optimizePars: Optimize parameters for acoustic analysis

In soundgen: Sound Synthesis and Acoustic Analysis

Optimize parameters for acoustic analysis

Description

This customized wrapper for optim attempts to optimize the parameters of segment or analyze by comparing the results with a manually annotated "key". This optimization function uses a single measurement per audio file (e.g., median pitch or the number of syllables). For other purposes, you may want to adapt the optimization function so that the key specifies the exact timing of syllables, their median length, frame-by-frame pitch values, or any other characteristic that you want to optimize for. The general idea remains the same, however: we want to tune function parameters to fit our type of audio and research priorities. The default settings of segment and analyze have been optimized for human non-linguistic vocalizations.

Usage

optimizePars(
  myfolder,
  key,
  myfun,
  pars,
  bounds = NULL,
  fitnessPar,
  fitnessFun = function(x) 1 - cor(x, key, use = "pairwise.complete.obs"),
  nIter = 10,
  init = NULL,
  initSD = 0.2,
  control = list(maxit = 50, reltol = 0.01, trace = 0),
  otherPars = list(plot = FALSE),
  mygrid = NULL,
  verbose = TRUE
)

Arguments

myfolder	path to where the .wav files live
key	a vector containing the "correct" measurement that we are aiming to reproduce
myfun	the function being optimized: either 'segment' or 'analyze' (in quotes)
pars	names of arguments to myfun that should be optimized
bounds	a list setting the lower and upper boundaries for possible values of optimized parameters. For ex., if we optimize smooth and smoothOverlap, reasonable bounds might be list(low = c(5, 0), high = c(500, 95))
fitnessPar	the name of output variable that we are comparing with the key, e.g. 'nBursts' or 'pitch_median'
fitnessFun	the function used to evaluate how well the output of myfun fits the key. Defaults to 1 - Pearson's correlation (i.e. 0 is perfect fit, 1 is awful fit). For pitch, log scale is more meaningful, so a good fitness criterion is "function(x) 1 - cor(log(x), log(key), use = 'pairwise.complete.obs')"
nIter	repeat the optimization several times to check convergence
init	initial values of optimized parameters (if NULL, the default values are taken from the definition of myfun)
initSD	each optimization begins with a random seed, and initSD specifies the SD of normal distribution used to generate random deviation of initial values from the defaults
control	a list of control parameters passed on to optim. The method used is "Nelder-Mead"
otherPars	a list of additional arguments to myfun
mygrid	a dataframe with one column per parameter to optimize, with each row specifying the values to try. If not NULL, optimizePars simply evaluates each combination of parameter values, without calling optim (see examples)
verbose	if TRUE, reports the values of parameters evaluated and fitness

Details

If your sounds are very different from human non-linguistic vocalizations, you may want to change the default values of other arguments to speed up convergence. Adapt the code to enforce suitable constraints, depending on your data.

Value

Returns a matrix with one row per iteration with fitness in the first column and the best values of each of the optimized parameters in the remaining columns.

Examples

## Not run: 
# Download 260 sounds from the supplements in Anikin & Persson (2017)
# - see http://cogsci.se/publications.html
# Unzip them into a folder, say '~/Downloads/temp'
myfolder = '~/Downloads/temp260'  # 260 .wav files live here

# Optimization of SEGMENTATION
# Import manual counts of syllables in 260 sounds from
# Anikin & Persson (2017) (our "key")
key = segmentManual  # a vector of 260 integers

# Run optimization loop several times with random initial values
# to check convergence
# NB: with 260 sounds and default settings, this might take ~20 min per iteration!
res = optimizePars(myfolder = myfolder, myfun = 'segment', key = key,
  pars = c('shortestSyl', 'shortestPause'),
  fitnessPar = 'nBursts', otherPars = list(method = 'env'),
  nIter = 3, control = list(maxit = 50, reltol = .01, trace = 0))

# Examine the results
print(res)
for (c in 2:ncol(res)) {
  plot(res[, c], res[, 1], main = colnames(res)[c])
}
pars = as.list(res[1, 2:ncol(res)])  # top candidate (best pars)
s = do.call(segment, c(myfolder, pars))  # segment with best pars
cor(key, as.numeric(s[, fitnessPar]))
boxplot(as.numeric(s[, fitnessPar]) ~ as.integer(key), xlab='key')
abline(a=0, b=1, col='red')

# Try a grid with particular parameter values instead of formal optimization
res = optimizePars(myfolder = myfolder, myfun = 'segment', key = segmentManual,
  pars = c('shortestSyl', 'shortestPause'),
  fitnessPar = 'nBursts', otherPars = list(method = 'env'),
  mygrid = expand.grid(shortestSyl = c(30, 40),
                       shortestPause = c(30, 40, 50)))
1 - res$fit  # correlations with key

# Optimization of PITCH TRACKING (takes several hours!)
key = as.numeric(log(pitchManual))
res = optimizePars(
  myfolder = myfolder,
    myfun = 'analyze',
    key = key,  # log-scale better for pitch
    pars = c('windowLength', 'silence'),
    bounds = list(low = c(5, 0), high = c(200, .2)),
    fitnessPar = 'pitch_median',
    nIter = 2,
    otherPars = list(plot = FALSE, loudness = NULL, novelty = NULL,
                     roughness = NULL, nFormants = 0),
    fitnessFun = function(x) {
      1 - cor(log(x), key, use = 'pairwise.complete.obs') *
      (1 - mean(is.na(x) & is.finite(key))) # penalize failing to detect f0
})

## End(Not run)

soundgen documentation built on Sept. 12, 2024, 6:29 a.m.