MULAN-LRE

This version of MULAN repository for extracting attribute features from raw audio files from LRE15 dataset. These articulatory attribute features (manner and place) are high-level speech descriptive features. More information you can find in the article below

Citation

@inproceedings{DBLP:conf/slt/KukanovHSL16,
  author    = {Ivan Kukanov and Ville Hautam{\"{a}}ki and
               Sabato Marco Siniscalchi and Kehuang Li},
  title     = {Deep learning with maximal figure-of-merit cost to advance multi-label
               speech attribute detection},
  booktitle = {2016 {IEEE} Spoken Language Technology Workshop, {SLT} 2016, San Diego,
               CA, USA, December 13-16, 2016},
  pages     = {489--495},
  year      = {2016},
  doi       = {10.1109/SLT.2016.7846308}
}

How to run?

1. fix path variable KALDI_ROOT in path.sh pointing to your installed Kaldi toolkit
2. be sure that all your bash files are runnable, fixing: run from the project folder chmod -R +x ./
3. fix paths dataset_dir, scan_sub_dir and out_dir in the scripts run_cnn.sh and run_dbn.sh

• dataset_dir: path to the LRE-15 dataset
• scan_sub_dir: subforlders to search wave files with the pcm extention
• out_dir: output path, where audio lists, fbank features and result attribute scores will be saved there

After that, you can run

$ ./run_dbn.sh 0 # 0 is the processing stage, from 0-2

or

$ ./run_cnn.sh 0 # 0 is the processing stage, from 0-3

Manner attribute scores will be saved in $out_dir/res/manner/scores.txt and place attributes in $out_dir/res/place/scores.txt in the next format:

utterance_id [ columns with attributes scores per each frame]

Columns in scores.txt correspond to the next type of attributes (you can find in data/dict/):

manner: [ fricative glides nasal other silence stop voiced vowel ]

place: [ coronal dental glottal high labial low mid other palatal silence velar ]