AutoSpeech2020

1st place solution https://www.automl.ai/competitions/2

Usage

git clone https://github.com/u1234x1234/AutoSpeech2020.git

class Model from the file model.py satisfies the interface, so you could just run the following line in order to reproduce the results:

python run_local_test.py -dataset_dir=path_to_dataset -code_dir=path_to_model_file

Please refer to the official detailed description of the evaluation protocol.

How it works

The basic ideas:

• Get a decent result as fast as possible with the simplest models, then train more elaborate ones.
• There's no single model that performs the best on the all datasets, so try different ones.

Used models:

1. Logistic Regression on features extracted with pretrained model trained on the speaker recognition task
2. Logistic Regression on features extracted with pretrained model trained on the music genre classification task
3. Logistic Regression on the combination of features from 1. 2.
4. AutoSpeech 2019 1st place solution by Hazza Cheng
5. AutoSpeech 2019 3rd place Solution by Kon
6. Fine-tuning of pretrained network from 1.

Then average the results from multiple models with geometric mean.

Acknowledgements

1. AutoSpeech 2019 1st place solution by Hazza Cheng, GPL Licence, Code modifications
2. AutoSpeech 2019 3rd place Solution by Kon, MIT Licence
3. Speaker recognition pretrained model by ClovaAI, MIT Licence
4. Musicnn by Jordi Pons, ISC Licence

Notice

Each subdirectory from 3rdparty contains subcomponents with separate copyright notices and license terms. Please refer to Licence provided in specific subdirectory.