Topcoder Soundscapes Marathon Match Academic Prize solution

This is a solution to the Soundscapes Marathon Match competition organised by Topcoder and National Geospatial-Intelligence Agency. The main goal of the competition was to geo-locate audio files consisting of non-speech ambient noise in one of nine target cities.

Solution

The final solution consists of an ensemble of nine CNN models fitted on mel spectrograms, with heavy use of data augmentations.

Key points:

• Raw audio converted to log-scaled mel spectrograms
• Random crops of mel spectrograms along time axis as input during training
• Five-fold group cross-validation
• Pretrained CNN models from timm
• SpecAugment and Mixup data augmentation techniques

Replicate results

The results were obtained using two Nvidia V100 GPUs, however the training script can be adapted to a different GPU setup. The code requires Docker and nvidia-docker installed.

Steps

1. Build a Docker image from the Dockerfile
2. Run the container. The training data expected to be in /data and output will be saved in /wdata. For instance, run

docker run --gpus all --ipc=host \
	-v ~/<path>/data/train:/data \
    -v ~/<path>/wdata:/wdata \
    -it <image-name>

3. Run train.sh <path-to-speech-files> <path-to-files-without-speech> to preprocess .flac files and fit the models. This should create 9 folders in /wdata/models containing model checkpoints
4. Run test.sh <path-to-files-without-speech> <path-to-output-file> to preprocess test .flac files and make predictions.