Trim each token to its voiced region

Keeps only the rows between a token's first and last voiced frame, so a downstream measurement spans the syllable rather than the whole recording. Frame-based trackers emit a frame every step across the entire file, with NA (or 0 Hz) f0 in silence; without trimming, "21 equidistant points across the token" spends part of the grid on leading and trailing silence. Sparse input such as a .PitchTier has no silent rows to trim, so this is a no-op there.

Usage

trim_to_voiced(df, token = "token", time = "time", f0 = "f0", min_run = 2)

Arguments

df

Long-format f0 data frame.

token, time, f0

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

min_run

Minimum number of consecutive voiced frames an edge anchor must belong to. Default 2.

Value

df with out-of-region rows removed, row order preserved.

Details

The edges are anchored on a run of at least min_run consecutive voiced frames, so one stray voiced frame in silence (a common octave-error artefact) cannot stretch the region. If no run is that long, the first and last voiced frames are used. Tokens with no voiced frame at all keep no rows — count them before and after if their loss should be reported.

Frames inside the region are kept whether or not they are voiced, so an internal unvoiced stretch (a medial voiceless stop, a creaky patch) stays in place and still reads as NA.

See also

resample_f0_equal(), which is normally applied after this; filter_interval_rows() for a TextGrid-interval region instead.