Deep Learning For Monaural Source Separation

Demo

Webpage: https://sites.google.com/site/deeplearningsourceseparation/

Experiments

MIR-1K experiment (singing voice separation)

1. Training code: codes/mir1k/train_mir1k_demo.m

2. Demo

• Download a trained model http://www.ifp.illinois.edu/~huang146/DNN_separation/model_400.mat
• Put the model at codes/mir1k/demo and go to the folder
• Run: codes/mir1k/demo/run_test_single_model.m

TIMIT experiment (speech separation)

1. Training code: codes/timit/train_timit_demo.m and codes/timit/train_timit_demo_mini_clip.m

2. Demo

• Download a trained model http://www.ifp.illinois.edu/~huang146/DNN_separation/timit_model_70.mat
• Put the model at codes/timit/demo and go to the folder
• Run: codes/timit/demo/run_test_single_model.m

TSP experiment (speech separation)

1. Training code: codes/TSP/train_TSP_demo_mini_clip.m

2. Demo

• Download a trained model http://www.ifp.illinois.edu/~huang146/DNN_separation/TSP_model_RNN1_win1_h300_l2_r0_64ms_1000000_softabs_linearout_RELU_logmel_trn0_c1e-10_c0.001_bsz100000_miter10_bf50_c0_d0_7650.mat
• Put the model at codes/TSP/demo and go to the folder
• Run the demo code at codes/TSP/demo/run_test_single_model.m

Denosing experiment

1. Put original FCJF0, FDAW0', FDML0, FECD0, 'FETB0', 'FJSP0', 'FKFB0', 'FMEM0', 'FSAH0', 'FSJK1', 'FSMA0', 'FTBR0', 'FVFB0' 'FVMH0 of the original TIMIT data under codes/denoising/Data/timit/

2. Training code: codes/denoising/train_denoising_demo.m

3. Demo

• Download a trained model http://www.ifp.illinois.edu/~huang146/DNN_separation/denoising_model_870.mat
• Put the model at codes/denoising/demo and go to the folder
• Run the demo code at codes/denoising/demo/run_test_single_model.m

Dependencies

1. The package is modified based on rnn-speech-denoising

2. The software depends on Mark Schmidt's minFunc package for convex optimization.

3. Additionally, we have included Mark Hasegawa-Johnson's HTK write and read functionsthat are used to handle the MFCC files.

4. We use HTK for computing features (MFCC, logmel) (HCopy).

5. We use signal processing functions from labrosa.

6. We use BSS Eval toolbox Version 2.0, 3.0 for evaluation.

7. We use MIR-1K for singing voice separation task.

8. We use TSP for speech separation task.

Work on your data:

1. To try the codes on your data, see mir1k, TSP settings - put your data into codes/mir1k/Wavfile or codes/TSP/Data/ accordingly.

2. Look at the unit test parameters below codes/mir1k/train_mir1k_demo.m, codes/TSP/train_TSP_demo_mini_clip.m (with minibatch lbfgs, gradient clipping)

3. Tune the parameters on the dev set and check the results.

Reference

1. P.-S. Huang, M. Kim, M. Hasegawa-Johnson, P. Smaragdis, "Joint Optimization of Masks and Deep Recurrent Neural Networks for Monaural Source Separation", IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 23, no. 12, pp. 2136–2147, Dec. 2015

2. P.-S. Huang, M. Kim, M. Hasegawa-Johnson, P. Smaragdis, "Singing-Voice Separation From Monaural Recordings Using Deep Recurrent Neural Networks," in International Society for Music Information Retrieval Conference (ISMIR) 2014.

3. P.-S. Huang, M. Kim, M. Hasegawa-Johnson, P. Smaragdis, "Deep Learning for Monaural Speech Separation," in IEEE International Conference on Acoustic, Speech and Signal Processing 2014.

Notes

The codes are tested using MATLAB R2015a

Related Implementations

source_separaton_ml_jeju