compareSounds: Compare two sounds

In soundgen: Sound Synthesis and Acoustic Analysis

Compare two sounds

Description

Computes similarity between two sounds based on comparing their spectrogram-like representations. If the input is audio, two methods of producing spectrograms are available: specType = 'linear' calls powspec for an power spectrogram with frequencies in Hz, and specType = 'mel' calls melfcc for an auditory spectrogram with frequencies in Mel. For more customized options, just produce your spectrograms or feature matrices (time in column, features like pitch, peak frequency etc in rows) with your favorite function before calling compareSounds because it also accepts matrices as input. To be directly comparable, the two matrices are made into matrices of the same size. In case of differences in sampling rates, only frequencies below the lower Nyquist frequency or below maxFreq are kept. In case of differences in duration, the shorter sound is padded with 0 (silence) or NA, as controlled by arguments padWith, padDir. Then the matrices are compared using methods like cross-correlation or Dynamic Time Warp.

Usage

compareSounds(
  x,
  y,
  samplingRate = NULL,
  windowLength = 40,
  overlap = 50,
  step = NULL,
  dynamicRange = 80,
  method = c("cor", "cosine", "diff", "dtw"),
  specType = c("linear", "mel")[2],
  specPars = list(),
  dtwPars = list(),
  padWith = NA,
  padDir = c("central", "left", "right")[1],
  maxFreq = NULL
)

Arguments

x, y	either two matrices (spectrograms or feature matrices) or two sounds to be compared (numeric vectors, Wave objects, or paths to wav/mp3 files)
samplingRate	if one or both inputs are numeric vectors, specify sampling rate, Hz. A vector of length 2 means the two inputs have different sampling rates, in which case spectrograms are compared only up to the lower Nyquist frequency
windowLength	length of FFT window, ms
overlap	overlap between successive FFT frames, %
step	you can override overlap by specifying FFT step, ms (NB: because digital audio is sampled at discrete time intervals of 1/samplingRate, the actual step and thus the time stamps of STFT frames may be slightly different, eg 24.98866 instead of 25.0 ms)
dynamicRange	parts of the spectra quieter than -dynamicRange dB are not compared
method	method of comparing mel-transformed spectra of two sounds: "cor" = Pearson's correlation; "cosine" = cosine similarity; "diff" = absolute difference between each bin in the two spectrograms; "dtw" = multivariate Dynamic Time Warp with dtw
specType	"linear" = power spectrogram with powspec, "mel" = mel-frequency spectrogram with melfcc
specPars	a list of parameters passed to melfcc
dtwPars	a list of parameters passed to dtw
padWith	if the duration of x and y is not identical, the compared spectrograms are padded with either silence (padWith = 0) or with NA's (padWith = NA) to have the same number of columns. Padding with NA implies that only the overlapping part is of relevance, whereas padding with 0 means that the added silent part is also compared with the longer sound, usually resulting in lower similarity (see examples)
padDir	if padding, specify where to add zeros or NAs: before the sound ('left'), after the sound ('right'), or on both sides ('central')
maxFreq	parts of the spectra above maxFreq Hz are not compared

Value

Returns a dataframe with two columns: "method" for the method(s) used, and "sim" for the similarity between the two sounds calculated with that method. The range of similarity measures is [-1, 1] for "cor", [0, 1] for "cosine" and "diff", and (-Inf, Inf) for "dtw".

Examples

data(orni, peewit, package = 'seewave')
compareSounds(orni, peewit)
# spectrogram(orni); playme(orni)
# spectrogram(peewit); playme(peewit)

## Not run: 
s1 = soundgen(formants = 'a', play = TRUE)
s2 = soundgen(formants = 'ae', play = TRUE)
s3 = soundgen(formants = 'eae', sylLen = 700, play = TRUE)
s4 = runif(8000, -1, 1)  # white noise
compareSounds(s1, s2, samplingRate = 16000)
compareSounds(s1, s4, samplingRate = 16000)

# the central section of s3 is more similar to s1 than is the beg/eng of s3
compareSounds(s1, s3, samplingRate = 16000, padDir = 'left')
compareSounds(s1, s3, samplingRate = 16000, padDir = 'central')

# padding with 0 penalizes differences in duration, whereas padding with NA
# is like saying we only care about the overlapping part
compareSounds(s1, s3, samplingRate = 16000, padWith = 0)
compareSounds(s1, s3, samplingRate = 16000, padWith = NA)

# comparing linear (Hz) vs mel-spectrograms produces quite different results
compareSounds(s1, s3, samplingRate = 16000, specType = 'linear')
compareSounds(s1, s3, samplingRate = 16000, specType = 'mel')

# pass additional control parameters to dtw and melfcc
compareSounds(s1, s3, samplingRate = 16000,
              specPars = list(nbands = 128),
              dtwPars = list(dist.method = "Manhattan"))

# use feature matrices instead of spectrograms (time in columns, features in rows)
a1 = t(as.matrix(analyze(s1, samplingRate = 16000)$detailed))
a1 = a1[4:nrow(a1), ]; a1[is.na(a1)] = 0
a2 = t(as.matrix(analyze(s2, samplingRate = 16000)$detailed))
a2 = a2[4:nrow(a2), ]; a2[is.na(a2)] = 0
a4 = t(as.matrix(analyze(s4, samplingRate = 16000)$detailed))
a4 = a4[4:nrow(a4), ]; a4[is.na(a4)] = 0
compareSounds(a1, a2, method = c('cosine', 'dtw'))
compareSounds(a1, a4, method = c('cosine', 'dtw'))

# a demo for comparing different similarity metrics
target = soundgen(sylLen = 500, formants = 'a',
                  pitch = data.frame(time = c(0, 0.1, 0.9, 1),
                                     value = c(100, 150, 135, 100)),
                  temperature = 0.001)
spec1 = soundgen:::getMelSpec(target, samplingRate = 16000)

parsToTry = list(
  list(formants = 'i',                                            # wrong
       pitch = data.frame(time = c(0, 1),                         # wrong
                          value = c(200, 300))),
  list(formants = 'i',                                            # wrong
       pitch = data.frame(time = c(0, 0.1, 0.9, 1),               # right
                                 value = c(100, 150, 135, 100))),
  list(formants = 'a',                                            # right
       pitch = data.frame(time = c(0,1),                          # wrong
                                 value = c(200, 300))),
  list(formants = 'a',
       pitch = data.frame(time = c(0, 0.1, 0.9, 1),               # right
                                 value = c(100, 150, 135, 100)))  # right
)

sounds = list()
for (s in 1:length(parsToTry)) {
  sounds[[length(sounds) + 1]] =  do.call(soundgen,
    c(parsToTry[[s]], list(temperature = 0.001, sylLen = 500)))
}
lapply(sounds, playme)

method = c('cor', 'cosine', 'diff', 'dtw')
df = matrix(NA, nrow = length(parsToTry), ncol = length(method))
colnames(df) = method
df = as.data.frame(df)
for (i in 1:nrow(df)) {
  df[i, ] = compareSounds(
    x = spec1,  # faster to calculate spec1 once
    y = sounds[[i]],
    samplingRate = 16000,
    method = method
  )[, 2]
}
df$av = rowMeans(df, na.rm = TRUE)
# row 1 = wrong pitch & formants, ..., row 4 = right pitch & formants
df$formants = c('wrong', 'wrong', 'right', 'right')
df$pitch = c('wrong', 'right', 'wrong', 'right')
df

## End(Not run)

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