Two-stage speech/music classifier with decision smoothing and sharpening in the EVS codec

In most internationally recognized standardized multi-mode codecs, signal classification is performed in a single step by either linear discrimination or SNR-based metrics. The speech/music classifier of the EVS codec achieves greater discrimination than these single-step models by combining Gaussian mixture modelling (GMM) with a series of context-based improvement layers. Additionally, unlike traditional GMM classifiers the EVS model adopts a short hangover period, allowing it to track transitions between music and speech. Misclassifications are mitigated by applying a novel decision smoothing and sharpening technique. The results in relatively static environments demonstrate that the new two-stage approach with selective hangover leads to classification accuracies comparable to speech/music classifiers with longer hangovers. They also show that the new approach leads to faster and more accurate switching of coding modes than conventional classifiers for more complex audio environments such as advertisements, jingles and speech superimposed on music.