Direct estimation and improvement of Leptokurtic shape parameters

Develop methods to directly estimate and improve the shape parameters of the Leptokurtic Generalized Gaussian distribution in the heavy-tailed neural FCASA system, whose diarization performance is sensitive to hyperparameter tuning and degrades for some tested parameter configurations.

Background

The heavy-tailed neural FCASA system replaces the Gaussian source-separation model with heavy-tailed distributions, including the Leptokurtic Generalized Gaussian distribution. Experiments show that some Leptokurtic shape-parameter settings, specifically beta in {0.4, 0.8, 1.2}, produce worse diarization performance than the Gaussian baseline, indicating sensitivity to hyperparameter selection.

The unresolved issue is whether and how these distributional shape parameters can be estimated directly and how the resulting parameterization can be improved to avoid performance degradation. The paper explicitly leaves this investigation for future work rather than resolving it experimentally or theoretically.