文件列表:

LICENSE (11357, 2017-12-16)
audio_dataset.py (19722, 2017-12-16)
create_taskfiles.sh (841, 2017-12-16)
custom_callbacks.py (1008, 2017-12-16)
custom_layers.py (17384, 2017-12-16)
data_setup_downsample1 (0, 2017-12-16)
data_setup_downsample1\configs (0, 2017-12-16)
data_setup_downsample1\configs\params_lstm_4561bd13e267026c3f3d1c936b15f709.yaml (152, 2017-12-16)
data_setup_downsample1\configs\params_lstm_46666e232751074bd609167dc440df8c.yaml (151, 2017-12-16)
data_setup_downsample1\configs\params_lstm_6a4fc9017283c9f89380f765a60087ce.yaml (151, 2017-12-16)
data_setup_downsample1\configs\params_lstm_b6da76df68cf530d091aa499d61143de.yaml (152, 2017-12-16)
data_setup_downsample1\configs\params_unfolded_snmf_364ccd17a3e187bcccd30cfaa6bd9422.yaml (223, 2017-12-16)
data_setup_downsample1\configs\params_unfolded_snmf_a23657edf96a44331501d773db837a1c.yaml (222, 2017-12-16)
data_setup_downsample1\configs\params_unfolded_snmf_a45e86a1cc146e1e9d7a7f8100d9d2d7.yaml (220, 2017-12-16)
data_setup_downsample1\configs\params_unfolded_snmf_ea1e7d485421e527486476ef696da2da.yaml (221, 2017-12-16)
data_setup_downsample1\params_data.yaml (509, 2017-12-16)
data_setup_downsample10 (0, 2017-12-16)
data_setup_downsample10\configs (0, 2017-12-16)
data_setup_downsample10\configs\params_lstm_4561bd13e267026c3f3d1c936b15f709.yaml (152, 2017-12-16)
data_setup_downsample10\configs\params_lstm_46666e232751074bd609167dc440df8c.yaml (151, 2017-12-16)
data_setup_downsample10\configs\params_lstm_6a4fc9017283c9f89380f765a60087ce.yaml (151, 2017-12-16)
data_setup_downsample10\configs\params_lstm_b6da76df68cf530d091aa499d61143de.yaml (152, 2017-12-16)
data_setup_downsample10\configs\params_unfolded_snmf_364ccd17a3e187bcccd30cfaa6bd9422.yaml (223, 2017-12-16)
data_setup_downsample10\configs\params_unfolded_snmf_a23657edf96a44331501d773db837a1c.yaml (222, 2017-12-16)
data_setup_downsample10\configs\params_unfolded_snmf_a45e86a1cc146e1e9d7a7f8100d9d2d7.yaml (220, 2017-12-16)
data_setup_downsample10\configs\params_unfolded_snmf_ea1e7d485421e527486476ef696da2da.yaml (221, 2017-12-16)
data_setup_downsample10\params_data.yaml (511, 2017-12-16)
download_toolboxes.sh (775, 2017-12-16)
enhance.py (63959, 2017-12-16)
plot_learning_curves_waspaa2017.ipynb (254483, 2017-12-16)
print_scores.py (13178, 2017-12-16)
run_waspaa2017.sh (2219, 2017-12-16)
score_audio.m (6306, 2017-12-16)
score_audio.sh (1317, 2017-12-16)
snmf.py (4814, 2017-12-16)
sparseNMF (0, 2017-12-16)
sparseNMF\sparse_nmf_exec.m (546, 2017-12-16)
sparseNMF\sparse_nmf_gpu.m (9660, 2017-12-16)
... ...

## Implementation of deep recurrent nonnegative matrix factorization (DR-NMF) for speech separation DR-NMF is a recurrent neural network constructed from the unfolded iterations of the iterative soft-thresholding algorithm (ISTA) applied to sparse NMF inference. Sparse NMF inference is the task of inferring the nonnegative sparse coefficients H given a nonnegative dictionary W such that WH approximates a nonnegative observation matrix X. For speech separation, the observation matrix X is the raw spectrogram of noisy audio, and the dictionary W is partitioned into speech and noise components. This partitioning of the dictionary W allows computation of an enhancement mask in the STFT domain. Read the paper here: https://arxiv.org/abs/1709.07124 [Presentation slides from WASPAA 2017 (received best student paper award)](https://www.dropbox.com/s/mib76xfwepuclwd/waspaa2017_drnmf.pdf?dl=0) [Download pretrained sparse NMF dictionaries and weights for the trained networks here](https://www.dropbox.com/s/u8bsw5orkw2l73s/pretrained_models.zip?dl=1) ### Instructions: Uses the [task 2 data from the 2nd CHiME Challenge](http://spandh.dcs.shef.ac.uk/chime_challenge/chime2013/chime2_task2.html), which is available from the [LDC (free for 2017 members, $50 for non-members)](https://catalog.ldc.upenn.edu/LDC2017S10). 1) Download required toolboxes by running `download_toolboxes.sh`. 2) Generate taskfiles by replacing the variable `chime2_path` in `create_taskfiles.sh` by your local CHiME2 path and running `create_taskfiles.sh`. 3) Use `enhance.py` to train, reconstruct, and score audio. Use the `run_waspaa2017.sh` script to replicate results from the WASPAA 2017 paper. Uses code from the following sources, which are automatically downloaded and unzipped by `download_toolboxes.sh`: - sparseNMF by Jonathan Le Roux from http://www.jonathanleroux.org/software/sparseNMF.zip (put Matlab files in "sparseNMF" directory) - BSS Eval by Emmanuel Vincent from http://bass-db.gforge.inria.fr/bss_eval/bss_eval.zip (put "bss-eval" directory in "evaluation" directory) - Matlab PESQ implementation by Y. Hu and P. Loizou from http://ecs.utdallas.edu/loizou/speech/composite.zip (put "composite" directory in "evaluation" directory) - Matlab STOI implementation by Cees Taal from http://ceestaal.nl/stoi.zip (put "stoi" directory in "evaluation" directory) - Matlab Voicebox toolbox by Mike Brookes from http://www.ee.ic.ac.uk/hp/staff/dmb/voicebox/voicebox.zip (put "voicebox" directory in "evaluation" directory)

近期下载者：

相关文件：

评论：[我要评论] [举报此文件]

收藏者：