# 14th place solution

First of all, we’d like to thank the hosts and Kaggle for organizing this competition.
Congratulations to
[@xyzdivergence](https://www.kaggle.com/xyzdivergence)
on achieving Kaggle Competitions Grandmaster—well deserved!
Model & Training
We used a SED architecture and trained on random 10-second audio segments. For the final ensemble, we relied solely on tf_efficientnetv2_m.in21k, trained with slightly varied configurations.
The best private score for a single SED model was 0.922 (not the selected submission), and 0.894 for a CNN model using the same backbone.
Our training pipeline consisted of several stages:
Stage 1: Pretraining on the training audio with true labels.
Stage 2: Knowledge distillation using both train audio and train soundscapes. We combined average pseudo labels across the full audio (with a 1-second stride, weighted at 0.3) and chunk-level pseudo labels from the teacher model (10-second chunks, weighted at 0.7).
We performed several rounds of distillation, selecting the best-performing teacher from the previous round based on leaderboard improvements.
Features
Melspectrogram settings:
Augmentations:
On waveform: sumix (=1)
On spectrogram: mixup (=1), 3 time/frequency masks (=0.5), horizontal flip (=0.5),  random erasing (=0.5)
Final Submission
The final submission was a simple average of three tf_efficientnetv2_m.in21k checkpoints, followed by smoothing using neighboring clips with weights of 0.1, 0.8, and 0.1.
To speed up inference, all models were converted to OpenVINO.