A MAP criterion for detecting the number of speakers at frame level in model-based single-channel speech separation

The problem of detecting the number of speakers for a particular segment occurs in many different speech applications. In single channel speech separation, for example, this information is often used to simplify the separation process, as the signal has to be treated differently depending on the number of speakers. Inspired by the asymptotic maximum a posteriori rule proposed for model selection, we pose the problem as a model selection problem. More specifically, we derive a multiple hypotheses test for determining the number of speakers at a frame level in an observed signal based on underlying parametric speaker models, trained a priori. The experimental results indicate that the suggested method improves the quality of the separated signals in a single-channel speech separation scenario at different signal-to-signal ratio levels.