark axes (see normalise_time_landmarks()) are usually the better route.

Usage

resample_f0_equal(
  df,
  n = 21,
  token = "token",
  time = "time",
  f0 = "f0",
  intensity = "intensity",
  method = c("linear", "nearest")
)

Arguments

df

Long-format f0 data frame.

n

Number of points per token (at least 2). Default 21.

token, time, f0

Column names. Defaults "token", "time", "f0".

intensity

Name of the optional intensity column; used only when present in df. Default "intensity".

method

How each point takes its value, mirroring the two interpolation choices Praat offers in Pitch: Get value at time...: "linear" (default, and Praat's own default) applies the rule in Details; "nearest" always takes the value of the frame the point falls in, and is NA when that frame is unvoiced, so nothing is ever computed.

Value

A data frame with the token, point (measurement number, 1...n), time (seconds, on the new grid), time_prop (proportional position 0-1 within the token), and f0 columns, plus intensity (when present) and any carried token-constant columns. Input columns named point or time_prop are dropped, since those names are computed here. Rows are sorted by time within each token; tokens keep their order of first appearance. The attribute dropped_columns names any columns that could not be carried.

Details

f0 at each new point follows Praat's own rule for Pitch: Get value at time... Linear (Sampled_getValueAtX() in Praat's source), so a contour resampled here matches one extracted by a Praat script querying the same times. Of the two native frames bracketing the point, call the closer one near and the other far:

  • both voiced: the value is linearly interpolated between them;

  • far unvoiced: the point takes the near frame's measured value, so no value is blended across an unvoiced frame while a usable measurement at the edge of one is not thrown away;

  • near unvoiced: the point is NA, since nothing was measured there.

A point landing exactly on a frame has a phase of 0 and therefore takes that frame's value exactly. At a point falling exactly midway between two frames the choice of near is a floating-point tie and may fall either way; it only changes the answer when exactly one of the two is unvoiced.

This rests on unvoiced frames being present as NA (or 0 Hz) rows, which is how the wrassp and .Pitch paths deliver them. Sparse input carries no such rows: a .PitchTier, or a CSV listing only voiced samples, represents an unvoiced stretch as nothing more than a wide gap between two voiced anchors, so both bracketing frames are voiced and the point is interpolated straight across. Praat does the same on a PitchTier. Add explicit NA rows if such stretches should read as unvoiced.

With method = "nearest" (Praat's other option) the middle case goes away: a point always takes the nearer frame's measured value, or NA when that frame is unvoiced. Every exported value is then a number the tracker produced, at the cost of a timing error of up to half a frame step.

If an intensity column is present, it is linearly interpolated across all frames with finite intensity (ends extended, matching how the extraction aligns the intensity track to the f0 frames).

Any other column is carried through when it is constant within every token (metadata, token_dropped, and similar token-level columns); per-frame columns that vary within a token cannot survive a change of grid and are dropped — their names are recorded in the dropped_columns attribute of the result.

Rows with a missing time are dropped, and rows sharing a frame time are collapsed to the first of them, so interpolation has strictly increasing anchors. A token left with fewer than two distinct frame times cannot be resampled and keeps its single remaining row, with time_prop = 0.5.

See also

normalise_time_token() for adding a proportional time column without changing the sampling grid.