Introduction
Audio clipping is a ubiquitous distortion in digital signal processing, occurring when the amplitude of a signal exceeds the dynamic range of a recording or transmission system. This results in a non-linear loss of information where signal peaks are flattened, introducing harsh harmonic distortion. Traditionally, recovering the original signal—a process known as declipping—has relied on either analytical priors, such as temporal or spectral sparsity, or supervised deep learning models trained on large datasets of paired clean and clipped audio.
However, supervised methods face two major hurdles: the scarcity of high-quality ground-truth data for specific recording environments and a lack of robustness when the test data distribution differs from the training set. This paper proposes a novel self-supervised learning (SSL) framework that recovers audio signals using only the clipped measurements themselves. By leveraging the principle of scale invariance through Equivariant Imaging (EI), the authors demonstrate that a model can learn to reconstruct lost signal peaks without ever needing to see a "clean" example.
Methodology
The proposed approach treats declipping as a non-linear inverse problem. The core innovation lies in adapting Equivariant Imaging—a technique typically used for linear inverse problems—to the non-linear nature of clipping.
Scale Invariance and Equivariant Loss
The fundamental assumption is that natural audio signals are scale-invariant: if a waveform is a valid audio signal, then the same waveform scaled by a constant factor $g$ is also a valid signal. The researchers define an equivariant loss ($\mathcal{L}_{EI}$) that enforces consistency across different scales.
Let $f_\theta$ be the reconstruction network and $\eta$ be the clipping operator. The training objective encourages the model to satisfy the condition: $$f_\theta(\eta(gx)) \approx g f_\theta(\eta(x))$$ By artificially shrinking a clipped signal and re-clipping it, the model learns the relationship between different levels of saturation. This "family of operators" trick allows the network to use information from less-saturated regions to infer the structure of more-saturated peaks.
Measurement Consistency in Saturated Regions
Standard measurement consistency losses ($\mathcal{L}_{MC}$) often fail in declipping because the gradient of the clipping operator is zero in saturated regions. To overcome this, the authors designed a modified consistency loss that provides a learning signal even when the signal is clipped, effectively pushing the model's predictions to exceed the clipping threshold where necessary.
Architectural Considerations
To ensure the network naturally respects the scale-invariance property, the authors utilized bias-free ReLU networks. These architectures are "homogeneous," meaning that scaling the input directly scales the output. Additionally, the model is provided with a binary mask indicating which samples are clipped, which significantly aids the reconstruction process.
Experimental Evaluation
The method was evaluated using both synthetic data and real-world audio.
- Synthetic Data: A random universal subspace of $\mathbb{R}^{100}$ was used to analyze the model's performance under controlled levels of complexity and saturation.
- Real-World Audio (GTZAN): The model was trained on 1-second segments of classical music from the GTZAN dataset, artificially clipped at a threshold of 0.1.
- Robustness Testing: To test for distribution shift, models trained on music were evaluated on a mixture of music and speech (using datasets like VCTK).
Results and Discussion
The self-supervised method demonstrated remarkable performance, nearly matching supervised benchmarks. On the GTZAN dataset, the SSL approach achieved a Signal Distortion Ratio (SDR) of 15.8 dB, compared to 16.43 dB for the supervised model.
The most significant finding was the model's robustness. When faced with a distribution shift (testing on a mix of music and speech after training only on music), the self-supervised model outperformed the supervised one (11.92 dB vs. 10.94 dB). This suggests that supervised models tend to overfit to the specific characteristics of their clean training data, whereas the SSL approach, by training on the distorted measurements themselves, adapts better to the actual signal characteristics at hand.
Conclusion
This research demonstrates that the geometric properties of a signal—specifically its symmetries—can serve as a powerful proxy for ground-truth data in non-linear inverse problems. While the self-supervised method currently exhibits higher variance and slightly lower peak performance than supervised learning in ideal conditions, its robustness to distribution shifts makes it a compelling choice for real-world audio restoration. Future work will focus on providing theoretical guarantees for non-linear Equivariant Imaging and exploring its application to other domains, such as phase retrieval.
Summary updated