The endsem project for MLSP-2020
Task-1 is the implimentation of the model from [1]. Here the goal is to classify given audio file which contains a single event-class - for example "dog_bark". There are a total of 10 classes we are interested in.
Task-2 is more interesting in the sense that here we detect the sequence of these event-classes. The network which was used is shown in the following figure. We train the network using CTC[2] loss. Spectogram Image Features (SIF) augmented with energy in each frame as described in [1] is used as features. And mean normalized edit-distance is used as a metric.
[1] P han, H., Hertel, L., Maaß, M., & Mertins, A. (2016). Robust Audio Event Recognition with 1-Max Pooling Convolutional Neural Networks. I NTERSPEECH.
[2]: Graves, S.Fernández, F.Gomez, and J.Schmidhuber. (2006). Connectionist temporal classification: l abelling unsegmented sequence data with recurrent neural networks.