Skip to contents

Detect voiced frames using Drugman's multi-branch VAD

trk_covarep_vad_drugman.Rd

Estimates per-frame speech activity as posterior probabilities using three independent ANN classifiers (MFCC-based, Sadjadi pitch-related, and CPP/SRH features) combined by geometric mean (Drugman et al. 2012) . Suitable for pre-filtering frames before pitch or voice quality analysis.

Usage

trk_covarep_vad_drugman(listOfFiles, beginTime = 0, endTime = 0, toFile = FALSE, explicitExt = "cvd", outputDirectory = NULL, verbose = TRUE)

Arguments

listOfFiles

Character vector of audio file paths. Any format supported by av is accepted; non-native inputs are transcoded automatically.

toFile

Logical. If TRUE, write SSFF output files and return the paths written invisibly. If FALSE, return an AsspDataObj. Default FALSE.

explicitExt

Character. Output file extension. Default "cvd".

beginTime

Start time for the extracted portion in seconds. Default: NULL (beginning of signal). Note: uses beginTime/endTime (seconds) matching DSP function conventions, unlike read_audio() which uses begin/end.

endTime

The end time of the section of the sound files that should be analysed (in seconds). Use 0 for end of file.

outputDirectory

The directory where the slice file should be stored. If not defiled (NULL), the sparse slice file will placed in the same folder as the media file.

verbose

Logical. Show a progress bar (sequential path) or a progress-aware parallel apply (pbapply/pbmcapply, if installed).

Value

If toFile = FALSE: an AsspDataObj with tracks:

vad_final

FLOAT, ensemble voicing posterior (geometric mean of three branches), 0–1, n_frames × 1. Recommended for downstream use.

vad_mfcc

FLOAT, MFCC-branch posterior, 0–1, n_frames × 1.

vad_sadjadi

FLOAT, Sadjadi pitch-feature posterior, 0–1, n_frames × 1.

vad_new

FLOAT, CPP/SRH-feature posterior, 0–1, n_frames × 1.

Frame rate: 200 Hz (5 ms hop, interpolated from 10 ms internal hop). If toFile = TRUE: character vector of output file paths, returned invisibly.

Details

Audio is resampled to 16 kHz internally. A threshold of 0.5 on vad_final gives a reasonable binary voiced/unvoiced decision; 0.7 is more conservative. Each branch applies an 11-frame median filter to the posterior before combination.

References

Drugman T, Soria-Olivas E, Perez-Córdoba JL, Alwan A (2012). “A comparative study of different feature sets for acoustic voice quality assessment.” IEEE Transactions on Audio, Speech, and Language Processing, 20(6), 1690–1703. doi:10.1109/TASL.2012.2188377 . Multi-branch voice activity detection using MFCC and Sadjadi features.

Examples

if (FALSE) { # \dontrun{
trk_covarep_vad_drugman(
  system.file("samples", "sustained", "a1.wav", package = "superassp"),
  toFile = FALSE
)
} # }