# 13rd solution summary

Hi,
As I have made a slide about my solution, I would like to post it here. Thank organizer and staffs for this great competition.
Here is the brief summary.
Inference:
Raw audio to log mel.
Apply sliding window to the log mel and predict each windows by CNN.
Feed temporal prediction given step2 to RNN
Feed temporal prediction given step2 to Rakeld + ExtraTrees
Average step3 and step4.
Training:
Train CNN by using only curated data.
Apply sliding window to the log mel in noisy dataset and predict each windows by CNN.
Feed temporal prediction given step2 to RNN. Then, I got pseudo labels for noisy dataset.
Average pseudo labels with original noisy labels by 0.75:0.25.
Train CNN by using all data.
Train some models to follow the inference steps.
Details in Training CNN:
Densenet 121
Spec Aug
Mixup
CLR + SWA + Snapshot ensemble.
I intensively tried SSL approach such as ICT, but it did not work.