An objective measure of quality for time-scale modification of audio

TL;DR

The paper proposes an objective quality measure for time-scale modified audio using hand-crafted PEAQ and TSM-specific features with a fully connected neural network trained on a subjective dataset. It reports mean RMSE 0.487 and mean Pearson correlation 0.865, and uses the measure to rank TSM algorithms.

Key findings

• The proposed measure achieved a mean RMSE of 0.487 and a mean Pearson correlation of 0.865, equivalent to the 98th and 82nd percentiles of subjective sessions respectively (abstract; Results states PCC 82nd and RMSE 98th).
• Among six alignment methods, interpolation of the reference magnitude spectrum to the length of the test magnitude spectrum gave the best performance; including reference files improved results.
• For TSM algorithm evaluation, the paper reports Elastique gives the highest objective quality for Solo instrument and voice signals, while the Identity Phase-Locking Phase Vocoder (IPL) gives the highest objective quality for music signals and the best overall quality.
• Table II overall mean OMOS: IPL 3.717, Driedger’s IPL 3.710, Elastique 3.694; for music IPL 3.810 and Elastique 3.691; for voice Elastique 3.617 and IPL 3.568; for solo Driedger’s IPL 3.833 and Elastique 3.773, which differs from the abstract/conclusion statement that Elastique is highest for solo.

Methods

• Dataset: training subset of 5,280 processed files from 88 reference files at 10 ratios using six TSM methods; testing subset of 240 files from 20 reference files using three additional methods; evaluation subset of 5,200 files (400 per method) at 20 ratios in 0.22 < β < 2.2. Subjective labels came from 42,529 ratings by 263 participants in 633 sessions, with a minimum of 7 ratings per file.
• Features: PEAQ Basic and Advanced MOVs plus nine TSM-specific features (phasiness, spectral similarity, transient, etc.); features were pruned if Pearson correlation between same-type features exceeded approximately 0.95, normalized to [0,1], and trained to normalized SMOS.
• Network: fully connected network with three hidden layers of 128 nodes, layer normalization, ReLU activation, residual connections around the second and third layers, and sigmoid output; trained for 800 epochs (extended if still improving), single batch, RMSE loss, AdamW, learning rate 1e-4, 10% validation, seeds 0–99; final alignment used interpolation to test including reference files.

Limitations

• Stated: future research includes evaluating a wider range of commercial and lesser-known TSM methods, comparing different implementations of the same method, and exploring alternative/deeper neural networks; initial testing gave ρte of 0.71 for a random forest and 0.65 for a CNN with blind data-driven features.
• Evident from text: the method assumes a constant time-scale ratio and single-channel audio; PEAQ assumes 48 kHz while the dataset uses 44.1 kHz; evaluation averages exclude β = 1 and β < 0.25; some TSM method results are noisy and were smoothed; the solo-signal ranking in the abstract/conclusion conflicts with the Results text and Table II.