Speech quality estimation with deep lattice networks.

Area of Science:

• Speech processing and signal analysis
• Machine learning for audio quality assessment
• Human-computer interaction and subjective evaluation

Background:

• Objective speech quality estimation typically maps similarity scores to Mean Opinion Scores (MOS) using fitted functions.
• Advanced models like Support Vector Regression (SVR) and deep neural networks offer multidimensional inputs but lack monotonic property between similarity and quality.
• Existing methods like PESQ and POLQA have limitations in accurately reflecting subjective speech quality.

Purpose of the Study:

• To investigate a multidimensional mapping function using Deep Lattice Networks (DLNs) for speech quality estimation.
• To incorporate monotonic constraints into the mapping of similarity features to MOS.
• To develop a more accurate and reliable objective speech quality assessment model.

Main Methods:

• Utilized Deep Lattice Networks (DLNs) for multidimensional mapping of speech features.
• Employed input features from ViSQOL (Virtual Speech Quality Objective Listener) for the DLN model.
• Trained and evaluated the DLN on diverse datasets including Voice over IP (VoIP) and codec degradations.

Main Results:

• Achieved a mean-square error of 0.24 in speech mapping, outperforming 1-D functions, SVR, PESQ, and POLQA.
• Demonstrated that DLNs can learn a well-calibrated quantile function for uncertainty measurement.
• Showcased improved mapping of similarity representations to human-interpretable scales, offering quantile intervals instead of point estimates.

Conclusions:

• Deep Lattice Networks provide a robust framework for monotonic speech quality mapping, enhancing prediction accuracy.
• The DLN model offers a valuable measure of uncertainty through quantile functions, improving the interpretability of predictions.
• This research advances objective speech quality estimation by integrating deep learning with monotonic constraints and uncertainty quantification.