A multi-task net for KWS and speaker detection. The description of this experiment is in the paper： “MTN-CBAM: Multi-Task Network with Convolutional Block Attention Module for Speaker Related Small-Footprint Keyword Spotting”
In order to run these Python scripts, the following libraries and packages are needed:
* Keras
* Librosa
* Numpy
* Pickle
* Matplotlib
When running these Python scripts, by default, it is expected to find two folders within this one: "HADataset" and "exp". The first would contain the hearing aid speech database that can be freely downloaded from Data Link The second folder is the working directory, where all files resulting from running the provided scripts are stored.
Thanks to Lopezespejo I, Tan Z, Jensen J, et al. Keyword Spotting for Hearing Assistive Devices Robust to External Speakers[C]. conference of the international speech communication association, 2019: 3223-3227. for providing this data set
run.sh demonstrates the running example
The pre-trained model of "MTN-CBAM" and "MTN-CBAM-2" using "2 * 2" convolution kernel is provided in "exp".
You can use "test_cbam.py" to test them, the specific command reference in "run.sh"