Stacked frequency-timeGRUs for continuous arousal recognition from musical audio

We address the problem of continuous arousal detection for emotion recognition in musical audio pieces where emotions are represented in the two-dimensional arousal-valence space. We propose a novel method which is a combination of two recurrent neural networks using mel-spectrogram features: A bidirectional GRU network along the frequency dimension as a feature extractor, stacked with a GRU network along the temporal dimension, which is unidirectional for real-time adaptability. The method is evaluated on the MediaEval2015- Emotion in Music Dataset, achieving an RMSE of 0.215 which is better than the results reported by real-time adaptable state-of-the-art models.