Speech quality estimation with deep lattice networks

Intrusive subjective speech quality estimation of mean opinion score (MOS) often involves mapping a raw similarity extracted from differences between the clean and degraded utterance onto MOS with fitted function. More recent models such as support vector regression (SVR) or deep neural networks use multidimensional input, which allows for more accurate prediction than one-dimensional (1-D) mappings but does not provide monotonic property that is expected quality. We investigate function using lattice (DLNs) to constraints input features provided by ViSQOL. The DLN improved 0.24 mean-square error on mixture datasets include voice over IP codec degradations, outperforming 1-D functions SVR well PESQ POLQA. Additionally, we show can be used learn quantile well-calibrated useful measure uncertainty. provides an data driven representations human interpretable scales, intervals predictions instead point estimates.