Interspeech 2023

Self-Supervised Dataset Pruning for Efficient Training in Audio Anti-spoofing

Abdul Hameed Azeemi, Ihsan Ayyub Qazi, and Agha Ali Raza

Lahore University of Management Sciences

A label-free pruning method uses wav2vec 2.0 representations and clustering to retain informative audio, outperforming supervised pruning metrics while reducing selection computation by 91%.

Paper ISCA Archive

24th Annual Conference of the International Speech Communication Association

Abstract illustration of audio representations forming informative clusters.

Introduction

Neural audio anti-spoofing systems are increasingly expensive to train. Existing data-pruning metrics can reduce dataset size, but they depend on labels and an initial model-training run to determine which examples are important.

We propose a self-supervised alternative that selects informative audio directly from pretrained speech representations. It removes both label dependence and the preliminary anti-spoofing training step from the pruning pipeline.

Wav2vec representation, clustering, data pruning, and anti-spoofing training. — Audio is embedded with wav2vec 2.0, clustered, ranked by centroid distance, and pruned before training the anti-spoofing model.

Pruning from pretrained speech representations

The method treats distance in self-supervised representation space as a proxy for example difficulty and informativeness.

1. Encode every utterance

Each 16 kHz audio sample is passed through a 95M-parameter wav2vec2-base model. Its contextual representations are flattened into a fixed-size embedding.

2. Cluster the representation space

K-means groups examples with similar latent speech characteristics. This reveals structure without using bonafide or spoof labels.

3. Score by centroid distance

Cosine distance to the nearest cluster centroid estimates example difficulty. Farther, less prototypical examples receive higher importance scores.

4. Retain informative audio

Within each cluster, the highest-distance fraction is kept. RawNet2 or AASIST-L is then trained on the resulting compact subset.

Evaluation

We evaluate on the ASVspoof 2019 Logical Access dataset using RawNet2 and AASIST-L. Equal error rate and minimum tandem detection cost are averaged over three runs at pruning fractions from 0% to 90%.

The comparison includes Random, EL2N, Forgetting Score, and Forgetting Norm. Unlike these supervised metrics, self-supervised pruning does not require labels or an initial anti-spoofing training run.

Model at 90% pruning	Self-supervised	Best baseline
RawNet2 EER / min-tDCF	12.84 / 0.34	15.29 / 0.41
AASIST-L EER / min-tDCF	10.30 / 0.27	13.74 / 0.34

Results

At the most aggressive setting—90% of training data removed—the self-supervised method achieves the best EER and min-tDCF on both architectures. For RawNet2, it lowers EER from 19.09 to 12.84 and min-tDCF from 0.56 to 0.34 compared with the second-best scoring metric.

Selection uses only 8.51% of the computation required by Forgetting Norm on RawNet2, corresponding to approximately 91% lower pruning time.

T-SNE comparison of audio subsets produced by six pruning methods. — T-SNE views at 80% pruning. Self-supervised selection preserves a distinctive mixture of attacks and bonafide speech compared with supervised metrics.

Sensitivity to the number of clusters

Across moderate pruning levels, performance is stable over a broad range of k-means cluster counts. Cluster choice becomes more influential only when very little data remains, where preserving every informative subpopulation is harder.

RawNet2 EER across cluster counts and pruning fractions. — RawNet2 EER for different values of k. The method remains comparatively stable until extreme pruning.

Conclusion

Pretrained speech representations provide a fast, label-free signal for dataset pruning. By keeping examples that are distinctive within the latent space, the method improves anti-spoofing performance under severe data constraints while avoiding an expensive scoring-model training run.

Citation

@inproceedings{azeemi23_interspeech,
  title     = {Self-Supervised Dataset Pruning for Efficient
               Training in Audio Anti-spoofing},
  author    = {Azeemi, Abdul Hameed and Qazi, Ihsan Ayyub
               and Raza, Agha Ali},
  booktitle = {Interspeech},
  pages     = {2773--2777},
  year      = {2023}
}