Skip to contents

Does a time column look already normalised to [0, 1]?

Source: R/time_norm.R
time_already_normalised.Rd

Heuristic used by the modelling functions (via resolve_time_norm()) to decide whether a time column is already proportional (per-token normalised to the unit interval) and should be used as-is rather than min-max rescaled within each token again.

Usage

time_already_normalised(time_values, tokens, eps = 1e-08)

Arguments

time_values

Numeric vector of time values.

tokens

Vector of token IDs, same length as time_values.

eps

Numeric slop allowed beyond the [0, 1] bounds. Default 1e-8.

Value

TRUE if the column looks already normalised, else FALSE.

Details

Returns TRUE only when all three hold:

  1. Every finite value lies in [0, 1] (within eps).

  2. The pooled values actually use the unit scale: pooled minimum at or below 0.05 and pooled maximum at or above 0.95.

  3. Tokens individually cover the interval: the median per-token span (max - min) is at least 0.9.

Condition 3 is the load-bearing one: raw time in seconds for citation tones (durations of, say, 0.2–0.9 s) can satisfy the first two, but its per-token spans equal the (variable, well under 0.9) durations, so it is correctly treated as unnormalised. Millisecond-scale time fails condition 1 outright, as do sequential landmark axes (<tier>_tseq, which run from 0 to the number of segments).

Two limits are worth knowing, both following from the fact that no value-based test can separate these cases:

  • Not detected: a set in which every token is a partial span of the proportional axis (e.g. a vowel-only subset where each token covers [0.3, 0.7]) looks exactly like ordinary variable-duration time, so it is rescaled per token as before. Pass time_normalised = "yes" to resolve_time_norm() (or to the fitters) to keep such an axis intact.

  • Detected: raw seconds in which every token happens to last just under one second, with a median duration of 0.9 s or more, satisfies all three conditions and is used as-is. Pass time_normalised = "no" to force the per-token rescale for such data.

See also

resolve_time_norm(), which applies this decision.