Acoustic scene classification with multi-temporal complex modulation spectrogram features and a convolutional LSTM network

Acoustic scene classification (ASC) is a mapping from an environmental sound recording to predefined classes representing the auditory scene of the recording. This paper proposes an ASC solution based on the combination of convolutional neural networks, long short term memory cells, and multi-temporal input encoding. The major novelty of the work is applying complex modulation spectrogram for feature extraction. We evaluate the complex modulation spectrogram as discriminant features, resulting in a 4.7% improvement in comparison with the commonly used Mel spectrogram. These features are computed for individual temporal segments of the audio recording to acquire a representation containing both spectral and temporal structure. Also, we derive a de-noising method which has not been used for ASC before but was beneficial in other speech processing tasks. This method leads to 1.5% improvement in prediction accuracy in comparison with a model without de-noising. The proposed model outperforms the state of the art methods by 7.5% in terms of the prediction accuracy for evaluation data in ASC on the DCASE 2017 dataset.