Deep clustering (DPCL) was proposed by a research team from Mitsubishi Electric Research Laboratory and Columbia University in August 2015.Deep clustering: Discriminative embeddings for segmentation and separation".

DPCL is a deep learning framework proposed for solving the problem of sound source separation. When trained on spectrogram features from a two-speaker mixture and tested on a set of speaker mixtures not used in training, the framework can infer a masking function that improves signal quality by approximately 6 dB. DPCL does not require class labels, making it potentially trainable on a wide range of voice types.