reapeR is an R package providing access to David Talkin’s Ropust
Epoch and Pitch EstimatoR (REAPER)
algorithm. REAPER is a C++ library
that provides precise pitch tracking and epoch estimation from speech
data. reapeR provides access to relevant parts of the algorithm, with
functions for running the pitch and epoch estimation algorithm (in bulk
or on a single sound file), and for converting REAPER output into files
that can be used with Praat and
EMU.
If you run into any problems with the package, I’m happy to hear about them.
reapeR is available on
CRAN and
can be installed in the usual way:
install.packages('reapeR')
You can now track pitch and estimate epochs using the reaper()
function. In the simplest use case, you just pass a single argument with
the name of a sound file.
library(reapeR)
reaper_out <- reaper('inst/extdata/1.wav')
The resulting object, reaper_out, is a list containing two elements
called pitch and epochs.
class(reaper_out)
#> [1] "list"
names(reaper_out)
#> [1] "pitch" "epochs"
pitch is a tibble with columns including information about the time of
each analysis frame, a column with information about whether the frame
is voiced, a column with the predicted F0, and a column
with the file name.
head(reaper_out$pitch)
#> time voiced f0 file
#> 1 0.005 0 NA inst/extdata/1.wav
#> 2 0.010 0 NA inst/extdata/1.wav
#> 3 0.015 0 NA inst/extdata/1.wav
#> 4 0.020 0 NA inst/extdata/1.wav
#> 5 0.025 0 NA inst/extdata/1.wav
#> 6 0.030 0 NA inst/extdata/1.wav
The predicted pitch track looks like this (zooming into the middle of the file where there is voicing):
plot(reaper_out$pitch$time, reaper_out$pitch$f0, type = 'l',
xlab = 'Time (s)', ylab = 'F0 (Hz)',
xlim = c(0.5, 1))
epochs is just a named vector of values telling you the locations of
glottal closure instants:
reaper_out$epochs
#> $`inst/extdata/1.wav`
#> [1] 0.6515419 0.6579138 0.6646485 0.6714966 0.6783447 0.6852154 0.6920408
#> [8] 0.6988435 0.7056463 0.7123809 0.7195011 0.7258730 0.7332426 0.7400454
#> [15] 0.7468254 0.7535601 0.7602494 0.7669161 0.7736055 0.7803175 0.7870975
#> [22] 0.7937868 0.8005896 0.8074149 0.8143311 0.8212925 0.8283900 0.8355102
#> [29] 0.8428345 0.8502721 0.8578911 0.8656462 0.8738095 0.8820862 0.8905669
#> [36] 0.8992290 0.9082086 0.9176871
Here they are plotted along with the sound file in question:
snd <- tuneR::readWave('inst/extdata/1.wav')
times <- seq(0, length(snd@left) / snd@samp.rate,
length.out = length(snd@left))
plot(times, snd@left, type = 'l', xlab = 'Time (s)', ylab = '', yaxt = 'n',
xlim = c(0.65, 0.95))
abline(v = reaper_out$epochs[[1]], col='lightgrey')
box()
If you want just a data frame with pitch values, or just a vector of
glottal closure instants, this can be controlled with the output
argument. You also have access to a number of other potentially useful
values estimated by REAPER. For example, the output gci_cand gives you
access to the normalized cross-correlation at each candidate glottal
closure instant, which could be one useful measure of voicing
probability:
gci_cand <- reaper('inst/extdata/1.wav', output = 'gci_cand')
graphics::par(mar = c(0,2,0,2), mfrow = c(2,1),
oma = c(5,4,3,4))
plot(times, snd@left, type = 'l', xlab = 'Time (s)', ylab = '', yaxt = 'n', xaxt = 'n')
plot(gci_cand$time, gci_cand$nccf, type = 'l')
mtext('CCF (norm.)', side=2, line=3.5, cex=0.8)
mtext('Time (s)', side=1, line=3, outer=T, cex=0.8)
The output probs gives a number of other measures related to voicing
probability.
probs <- reaper('inst/extdata/1.wav', output = 'probs')
This includes frame-wise probabilities of voicing onset an voicing offset:
graphics::par(mar = c(0,2,0,2), mfrow = c(2,1),
oma = c(5,4,3,4))
plot(times, snd@left, type = 'l', xlab = 'Time (s)', ylab = '', yaxt = 'n', xaxt = 'n', xlim = c(0.55, 1.1))
plot(probs$time, probs$voice_onset_prob, type = 'l',
xlim = c(0.55, 1.1))
mtext('Probability of voicing onset', side=2, line=3.5, cex=0.8)
mtext('Time (s)', side=1, line=3, outer=T, cex=0.8)
graphics::par(mar = c(0,2,0,2), mfrow = c(2,1),
oma = c(5,4,3,4))
plot(times, snd@left, type = 'l', xlab = 'Time (s)', ylab = '', yaxt = 'n', xaxt = 'n', xlim = c(0.55, 1.1))
plot(probs$time, probs$voice_offset_prob, type = 'l',
xlim = c(0.55, 1.1))
mtext('Probability of voicing offset', side=2, line=3.5, cex=0.8)
mtext('Time (s)', side=1, line=3, outer=T, cex=0.8)
This output also includes REAPER’s “pseudo-probability of voicing” measure, which corresponds to normalized bandpassed RMS amplitude:
graphics::par(mar = c(0,2,0,2), mfrow = c(2,1),
oma = c(5,4,3,4))
plot(times, snd@left, type = 'l', xlab = 'Time (s)', ylab = '', yaxt = 'n', xaxt = 'n')
plot(probs$time, probs$prob_voiced, type = 'l')
mtext('Pseudo-probability of voicing', side=2, line=3.5, cex=0.8)
mtext('Time (s)', side=1, line=3, outer=T, cex=0.8)
For debugging, you can also get the resids output, containing things
like filtering residuals.
reaper() accepts arguments like start and end times, channel to be
used for analysis, pitch floor and ceiling, analysis interval etc. These
arguments can be inspected by calling ?reaper from the console.
If you want to run REAPER on all files in a directory, you can do so
with the reaper_bulk() function, where you pass the path to a
directory instead of the path to a sound file:
reaper_output <- reaper_bulk('inst/extdata')
reaper_bulk() has a few more interesting options.
There’s the argument hirst2pass, which can be set to TRUE to use the
Hirst/De Looze two-pass procedure to estimate suitable pitch floor and
ceiling values. In this case, REAPER is run with very liberal floor and
ceiling values on the directory, and the resulting pitch is used to
estimate more suitable values for a second pass.
Additionally, the praat_output argument can be set to TRUE, in which
case the REAPER estimated pitch is also saved as .Pitch files that can
be read in Praat (to a directory specified with praat_output_dir).
If you already have REAPER files lying around, the library also comes with several functions to work with REAPER output files.
read_pitch_out() will read an ASCII-formatted REAPER pitch output file
(saved with -f filename -a) as a well-formatted tibble.
pitchOut <- read_pitch_out('inst/extdata/pitch')
head(pitchOut)
#> # A tibble: 6 × 3
#> time voiced f0
#> <dbl> <dbl> <dbl>
#> 1 0.005 0 NA
#> 2 0.01 0 NA
#> 3 0.015 0 NA
#> 4 0.02 0 NA
#> 5 0.025 0 NA
#> 6 0.03 0 NA
read_epochs_out() will read an ASCII-formatted REAPER epochs
(‘pitchmarks’) output file (saved with -p filename -a) as a vector.
epochsOut <- read_epochs_out('inst/extdata/epochs')
epochsOut
#> [1] 0.651542 0.657914 0.664649 0.671497 0.678345 0.685215 0.692041 0.698844
#> [9] 0.705646 0.712381 0.719501 0.725873 0.733243 0.740045 0.746825 0.753560
#> [17] 0.760249 0.766916 0.773605 0.780317 0.787098 0.793787 0.800590 0.807415
#> [25] 0.814331 0.821293 0.828390 0.835510 0.842834 0.850272 0.857891 0.865646
#> [33] 0.873810 0.882086 0.890567 0.899229 0.908209 0.917687
If you’re interested in working with this output in Praat, the
write_praat_pitch() function will save a tibble with pitch output as a
.Pitch file, and the write_praat_epochs() function will save a
vector with epoch output as a .PointProcess file.
If you want to import REAPER-estimated pitch into an EMU
database, you can import a
tibble with pitch output to a database using the reaper2emuDB()
function.