Inter-channel information-based noise detector

Dataset preparation

Download ToyADMOS2 https://github.com/nttcslab/ToyADMOS2-dataset

Run mixer.py

Here, consider that we only used model A for both toy car and toy train.

After running, ./processed_data will be made

Noise detection

Basically, you modify baseline.yaml (or infer2others.yaml) and run python code.

Each variable in yaml file has below meaning and options.

<baseline.yaml>

machine_type: toy car / toy train

machine type that you want to experiment with

bg: 1 / 2 / 3 / 4

background noise that you want to experiment with

rseed: (recommended) 0 / 10 / 20

random seed that you want to experiment with

input_feature: STFT / IID / IPD / sinIPD / IID+sinIPD

input feature that you want to experiment with

==========================================================================

<infer2others.yaml>

bg_from : 1 / 2 / 3 / 4

background noise that is used for train and validation

bg_to : 1 / 2 / 3 / 4 (different from bg_from)

background noise that will be tested in

==========================================================================

other parameters

batch_size : (recommended) 16

batch size

short_cut : (recommended) True

you can pass data loading process if you make this True and data is already made as 'dataset_prepared'

epochs : (recommended) 20

epochs

exclude : (recommended) [aL, aM, aH, bL, bM, bH, cL, cM, cH, dL, dM, dH]

machine condition that will be excluded during train and validation

on-site

Modify machine_type, bg, rseed, input_feature of baseline.yaml and run noise_detection.py

results can be found in ./dataset_prepared/.../result/

other site

Modify machine_type, bg_from, bg_to, rseed, input_feature of infer2others.yaml and run noise_detection_infer2others.py

results can be found in ./infer2others/.../result/

draw

Modify machine_type, bg, rseed baseline.yaml and run draw.py

You can draw IID+sinIPD / intensity and phase spectrum