This repository provides scripts required for training "Boosted Locality Sensitive Hash (BLSH)" functions and the kNN-based source separation algorithm. The BLSH method was introduced in [1].
- Helper functions to load TIMIT speakers and Duan and DEMAND noise datasets.
- A python script to generate training and testing wavefiles and spectral features.
- A python script to check performance of oracle or projected kNN on various features.
- A python script to train weak learners on various features.
- Helper functions.
Make sure that pip and python3 are installed (The program was written using Python 3.6) and install the script's dependencies. Note: Librosa is used for audio reading and writing but can be replaced with other packages such as scipy.signal. Matplotlib can be removed if not plotting the results.
- To generate the clean source, noises, and mixture wavefiles and the spectral features,
python generate_data.py --make_wavefiles --option
where options are: None for STFT, --use_mel for mel spectrograms, --use_mfcc for MFCC
- To construct hash functions in the form of weak learners,
python train_weak_learners.py --use_stft
- For other options,
python train_weak_learners.py -h
- Eg. to test the performance of kNN procedure using ground truth STFT dictionary on closed set,
python check_performance.py -o --use_stft --is_closed
- For other options,
python check_performance.py -h
- TIMIT (https://catalog.ldc.upenn.edu/LDC93S1)
- Duan (http://www2.ece.rochester.edu/~zduan/data/noise/)
- DEMAND (https://zenodo.org/record/1227121#.Xbm4X797leg)
- S. Kim, H. Yang, and M. Kim. "Boosted Locality Sensitive Hashing: Discriminative Binary Codes for Source Separation" (https://arxiv.org/abs/2002.06239).
This material is based upon work supported by the National Science Foundation under Award Number:1909509.