VCM

The original code to this VCM is to be found here.

The code was cleaned and tested by William N. Havard (07&10/01/22).

Installation

Clone this repository. for example if you use ssh:

git clone git@github.com:LAAC-LSCP/vcm.git

If you which to install the dependencies directly you can run:

pip install -r requirements.txt

If you want to keep everything in a conda environment:

conda create -p /my/environment/vcm pip
conda activate /my/environment/vcm
pip install -r requirements.txt

You will need the SMILExtract binary file to run vcm, download it for example in you vcm directory:

wget https://github.com/georgepar/opensmile/blob/master/bin/linux_x64_standalone_static/SMILExtract?raw=true -O SMILExtract
chmod u+x SMILExtract

Run

If you use conda, remember to check that your environment is activated

conda activate /my/environment/vcm
(vcm) myname@mycomputer$

Vcm is used with the vcm.sh script:

usage: vcm.py [-h] -a INPUT_AUDIO_PATH -r INPUT_RTTM_PATH -s
              SMILEXTRACT_BIN_PATH [-o OUTPUT_VCM_PATH]
              [-x AUDIO_EXTENSION] [--all-children] [--remove-others]
              [--from-batched-vtc] [--keep-temp] [--reuse-temp]
              [--temp-dir TEMP_DIR] [--skip-done] [-j N_JOBS]

optional arguments:
  -h, --help            show this help message and exit
  
  -a INPUT_AUDIO_PATH, --input-audio-path INPUT_AUDIO_PATH
                        Path to the audio file to be processed.
  -r INPUT_RTTM_PATH, --input-rttm-path INPUT_RTTM_PATH
                        Path to the VTC output of the file to be processed.
  -s SMILEXTRACT_BIN_PATH, --smilextract-bin-path SMILEXTRACT_BIN_PATH
                        Path to smilextract SMILExtract (v2.3) binary.
  -o OUTPUT_VCM_PATH, --output-vcm-path OUTPUT_VCM_PATH
                        Output path were the results of the VCM should be
                        stored. Default: Same as RTTM file.
  -x AUDIO_EXTENSION, --audio-extension AUDIO_EXTENSION
                        Audio files file extension (no extension '' also
                        accepted). Default: '.wav'
                        
  --all-children        Should speech segment produced by other children
                        than the key child (KCHI) be analysed. (Default:
                        False.)
  --remove-others       Should the VTC annotations for the other speakers
                        be removed from the VCMoutput file. If Segments
                        from speaker-type SPEECH, MAL, FEM, etc. will be
                        removed. (Default: False.)
  --from-batched-vtc    Whether the VTC files were generated using
                        LSCP/LAAC batch-voice-type-classifier or not./!\
                        LSCP/LAAC specific, you shouldn't be needing this
                        option. (Default: False.)
                        
  --keep-temp           Whether temporary file should be kept or not.
                        (Default: False.)
  --reuse-temp          Whether temporary file should be reused instead of
                        being recomputed. (Default: False.)
  --temp-dir TEMP_DIR   Set path to temporary directory. (Default:
                        `../tmp`).
  --skip-done           Whether RTTM for which a VCM file already exists
                        should be skipped. (Default: False.)
                        
  -j N_JOBS, -J N_JOBS, --n-jobs N_JOBS
                        Number of parallel jobs to run.

Launch you computation:

./vcm.sh -a audio/file/or/directory -r rttm/file/or/directory -s path/SMILExtract -o output/path -j 8

Tests

To test the installation, run the following command:

pytest tst --smilextract-bin-path=/scratch2/whavard/PACKAGES/opensmile/bin/linux_x64_standalone_static/SMILExtract

The original code was written in Python 2 and used an unknown version of Pytorch (presumably 0.3.0). The code however runs seemlessly with Python 3.9 and Pytorch 1.11. requirements.txt is only given for reproducibility purposes, and packages with lower version numbers might work as well.

Notes

• example.rttm was extracted using Marvin's VTC as packaged by ALICE
• Feature extraction is done using OpenSMILE v2.3. Code was run and tested with the precompile version: opensmile/bin)/linux_x64_standalone_static/
• LSCP/LAAC specific: To process RTTM files generated with the batched version of the voice-type-classificier use the following option --from-batched-vtc.

References

• Extracted features (88 eGeMAPS) reference: "The Geneva Minimalistic Acoustic Parameter Set (GeMAPS) for Voice Research and Affective Computing". Eyben et al., 2015.
• VCM reference: "VCMNet: Weakly Supervised Learning for Automatic Infant Vocalisation Maturity Analysis" . Futaisi et al., 2019. DOI: 10.1145/3340555.3353751
• VTC code ( Lavechin et al.)
• ALICE Code (Räsänen et al.)