Single-channel source separation of audio signals using Bark Scale Wavelet Packet Decomposition

We address the problem of blind source separation from a single channel audio source using statistical model of the sources. We modify the bark scale aligned wavelet packet decomposition, to approximately acquire shift invariance. We allow oversampling in some decomposition nodes to equalize sample rate in all terminal nodes. Statistical models are trained from samples of each source separately. The separation is performed using these models. Experimental results show improved performance compared to a competing algorithm using synthetic and real audio examples.