Segmentation, classification and clustering of an Italian broadcast news corpus

摘要: This work reports on preliminary activity at ITC-irst the problem of acoustic segmentation, classification and clustering an Italian audio broadcast news corpus. The approach is based following stages. First, input data stream segmented by detecting spectral changes through Bayesian Information Criterion (BIC). Second, segments are classified in terms conditions, modeled mixtures Gaussians. Finally, from same speakers clustered, using again BIC. The scheme proposed for automatic causes a degradation recognition error rate, with respect to fully supervisioned experiment, equal 1.3% before adaptation, 3.4% after adaptation.