High-quality sound restoration is a critical process for preserving audio recordings, particularly those suffering from age, damage, or recording imperfections. Among the most common and disruptive artifacts are clicks and pops—short, sharp transient noises that can stem from vinyl surface defects, digital conversion errors, or microphone handling. Removing these artifacts without degrading the underlying audio requires sophisticated algorithms that can distinguish between genuine signal content and noise. This article provides an in-depth comparison of the leading click removal algorithms, examining their strengths, weaknesses, and ideal use cases to help audio engineers and restoration specialists make informed choices.

Understanding Click Removal in Audio Restoration

Clicks and pops manifest as brief, high-energy transients that occupy a wide frequency range. They can be isolated, occurring as single anomalies, or appear in bursts, often correlated with physical surface damage on vinyl records. Effective click removal algorithms must accomplish three tasks: detection of the click location, estimation of the corrupted sample values, and interpolation or replacement of the damaged audio. The challenge lies in performing these steps without introducing audible artifacts such as “warbling,” high-frequency roll-off, or unnatural smoothing of natural attack transients (like drum hits or plosive consonants).

Modern restoration tools integrate multiple detection strategies, including energy thresholding, spectral analysis, and statistical outlier detection. The choice of algorithm depends on the nature of the click, the density of defects, and the fidelity requirements of the final product. Below we explore the most widely used algorithms, from classic digital signal processing techniques to newer machine-learning-based approaches.

Top Click Removal Algorithms

1. Spectral Subtraction

Spectral subtraction is a frequency-domain method that estimates the noise spectrum of clicks and subtracts it from the overall signal. The algorithm first computes a short-time Fourier transform (STFT) of the audio, identifies frames where clicks are present (often by detecting sudden energy increases), and then attenuates the corresponding frequency bins. This technique is effective for isolated, high-amplitude clicks because it preserves the underlying tonal content in unaffected bands.

However, spectral subtraction can introduce “musical noise” artifacts if the subtraction threshold is too aggressive. It also struggles with clicks that have energy spread across many frequency bins, as reducing those bins may remove part of the desired signal. Careful parameter tuning—particularly the over-subtraction factor and noise flooring level—is essential to balance click removal with audio fidelity. Spectral subtraction is often used as a first pass before more advanced refinement algorithms in professional restoration suites.

For further reading on the mathematical foundations of spectral subtraction in audio enhancement, refer to Wikipedia’s Spectral Subtraction article.

2. Median Filtering

Median filtering operates in the time domain and is one of the simplest click removal methods. The algorithm slides a window of a predefined length (e.g., 3–7 samples) across the audio waveform. If the current sample deviates significantly from the median of the window, it is replaced by the median value. This non-linear approach efficiently suppresses isolated spikes while preserving sharp edges to some degree.

The primary advantage of median filtering is its speed and ease of implementation—it requires no frequency analysis or complex parameter tuning. However, the method has significant limitations. First, it can blur rapid transients that are part of the intended audio, such as cymbal crashes or plucked strings, leading to a loss of “air” and high-frequency detail. Second, the fixed window size makes it less adaptable to clicks of varying duration. For dense click patterns (popcorn noise), the median filter may fail to differentiate noise from valid signal, causing audible distortion. Despite these drawbacks, median filtering remains a useful “quick fix” for lightly damaged recordings and is often employed as a real-time monitor effect.

3. Wavelet-Based Denoising

Wavelet-based methods decompose the audio signal into multiple frequency bands at different time resolutions using wavelet transforms. Clicks, being transient events, appear as high-magnitude coefficients in the high-resolution (high-frequency) bands. The algorithm applies thresholding or shrinkage to these coefficients—hard thresholding to remove strong clicks, or soft thresholding to reduce background noise more gently—then reconstructs the signal from the modified coefficients.

The multi-scale nature of wavelet transforms allows them to capture clicks that vary in duration and frequency content more effectively than fixed-window FFT methods. Wavelet denoising is particularly adept at preserving sharp musical transients because the thresholding can be tailored per frequency band. It is widely considered the gold standard in professional restoration tools such as iZotope RX and CEDAR Audio systems. The main drawback is computational complexity; processing high-resolution audio can be slow without hardware acceleration. Additionally, choosing the correct wavelet family (e.g., Daubechies, Symlet, Coiflet) and threshold parameters requires experience.

A comprehensive overview of wavelet transform applications in audio processing can be found at MathWorks’ wavelet audio denoising page.

4. Adaptive Interpolation

Adaptive interpolation algorithms divide the click removal task into two stages: detection and correction. Detection is typically performed using an energy-based or spectral anomaly detector. Once a click is located, the corrupted samples are replaced using interpolation from neighboring uncorrupted samples. Common interpolation methods include linear interpolation, cubic spline interpolation, and autoregressive (AR) modeling.

AR-based interpolation treats the audio as a time series and predicts the missing samples using a linear prediction model trained on the surrounding context. This approach often yields the most natural-sounding results for isolated clicks because it preserves the local statistical properties of the signal, such as pitch and formant structure. Adaptive interpolation is particularly effective for vocal recordings and spoken word, where median filtering would introduce unnatural timbre changes. However, it requires accurate click detection; false positives can lead to unnecessary interpolation and smearing of legitimate transients. The method also struggles when clicks occur in rapid succession, as the prediction model may not have enough clean data to make reliable estimates.

5. Deep Learning-Based Click Removal

Recent advances in artificial intelligence have led to neural network models capable of click removal with unprecedented accuracy. These models are typically trained on large datasets of clean recordings paired with synthetic click-corrupted versions. Architectures such as U-Nets, WaveNet, and transformer-based speech enhancement networks learn to separate click artifacts from the underlying signal in the time-frequency domain.

Deep learning approaches excel in handling complex click patterns, including non-stationary clicks correlated with signal activity. They can also restore heavily degraded recordings where traditional methods fail. However, they require significant computational resources for both training and inference, and the models may generalize poorly to unseen click types or recording conditions. Additionally, the decisions made by neural networks are not easily interpretable, which can be a concern in archival restoration where provenance and transparency are valued. Nonetheless, deep learning is increasingly integrated into commercial products and is expected to become the dominant method in the coming years.

For an overview of deep learning techniques in audio restoration, see the research article “Deep Learning for Audio Restoration: A Survey” on arXiv.

Comparison of Algorithms

Accuracy

Wavelet-based denoising and deep learning models typically achieve the highest perceptual accuracy, particularly on musical recordings. Spectral subtraction ranks next, though it can introduce frequency-domain artifacts. Median filtering and adaptive interpolation are effective only for specific click types; median filtering is best for isolated, high-amplitude spikes, while adaptive interpolation works well for short gaps in otherwise clean material.

Speed and Computational Efficiency

Median filtering is the fastest, operating in real time even on low-power hardware. Spectral subtraction and adaptive interpolation are moderately fast and can be parallelized. Wavelet-based denoising is slower, especially at high sample rates. Deep learning models require GPU acceleration for real-time use; otherwise they are CPU-intensive.

Parameter Sensitivity

Median filtering has a single parameter (window size) and is easy to set. Adaptive interpolation depends heavily on detection thresholds. Spectral subtraction requires careful tuning of multiple parameters (FFT size, overlap, over-subtraction factor, noise floor). Wavelet denoising requires selecting the wavelet type and threshold method. Deep learning models have no user-adjustable parameters at inference time but require extensive training data and hyperparameter tuning during development.

Risk of Artifacts

Median filtering can cause “smearing” of transients and high-frequency loss. Spectral subtraction may introduce musical noise. Adaptive interpolation, if triggered falsely, can smooth out legitimate attacks. Wavelet-based methods rarely introduce artifacts when properly configured but can produce “birdie” sounds if thresholding is too aggressive. Deep learning models sometimes hallucinate small artifacts, but these are typically less noticeable than traditional processor artifacts.

Practical Considerations for Choosing an Algorithm

The optimal click removal algorithm depends on the specific characteristics of the source material and the restoration goals. For archival transfers of rare vinyl recordings, where every sonic detail matters, wavelet-based denoising combined with manual click detection in a spectral editor is the preferred workflow. For live concert recordings with occasional microphone handling noise, spectral subtraction with a high-quality noise gate may suffice.

Processing power is another consideration. If you need real-time click removal for a broadcast feed, median filtering or a lightweight adaptive interpolation engine running on a DSP chip is more practical than a deep learning model. Conversely, for offline restoration of a precious master tape, waiting for a neural network to process the file is acceptable if the improvement in quality justifies it.

It is also worth evaluating the density of clicks. A record with hundreds of small ticks per side will benefit from a robust detector capable of batch processing—spectral subtraction or wavelet methods can handle high density better than median filtering. Sparse, loud clicks (such as those from digital glitches) are efficiently removed by adaptive interpolation without affecting adjacent audio.

Finally, consider the musical genre. Classical music with delicate harmonic textures requires an algorithm that preserves subtle overtones; wavelet and deep learning models excel here. Rock or electronic music with heavy dynamics may tolerate median filtering or spectral subtraction without noticeable degradation, allowing for faster processing.

The field of audio restoration is shifting toward machine learning approaches, driven by the availability of large training datasets and advances in generative models. Self-supervised learning methods, where the model learns to detect and remove clicks without paired clean/corrupted data, are emerging to address the lack of ground-truth recordings for historical material. Additionally, real-time neural network inference on mobile devices is becoming feasible thanks to quantization and model pruning techniques.

Another promising direction is the integration of click removal into broader restoration frameworks that also handle noise reduction, de-clipping, and bandwidth extension. For example, end-to-end models that take a raw corrupted waveform and output a clean waveform are being developed, reducing the need for pipeline engineering. However, these models still face challenges with rare click types and adversarial examples, so traditional algorithms remain valuable as fallbacks.

Open-source libraries like Guitarix and Audacity have incorporated wavelet and spectral subtraction click removal, making these techniques accessible to hobbyists and small studios. Commercial products continue to refine their algorithms, with iZotope RX’s “Repair Assistant” using machine learning to suggest optimal settings based on the audio content.

Conclusion

Click removal is a nuanced discipline where no single algorithm dominates all scenarios. Spectral subtraction offers a balance between effectiveness and complexity, making it suitable for many everyday restoration tasks. Median filtering provides a fast, simple solution for minor defects but falls short on critical work. Wavelet-based denoising remains the professional standard for high-fidelity restoration, while adaptive interpolation excels in preserving naturalness for sparse clicks. Deep learning is rapidly advancing and will likely become the default choice as models become more efficient and generalizable.

Audio engineers should evaluate each algorithm based on the recording’s condition, the desired output quality, and available processing resources. By understanding the strengths and limitations of each method, restoration specialists can choose—or combine—algorithms to achieve transparent click removal that lets the original performance shine through. As technology evolves, the line between restoration and enhancement will continue to blur, ultimately giving us the ability to hear historical recordings as they were meant to be heard.