Speech and music pitch trajectory classification using recurrent neural networks for monaural speech segregation

摘要

In this paper, we propose speech/music pitch classification based on recurrent neural network (RNN) for monaural speech segregation from music interferences. The speech segregation methods in this paper exploit sub-band masking to construct segregation masks modulated by the estimated speech pitch. However, for speech signals mixed with music, speech pitch estimation becomes unreliable, as speech and music have similar harmonic structures. In order to remove the music interference effectively, we propose an RNN-based speech/music pitch classification. Our proposed method models the temporal trajectories of speech and music pitch values and determines an unknown continuous pitch sequence as belonging to either speech or music. Among various types of RNNs, we chose simple recurrent network, long short-term memory (LSTM), and bidirectional LSTM for pitch classification. The experimental results show that our proposed method significantly outperforms the baseline methods for speech-music mixtures without loss of segregation performance for speech-noise mixtures.