Inroduction

This is the source code of "Detect what you want: Target Sound Detection"

first step ---> data process

The process of getting data, as Figure 1 shows: Please first download the Urbansound-sed dataset and Urbansound8K dataset.
After that, please run the following commond:
python data/extract_feature.py
if a mixture audio include N envents, so that we can generate N samples. sample style： filename_event target_event mel_feature time eg. soundscape_test_bimodal0_gun_shot.wav gun_shot [[0,1],[1,1]] [[],[]] In our experiments, we randomly choose embedding according to class name.

you can also use the feature we extracted, please download it from google dirve https://drive.google.com/drive/folders/1oMfoDX_UQgalKMQstUx3D8kSwOvP8KXz?usp=sharing. you only need to place the feature file to correct path, then you can run it.

How to run code

bash bash/tsd.sh

Join train

please run bash/jtsd.sh

TO do list

We will upload our extracted feature and trained model in the future.
We only realease the fully supervised training code, the weakly training code is similar with this. If you need it, please let me known.

reference

https://github.com/RicherMans/CDur https://www.kaggle.com/adinishad/urbansound-classification-with-pytorch-and-fun https://github.com/qiuqiangkong/audioset_tagging_cnn