SPECTRA, a method for few-shot class-incremental audio classification, reported higher average accuracy and a lower performance drop than TAPE on NSynth-100, FSC-89 and LS-100. The document is an arXiv version 1 preprint dated 25 August 2026.
The standard evaluations used five sessions, with five classes per session and five labeled examples, or shots, for each class. The reported values were averaged over 50 seeds.
The reported gap
The comparison uses two summary measures. Average accuracy, or AA, is the mean across sessions. Performance drop, or PD, is the first-session accuracy minus the last-session accuracy. Higher AA and lower PD are the more favorable pattern.
On NSynth-100, SPECTRA reported an AA of 96.52, against 93.48 for TAPE. Its PD was 3.23, compared with 5.58 for TAPE.
On FSC-89, SPECTRA's AA was 71.12, compared with TAPE's 69.24, while its PD was 18.77 against 19.96. On LS-100, SPECTRA reported AA of 92.58 and PD of 8.67, versus 85.49 and 11.50 for TAPE.
At session level, SPECTRA led TAPE at every reported standard-protocol session. On NSynth-100, the margin grew from +1.5 at S0 to +3.9 at S4.
What SPECTRA combines
SPECTRA combines a trainable residual adapter, replay of old classes from low-rank subspaces, and transductive optimal-transport refinement of class prototypes. The audio-language encoder remains frozen.
During training, the adapter, learned prototypes and reference anchors are updated. The encoder stays frozen and the transform remains closed-form. The optimal-transport refinement is applied only at inference, without gradients or parameter updates.
The pattern held under harder protocols
The paper also reports tests with more classes handled in each session. The harder protocol uses 10 classes per session over 10 sessions for NSynth-100 and LS-100, and 10 classes per session over eight sessions for FSC-89.
On NSynth-100, SPECTRA's AA/PD pair was 92.45/8.77, versus 87.66/14.11 for TAPE. On LS-100, the pairs were 84.26/17.31 for SPECTRA and 75.53/24.43 for TAPE. FSC-89 showed the same direction, with 58.02/25.44 for SPECTRA and 55.22/26.66 for TAPE.
Replay structure was part of the comparison
An ablation chain on NSynth-100 listed TAPE at 93.5/5.6 for AA/PD, followed by 95.5/4.3 for the adapter configuration, 95.8/4.7 for the transport configuration, 95.8/4.6 for Gaussian replay and 96.5/3.2 for subspace replay.
On FSC-89, the TAPE and subspace-replay configurations were 69.2/20.0 and 71.1/18.8. In a separate replay-structure comparison, the paper reported that subspace replay beat Gaussian replay on both metrics in two datasets. On NSynth-100, Gaussian replay barely improved over no replay. The reported comparison favored preserving low-rank class structure over replay alone in those analyses.
What the results do not settle
The reported values are point means, and uncertainty intervals are not supplied. The paper says its ablations statistically validate each component, but the supplied material does not name the tests or report p-values, confidence intervals or standard deviations.
The evidence is limited to computational comparisons on NSynth-100, FSC-89 and LS-100 under the reported class-incremental protocols, with the encoder frozen. It does not establish performance outside these benchmarks or in a production system.
Funding and conflicts of interest are not reported. Neither are data availability, code availability or supplementary materials.
Paper data and sources
Original title: SPECTRA: Subspace-Preserving Embedding Calibration, Transport, and Replay for Fully Few-Shot Class-Incremental Audio Classification
Authors: Giries Abu Ayoub, Loay Mualem, Simon Korman
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text