optimize_energy_detector {ohun}

Optimize energy-based sound event detection

Description

Optimize energy-based sound event detection under different correlation threshold values

Usage

optimize_energy_detector(reference, files = NULL, threshold = 5,
peak.amplitude = 0, hop.size = 11.6, wl = NULL, smooth = 5, hold.time = 0,
min.duration = NULL, max.duration = NULL, thinning = 1, cores = 1, pb = TRUE,
 by.sound.file = FALSE, bp = NULL, path = ".", previous.output = NULL, envelopes = NULL,
 macro.average = FALSE, min.overlap = 0.5)

Arguments

reference

Selection table (using the warbleR package's format, see selection_table) or data frame with columns for sound file name (sound.files), selection number (selec), and start and end time of sound event (start and end). It must contain the reference selections that will be used for detection optimization.

files

Character vector indicating the sound files that will be analyzed. Optional. If not supplied the function will work on the sound files in 'reference'. It can be used to include sound files with no target sound events. Supported file formats:'.wav', '.mp3', '.flac' and '.wac'. If not supplied the function will work on all sound files (in the supported format) in 'path'.

threshold

A numeric vector specifying the amplitude threshold for detecting sound events (in %). Default is 5. Several values can be supplied for optimization.

peak.amplitude

Numeric vector of length 1 with the minimum peak amplitude value. A detection below that value would be excluded. Peak amplitude is the maximum sound pressure level (in decibels) across the sound event (see sound_pressure_level). This can be useful when expecting higher peak amplitude in the target sound events compared to non-target sound events or when keeping only the best examples of the target sound events (i.e. high precision and low recall). Default is 0. Several values can be supplied for optimization.

hop.size

A numeric vector of length 1 specifying the time window duration (in ms). Default is 11.6 ms, which is equivalent to 512 wl for a 44.1 kHz sampling rate. Ignored if 'wl' is supplied.

wl

A numeric vector of length 1 specifying the window length of the spectrogram. Default is NULL. If supplied, 'hop.size' is ignored. Used internally for bandpass filtering (so only applied when 'bp' is supplied).

smooth

A numeric vector of length 1 to smooth the amplitude envelope with a sum smooth function. It controls the time 'neighborhood' (in ms) in which amplitude samples are smoothed (i.e. averaged with neighboring samples). Default is 5. 0 means no smoothing is applied. Note that smoothing is applied before thinning (see 'thinning' argument). The function envelope is used internally which is analogous to sum smoothing in env. This argument is used internally by get_envelopes. Several values can be supplied for optimization.

hold.time

Numeric vector of length 1. Specifies the time range (in ms) at which selections will be merged (i.e. if 2 selections are separated by less than the specified 'hold.time' they will be merged in to a single selection). Default is 0 (no hold time applied). Several values can be supplied for optimization.

min.duration

Numeric vector giving the shortest duration (in ms) of the sound events to be detected. It removes sound events below that threshold. Several values can be supplied for optimization.

max.duration

Numeric vector giving the longest duration (in ms) of the sound events to be detected. It removes sound events above that threshold. Several values can be supplied for optimization.

thinning

Numeric vector in the range 0~1 indicating the proportional reduction of the number of samples used to represent amplitude envelopes (i.e. the thinning of the envelopes). Usually amplitude envelopes have many more samples than those needed to accurately represent amplitude variation in time, which affects the size of the output (usually very large R objects / files). Default is 1 (no thinning). Higher sampling rates may afford higher size reduction (e.g. lower thinning values). Reduction is conducted by interpolation using approx. Note that thinning may decrease time precision, and the higher the thinning the less precise the time detection. Several values can be supplied for optimization.

cores

Numeric. Controls whether parallel computing is applied. It specifies the number of cores to be used. Default is 1 (i.e. no parallel computing).

pb

Logical argument to control progress bar and messages. Default is TRUE.

by.sound.file

Logical argument to control whether performance diagnostics are summarized across sound files (when by.sound.file = FALSE and more than 1 sound file is included in 'reference') or shown separated by sound file. Default is FALSE.

bp

Numeric vector of length 2 giving the lower and upper limits of a frequency bandpass filter (in kHz). Default is NULL. This argument is used internally by get_envelopes. Not used if 'envelopes' are supplied. Bandpass is done using the function ffilter, which applies a short-term Fourier transformation to first create a spectrogram in which the target frequencies are filtered and then is back transformed into a wave object using a reverse Fourier transformation.

path

Character string containing the directory path where the sound files are located. The current working directory is used as default.

previous.output

Data frame with the output of a previous run of this function. This will be used to include previous results in the new output and avoid recalculating detection performance for parameter combinations previously evaluated.

envelopes

An object of class 'envelopes' (generated by get_envelopes) containing the amplitude envelopes of the sound files to be analyzed. If 'files' and 'envelopes' are not supplied then the function will work on all supported format sound files in the working directory.

macro.average

Logical argument to control if diagnostics are first calculated for each sound file and then averaged across sound files, which can minimize the effect of unbalanced sample sizes between sound files. If FALSE (default) diagnostics are based on aggregated statistics irrespective of sound files. The following indices can be estimated by macro-averaging: overlap, mean.duration.true.positives, mean.duration.false.positives, mean.duration.false.positives, mean.duration.false.negatives, proportional.duration.true.positives, recall and precision (f.score is always derived from recall and precision). Note that when applying macro-averaging, recall and precision are not derived from the true positive, false positive and false negative values returned by the function.

min.overlap

Numeric. Controls the minimum amount of overlap required for a detection and a reference sound for it to be counted as true positive. Default is 0.5. Overlap is measured as intersection over union.

Details

This function takes a selections data frame or 'selection_table' ('reference') estimates the detection performance of a energy detector under different detection parameter combinations. This is done by comparing the position in time of the detection to those of the reference selections in 'reference'. The function returns several diagnostic metrics to allow user to determine which parameter values provide a detection that more closely matches the selections in 'reference'. Those parameters can be later used for performing a more efficient detection using energy_detector.

Value

A data frame in which each row shows the result of a detection job with a particular combination of tuning parameters (including in the data frame). It also includes the following diagnostic metrics:

• true.positives: number of sound events in 'reference' that correspond to any detection. Matching is defined as some degree of overlap in time. In a perfect detection routine it should be equal to the number of rows in 'reference'.

• false.positives: number of detections that don't match any of the sound events in 'reference'. In a perfect detection routine it should be 0.

• false.negatives: number of sound events in 'reference' that were not detected (not found in 'detection'. In a perfect detection routine it should be 0.

• splits: number of detections overlapping reference sounds that also overlap with other detections. In a perfect detection routine it should be 0.

• merges: number of detections that overlap with two or more reference sounds. In a perfect detection routine it should be 0.

• mean.duration.true.positives: mean duration of true positives (in ms). Only included when time.diagnostics = TRUE.

• mean.duration.false.positives: mean duration of false positives (in ms). Only included when time.diagnostics = TRUE.

• mean.duration.false.negatives: mean duration of false negatives (in ms). Only included when time.diagnostics = TRUE.

• overlap: mean intersection over union overlap of true positives.

• proportional.duration.true.positives: ratio of duration of true positives to th duration of sound events in 'reference'. In a perfect detection routine it should be 1. Based only on true positives that were not split or merged. Only included when time.diagnostics = TRUE.

• duty.cycle: proportion of a sound file in which sounds were detected. Only included when time.diagnostics = TRUE and path is supplied.

• recall: Proportion of sound events in 'reference' that were detected. In a perfect detection routine it should be 1.

• precision: Proportion of detections that correspond to sound events in 'reference'. In a perfect detection routine it should be 1.

Author(s)

Marcelo Araya-Salas ([e-mail]).

References

Araya-Salas, M., Smith-Vidaurre, G., Chaverri, G., Brenes, J. C., Chirino, F., Elizondo-Calvo, J., & Rico-Guevara, A. 2022. ohun: an R package for diagnosing and optimizing automatic sound event detection. BioRxiv, 2022.12.13.520253. https://doi.org/10.1101/2022.12.13.520253

Examples

# Save example files into temporary working directory
data("lbh1", "lbh2", "lbh_reference")
tuneR::writeWave(lbh1, file.path(tempdir(), "lbh1.wav"))
tuneR::writeWave(lbh2, file.path(tempdir(), "lbh2.wav"))

# using smoothing and minimum duration
optimize_energy_detector(
  reference = lbh_reference, path = tempdir(),
  threshold = c(6, 10), smooth = 6.8, bp = c(2, 9), hop.size = 6.8,
  min.duration = 90
)

# with thinning and smoothing
optimize_energy_detector(
  reference = lbh_reference, path = tempdir(),
  threshold = c(6, 10, 15), smooth = c(7, 10), thinning = c(0.1, 0.01),
  bp = c(2, 9), hop.size = 6.8, min.duration = 90
)

# by sound file
(opt_ed <- optimize_energy_detector(
  reference = lbh_reference,
  path = tempdir(), threshold = c(6, 10, 15), smooth = 6.8, bp = c(2, 9),
  hop.size = 6.8, min.duration = 90, by.sound.file = TRUE
))

# summarize
summarize_diagnostic(opt_ed)

# using hold time
(op_ed <- optimize_energy_detector(
  reference = lbh_reference,
  threshold = 10, hold.time = c(100, 150), bp = c(2, 9), hop.size = 6.8,
  path = tempdir()
))

# including previous output in new call
optimize_energy_detector(
  reference = lbh_reference, threshold = 10,
  hold.time = c(50, 200), previous.output = op_ed, smooth = 6.8,
  bp = c(2, 9), hop.size = 7, path = tempdir()
)

# having and extra file in files (simulating a file that should have no detetions)
sub_reference <- lbh_reference[lbh_reference$sound.files != "lbh1.wav", ]

optimize_energy_detector(
  reference = sub_reference, files = unique(lbh_reference$sound.files),
  threshold = 10, hold.time = c(1, 150), bp = c(2, 9), smooth = 6.8,
  hop.size = 7, path = tempdir()
)