TSM

TSM predicts subjective mean opinion scores (SMOS) for time-scaled audio by computing an objective measure of quality (OMOQ) from handcrafted and PEAQ features using a fully connected neural network implemented in MATLAB.

Key Features:

• Dataset: Uses a dataset of time-scaled audio signals annotated with subjective quality labels (SMOS) as the training and evaluation reference.
• Objective Measure (OMOQ): Computes an improved objective measure of quality that combines handcrafted features and a fully connected neural network to predict SMOS.
• Feature Set: Incorporates basic and advanced PEAQ features plus nine TSM-specific features tailored to time-scale modification artifacts.
• Alignment Techniques: Implements six methods to align reference and test magnitude spectra, including an interpolation method that resamples the reference magnitude spectrum to the test length and yields superior prediction performance.
• Performance Metrics: Reports mean RMSE of 0.490 and mean Pearson correlation of 0.864 against SMOS, corresponding to the 97th percentile for RMSE and the 82nd percentile for Pearson correlation in subjective evaluations.
• Algorithm Evaluation: Uses OMOQ to compare TSM algorithms, identifying Elastique as highest-quality for solo instrument and voice signals and the identity phase-locking phase vocoder as highest-quality for music signals and overall.
• MATLAB implementation: Implemented for MATLAB.

Scientific Applications:

• TSM algorithm benchmarking: Objectively compares time-scale modification algorithms across audio types using OMOQ and SMOS correlations.
• Perceptual quality prediction: Predicts perceived audio quality (SMOS) from objective features for time-scaled signals.
• Audio engineering research: Supports evaluation of tempo and length modification methods for solo instruments, voice, and music to guide algorithm selection and development.

Methodology:

Extracts basic and advanced PEAQ features plus nine TSM-specific handcrafted features, aligns reference and test magnitude spectra using six methods including interpolation, inputs features to a fully connected neural network to predict SMOS, and evaluates predictions with RMSE and Pearson correlation against subjective labels.