Complete and separate: Conditional separation with missing target source attribute completion
This work addresses a limitation in conditional source separation for audio processing, offering a more flexible and easier-to-use alternative, though it is incremental as it builds on existing multi-conditional frameworks.
The paper tackles the problem of source separation when only partial semantic information about the target source is available, by training a model to complete missing semantic data and using it to enhance a multi-conditional separation network, resulting in performance that approaches an oracle model and matches specialized single conditional models.
Recent approaches in source separation leverage semantic information about their input mixtures and constituent sources that when used in conditional separation models can achieve impressive performance. Most approaches along these lines have focused on simple descriptions, which are not always useful for varying types of input mixtures. In this work, we present an approach in which a model, given an input mixture and partial semantic information about a target source, is trained to extract additional semantic data. We then leverage this pre-trained model to improve the separation performance of an uncoupled multi-conditional separation network. Our experiments demonstrate that the separation performance of this multi-conditional model is significantly improved, approaching the performance of an oracle model with complete semantic information. Furthermore, our approach achieves performance levels that are comparable to those of the best performing specialized single conditional models, thus providing an easier to use alternative.