Traditional audio noise-reduction methods often analyze a signal in either the time domain or frequency domain.

But real-world audio is more complicated.

Speech contains rapidly changing transients, harmonics, pauses, consonants, and broadband components. Noise can also change over time and may appear at different time scales.

This creates a fundamental challenge:

How can we separate useful audio information from noise when both their frequency and time characteristics are changing?

Wavelet denoising provides one approach.

Instead of analyzing the entire signal with a single fixed frequency resolution, wavelet methods decompose the signal into different scales. This allows the algorithm to examine both where signal energy occurs and at what time scale it occurs.

For audio processing, this makes wavelet denoising particularly interesting for signals containing:

• Transient noise
• Impulsive noise
• Broadband noise
• Non-stationary noise
• Short-duration disturbances
• Speech with rapidly changing spectral characteristics

Wavelet denoising is therefore an important technique to understand when building a broader audio noise reduction and speech enhancement toolkit.

1. What Is Wavelet Denoising?

Wavelet denoising is a signal-processing technique that uses a wavelet transform to decompose an audio signal into components at different scales.

The basic idea is:

Noisy Audio

↓

Wavelet Decomposition

↓

Wavelet Coefficients

↓

Threshold / Modify Noise-Dominated Coefficients

↓

Wavelet Reconstruction

↓

Denoised Audio

The method relies on the observation that useful signal structures and noise may produce different distributions of wavelet coefficients.

Large coefficients often correspond to important signal structures, while many small coefficients may be associated with noise.

The algorithm can therefore suppress coefficients that are likely to be noise while retaining coefficients that contain useful audio information.

2. Why Wavelets Are Useful for Audio

Audio signals are inherently time-varying.

Consider a speech signal.

A vowel may remain relatively stable for tens or hundreds of milliseconds, while consonants such as plosive or fricative sounds can change much more rapidly.

A conventional Fourier transform provides frequency information but does not directly indicate when a particular frequency component occurs.

STFT improves this by analyzing short-time segments, but its time-frequency resolution is determined by the selected analysis window.

Wavelet analysis takes another approach.

It uses different scales to examine signal structures at different resolutions.

This is commonly described as:

Multiresolution analysis

3. Wavelet Transform vs. Fourier Transform

The difference can be summarized conceptually.

The important point is not that wavelets are universally better than FFT or STFT.

Each method is useful for different signal-processing problems.

For example:

• FFT is excellent for frequency analysis.
• STFT is widely used for speech enhancement.
• Wavelets can be useful for transient and multi-scale signal analysis.
• Adaptive filters are useful when a correlated reference signal is available.

4. Continuous Wavelet Transform

The Continuous Wavelet Transform (CWT) represents a signal using scaled and shifted versions of a wavelet function.

A simplified expression is:

W(a,b) = (1 / √|a|) ∫ x(t) ψ[(t − b)/a] dt*

where:

• x(t) = input signal
• ψ(t) = mother wavelet
• a = scale parameter
• b = time-shift parameter
• ψ* = complex conjugate

The two key parameters are:

Scale

Controls the size of the wavelet.

Translation

Determines where the wavelet is positioned in time.

This allows the transform to examine local structures at multiple scales.

5. Discrete Wavelet Transform

For practical digital signal processing, the Discrete Wavelet Transform (DWT) is often more useful because it reduces computational complexity.

The DWT decomposes the signal into different frequency bands or scales.

A typical decomposition looks like:

Input Audio

↓

Level 1

• Approximation A1
• Detail D1

↓

Level 2

• Approximation A2
• Detail D2

↓

Level 3

• Approximation A3
• Detail D3

And so on.

The approximation contains lower-frequency information, while detail coefficients represent progressively different higher-frequency components.

6. Wavelet Decomposition

A common DWT implementation can be understood using a pair of filters:

• Low-pass filter
• High-pass filter

The input signal is passed through both.

The low-pass branch produces the approximation coefficients.

The high-pass branch produces the detail coefficients.

The process can then be repeated on the approximation component.

Conceptually:

Input

→ Low-pass → Approximation

→ High-pass → Detail

Then:

Approximation

→ Low-pass → Lower-frequency Approximation

→ High-pass → Detail at next scale

This creates a hierarchical representation of the signal.

7. Why This Helps with Noise Reduction

Suppose the input is:

y(t) = x(t) + n(t)

where:

• x(t) = clean audio
• n(t) = noise

After wavelet decomposition:

Y → Wavelet Coefficients

Some coefficients will contain strong signal information.

Others may contain mostly noise.

Instead of filtering the original waveform directly, the denoising algorithm modifies the wavelet coefficients.

The general process is:

Noisy Signal

→ Wavelet Decomposition

→ Coefficient Analysis

→ Noise Suppression

→ Wavelet Reconstruction

This is the central idea behind wavelet denoising.

8. Wavelet Thresholding

Thresholding is one of the most important techniques in wavelet denoising.

A threshold λ is selected.

Wavelet coefficients with relatively small magnitudes are treated as more likely to represent noise.

Two common thresholding strategies are:

• Hard thresholding
• Soft thresholding

9. Hard Thresholding

Hard thresholding keeps coefficients whose magnitude is above the threshold and removes coefficients below it.

The basic rule is:

ŵ = w, if |w| ≥ λ

ŵ = 0, if |w| < λ

where:

• w = original wavelet coefficient
• ŵ = processed coefficient
• λ = threshold

This creates a simple decision:

Small coefficient → remove

Large coefficient → keep

Hard thresholding can preserve strong signal features effectively, but the abrupt transition at the threshold can sometimes introduce artifacts.

10. Soft Thresholding

Soft thresholding reduces the magnitude of coefficients above the threshold.

The commonly used expression is:

ŵ = sign(w)(|w| − λ)

when:

|w| > λ

Otherwise:

ŵ = 0

In other words:

1. Small coefficients are removed.
2. Large coefficients are retained.
3. Large coefficients are also slightly reduced.

Soft thresholding produces a smoother transition and is widely used in practical wavelet denoising.

11. Hard vs. Soft Thresholding

The appropriate choice depends on the signal and the desired audio characteristics.

12. How Is the Threshold Selected?

Threshold selection is critical.

If the threshold is too low:

Too much noise remains.

If the threshold is too high:

Important audio information may be removed.

A commonly discussed threshold is the universal threshold:

λ = σ√(2 ln N)

where:

• σ = estimated noise standard deviation
• N = number of samples

The formula is useful as a starting point, but it is not necessarily optimal for every audio application.

More sophisticated methods can use:

• Level-dependent thresholds
• SURE-based threshold selection
• Bayesian thresholding
• Adaptive thresholds
• Noise-level estimation from high-frequency coefficients

13. Estimating Noise Level

The threshold normally depends on an estimate of the noise level.

One practical approach is to estimate noise from detail coefficients at a high-frequency wavelet level.

A robust estimator can use the median absolute deviation:

σ̂ = median(|d|) / 0.6745

where d represents selected detail coefficients.

The constant approximately normalizes the median absolute deviation for a Gaussian-noise assumption.

This gives the system an estimate of the underlying noise standard deviation.

The estimate can then be used to determine appropriate thresholds.

14. Why the Wavelet Type Matters

Wavelet denoising does not use a single universal wavelet.

Common families include:

• Haar
• Daubechies
• Symlets
• Coiflets
• Biorthogonal wavelets

Different wavelets have different mathematical properties.

For audio signals, useful considerations include:

• Smoothness
• Compact support
• Symmetry
• Number of vanishing moments
• Computational complexity
• Similarity to the signal structure

There is no universally optimal wavelet for every audio application.

The choice should be validated against the actual signal and noise conditions.

15. Haar Wavelet

The Haar wavelet is the simplest wavelet.

It is computationally efficient and easy to understand.

However, its piecewise-constant structure may not model smooth audio waveforms as naturally as some other wavelet families.

It can nevertheless be useful for:

• Educational demonstrations
• Simple embedded implementations
• Signals containing abrupt changes

16. Daubechies Wavelets

Daubechies wavelets are widely used in signal processing.

They provide compact support and multiple vanishing moments, allowing them to represent different signal structures efficiently.

Different orders of Daubechies wavelets provide different time-frequency characteristics.

For audio processing, the appropriate order should be determined experimentally rather than assuming that a higher order is always better.

17. Symlets and Coiflets

Symlets are designed to provide improved symmetry compared with many Daubechies wavelets.

This can be useful when reconstruction symmetry and phase behavior are important.

Coiflets provide another family with useful mathematical properties for representing signal features.

Again, wavelet selection is application-dependent.

The best choice depends on:

• Signal type
• Noise characteristics
• Sampling rate
• Desired reconstruction quality
• Computational constraints

18. Wavelet Denoising vs. Spectral Subtraction

Wavelet denoising and spectral subtraction both attempt to separate useful audio from noise, but they operate differently.

Neither method should be considered universally superior.

They provide different ways to represent and suppress unwanted signal components.

19. Wavelet Denoising vs. Wiener Filtering

Wiener filtering is based on estimated signal and noise power.

Wavelet denoising is based on the distribution of signal information across wavelet scales.

This makes wavelet denoising another useful tool in the overall audio DSP toolbox.

20. Wavelet Denoising for Speech

Speech is particularly interesting because it contains both relatively stable and rapidly changing components.

For example:

• Vowels contain strong harmonic structures.
• Consonants can contain broadband energy.
• Plosives contain short transients.
• Fricatives may contain strong high-frequency components.

An overly aggressive wavelet threshold can remove some of these useful speech components.

Therefore, speech denoising should balance:

Noise suppression

against

Speech preservation

This is especially important when the final application is:

• Speech recognition
• Voice communication
• Meeting transcription
• Voice assistants
• Audio recording

A cleaner waveform is not automatically a better speech-recognition input.

21. Wavelet Denoising for Transient Noise

One potential advantage of wavelet analysis is its ability to localize signal structures in time.

Consider a short impulsive disturbance:

Click → Silence → Speech

A conventional frequency-domain filter may spread the disturbance across a wider time region depending on the analysis method.

Wavelet analysis can identify localized coefficients associated with the transient.

This makes wavelet-based processing interesting for:

• Click removal
• Impulsive interference
• Mechanical transients
• Switching noise
• Short-duration disturbances

However, the exact performance depends strongly on the selected wavelet, decomposition level, and threshold strategy.

22. Decomposition Level

The number of decomposition levels is another important parameter.

If the signal is decomposed too shallowly:

The algorithm may not capture enough multi-scale information.

If it is decomposed too deeply:

Computational cost increases and useful signal information may become distributed across too many scales.

The appropriate level depends on:

• Sampling rate
• Signal bandwidth
• Target application
• Wavelet family
• Noise spectrum

For example, a speech system operating at 16 kHz and a full-band audio system operating at 48 kHz will generally require different considerations.

23. A Practical Wavelet Denoising Pipeline

A typical implementation looks like:

Step 1 — Capture audio

MEMS Microphone → Digital Audio

Step 2 — Normalize / pre-process

Remove DC offset or perform required signal conditioning.

Step 3 — Wavelet decomposition

Apply DWT to obtain approximation and detail coefficients.

Step 4 — Estimate noise

Estimate noise level from appropriate wavelet coefficients.

Step 5 — Calculate threshold

Select a threshold according to the noise estimate and application.

Step 6 — Threshold coefficients

Use hard or soft thresholding.

Step 7 — Reconstruct

Apply the inverse wavelet transform.

Step 8 — Evaluate

Measure:

• Noise reduction
• SNR improvement
• Speech quality
• Transient preservation
• Recognition accuracy

24. Basic Python Example

The following example demonstrates a simple wavelet-denoising workflow using PyWavelets.

import numpy as np
import pywt

def wavelet_denoise(
    audio,
    wavelet="db8",
    level=5
):
    # Wavelet decomposition
    coeffs = pywt.wavedec(
        audio,
        wavelet,
        level=level
    )

    # Estimate noise level from the finest detail coefficients
    detail = coeffs[-1]

    sigma = np.median(
        np.abs(detail)
    ) / 0.6745

    # Universal threshold
    threshold = sigma * np.sqrt(
        2 * np.log(len(audio))
    )

    # Soft thresholding
    filtered_coeffs = [
        coeffs[0]
    ]

    for detail_coeff in coeffs[1:]:
        filtered_coeffs.append(
            pywt.threshold(
                detail_coeff,
                threshold,
                mode="soft"
            )
        )

    # Reconstruct signal
    denoised = pywt.waverec(
        filtered_coeffs,
        wavelet
    )

    return denoised[:len(audio)]

This example demonstrates the basic principle rather than providing a production-ready audio algorithm.

A practical implementation should additionally consider:

• Boundary effects
• Level-dependent thresholds
• Real-time buffering
• Signal scaling
• Quantization
• Latency
• Audio quality evaluation

25. Boundary Effects

Wavelet transforms operate on finite-length signal blocks.

How the algorithm handles the boundaries can affect reconstruction quality.

Possible extension methods include:

• Zero padding
• Symmetric extension
• Periodic extension
• Constant extension

The choice can affect artifacts at the beginning and end of processed blocks.

For offline processing this may be manageable.

For real-time systems, block boundaries need particular attention.

26. Wavelet Denoising in Real-Time Audio

Wavelet denoising can be implemented in real-time, but real-time processing introduces additional constraints.

Engineers must consider:

Block size

Smaller blocks reduce latency but provide less data for analysis.

Decomposition level

More levels require more processing.

Memory

The algorithm needs buffers for wavelet coefficients.

Computational load

The DSP must complete processing before the next audio block arrives.

Reconstruction

The overlap and boundary strategy must avoid audible discontinuities.

Power consumption

This is particularly important in low-power embedded devices.

Therefore:

A wavelet algorithm that works well offline may require substantial optimization before being deployed in a low-power real-time audio product.

27. Wavelet Denoising with MEMS Microphones

Wavelet denoising is not a microphone technology by itself.

It is a signal-processing technique applied after acoustic capture.

A simplified system is:

Acoustic Environment

↓

MEMS Microphone

↓

Analog / Digital Front End

↓

DSP

↓

Wavelet Decomposition

↓

Noise Suppression

↓

Enhanced Audio

The final result depends on both the microphone input quality and the processing algorithm.

A microphone with poor signal-to-noise performance or severe acoustic interference limits how much any subsequent algorithm can recover.

This is why microphone selection and DSP should be considered together.

28. Wavelet Denoising with Microphone Arrays

Wavelet denoising can also be used with multi-microphone systems.

A possible architecture is:

MEMS Microphone Array

↓

Synchronized Multi-Channel Capture

↓

Beamforming

↓

Wavelet Denoising

↓

Speech Enhancement

Alternatively, wavelet processing can be performed independently on each microphone channel before spatial processing.

The optimal architecture depends on:

• Noise type
• Spatial characteristics
• Processing resources
• Required latency
• Number of microphone channels

SISTC’s multi-microphone designs use synchronized digital microphone channels and integrate noise suppression and array signal processing, providing a hardware foundation for more advanced multi-channel audio algorithms.

29. Wavelet Denoising and AI

Wavelet processing and AI can also be combined.

For example:

Audio Input

↓

Wavelet Decomposition

↓

Wavelet Features

↓

Neural Network

↓

Coefficient Enhancement

↓

Wavelet Reconstruction

In this architecture, the neural network does not necessarily operate directly on the raw waveform.

Instead, it can learn how to modify wavelet coefficients.

This approach can combine:

• Multi-resolution signal representation
• Learned nonlinear processing
• Noise classification
• Adaptive coefficient suppression

Such hybrid approaches can be useful when conventional thresholding is insufficient.

30. Wavelet Transform as AI Input

Wavelet transforms can also provide useful features for machine-learning systems.

A neural network may use:

• Wavelet coefficients
• Wavelet energy
• Multi-scale features
• Time-scale distributions

as model inputs.

This can help AI systems distinguish between:

• Speech
• Noise
• Transients
• Mechanical sounds
• Environmental events

The specific benefit depends on the model architecture and training data.

31. Common Engineering Mistakes

Mistake 1: Assuming a higher decomposition level is always better

More levels do not automatically produce better audio.

Mistake 2: Using one threshold for every scale

Different wavelet levels may have very different noise and signal characteristics.

Level-dependent thresholds can therefore be more appropriate.

Mistake 3: Choosing a wavelet arbitrarily

Different wavelets have different properties.

Testing should be based on the actual audio signal and application.

Mistake 4: Using excessive thresholding

If the threshold is too aggressive, speech harmonics and transients may be removed.

Mistake 5: Evaluating only noise reduction

An algorithm can reduce noise while simultaneously damaging speech quality.

Evaluation should consider both.

32. How to Evaluate Wavelet Denoising

A practical evaluation should combine objective measurements with listening tests.

Useful metrics can include:

SNR

Measures the relative level of signal and noise.

STOI

Useful for evaluating speech intelligibility.

PESQ / POLQA

Can be considered depending on the evaluation framework and application. Metric selection should reflect the relevant standards and use case.

Spectral analysis

Compare:

• Original noisy signal
• Denoised signal
• Reference clean signal

Listening tests

Human listening remains important because numerical metrics do not capture every type of artifact.

Application-level performance

For voice systems, also evaluate:

• Speech recognition accuracy
• Wake-word detection
• Transcription quality
• Communication intelligibility

33. Wavelet Denoising vs. AI Noise Reduction

Wavelet denoising has several practical characteristics.

Wavelet-based processing

• Mathematically interpretable
• Relatively lightweight
• Does not require training data
• Suitable for many DSP implementations
• Can provide strong time-scale localization

AI-based processing

• Can model more complex nonlinear relationships
• Can learn from large datasets
• Potentially handles complex non-stationary noise
• Requires training and model optimization
• May require more computational resources

This is not necessarily an either/or decision.

A practical product can combine classical DSP and AI.

34. Where Wavelet Denoising Fits in a Modern Audio Pipeline

A modern audio system could use:

MEMS Microphone

↓

Multi-Channel Capture

↓

Beamforming

↓

AEC / Adaptive Filtering

↓

Wavelet or Spectral Processing

↓

AI Speech Enhancement

↓

AGC / Post-Processing

↓

Voice Recognition / Communication

The exact architecture will vary according to the application.

For example, a low-power embedded voice device may prioritize computational efficiency, while a professional network audio system may have more processing resources available.

35. When Should Engineers Consider Wavelet Denoising?

Wavelet denoising can be worth considering when:

• Noise is non-stationary
• Transient interference is important
• Multi-scale signal characteristics matter
• Time localization is important
• A lightweight DSP approach is preferred
• Spectral-only processing is insufficient

It may be less attractive when:

• The noise is well characterized by simple spectral filtering
• Ultra-low computational complexity is required
• The application already uses a highly optimized AI enhancement model
• Real-time constraints make multi-level processing impractical

Again, the correct choice depends on the entire system.

36. Wavelet Denoising in Audio Engineering: The Bigger Picture

Wavelet denoising is one part of a much larger family of audio noise-reduction methods.

The progression can be viewed as:

Time-Domain Filtering

↓

FFT / STFT Analysis

↓

Spectral Subtraction

↓

Wiener Filtering

↓

Adaptive Filtering

↓

Wavelet Denoising

↓

NMF

↓

Deep Learning

↓

Hybrid DSP + AI

Each technique uses a different representation or assumption about the audio signal.

This is why modern audio engineering is increasingly focused not on finding one universal noise-reduction algorithm, but on selecting the right combination of:

Microphone + Acoustic Design + Spatial Processing + DSP + AI

Frequently Asked Questions

What is wavelet denoising?

Wavelet denoising is an audio-processing technique that decomposes a signal into multiple scales using a wavelet transform, suppresses coefficients associated with noise, and reconstructs the enhanced signal.

How does wavelet denoising reduce audio noise?

It analyzes the audio at multiple time scales and applies thresholding or other coefficient-processing methods to reduce components that are likely to represent noise.

What is wavelet thresholding?

Wavelet thresholding modifies wavelet coefficients according to a selected threshold. Small coefficients may be removed, while larger coefficients are retained or reduced.

What is the difference between hard and soft thresholding?

Hard thresholding removes coefficients below the threshold while preserving larger coefficients. Soft thresholding also reduces the magnitude of coefficients above the threshold, generally producing a smoother result.

Is wavelet denoising better than Wiener filtering?

Neither is universally better. Wiener filtering uses statistical estimates of signal and noise power, while wavelet denoising works in a multi-resolution wavelet representation. Their effectiveness depends on the signal, noise, and application.

Can wavelet denoising be used for speech?

Yes. Wavelet denoising can be used for speech enhancement, but threshold selection must be carefully controlled to avoid removing speech harmonics, consonants, and transient information.

Can wavelet denoising work with MEMS microphones?

Yes. Wavelet denoising can be implemented after a MEMS microphone captures the audio signal. The algorithm can run on a DSP, MCU, application processor, or other suitable processing platform.

Can wavelet denoising be combined with microphone arrays?

Yes. Wavelet processing can be combined with synchronized multi-microphone acquisition and beamforming to provide both spatial and time-scale processing.

Is wavelet denoising suitable for real-time audio?

It can be, but the decomposition level, block size, computational resources, latency, and reconstruction method must be carefully designed.

Can wavelet denoising be combined with AI?

Yes. Wavelet coefficients can be used as features or processed by neural networks as part of a hybrid DSP + AI speech-enhancement system.

Conclusion

Wavelet denoising provides a different way to approach audio noise reduction by analyzing signals across multiple time scales.

Instead of relying only on frequency-domain filtering, it decomposes audio into wavelet coefficients and selectively suppresses components that are likely to represent noise.

Its most important concepts include:

• Wavelet transform
• Discrete Wavelet Transform (DWT)
• Multiresolution analysis
• Wavelet decomposition
• Hard thresholding
• Soft thresholding
• Noise-level estimation
• Adaptive or level-dependent thresholds
• Wavelet reconstruction

For audio engineers, the most important lesson is that wavelet denoising is not simply about choosing a wavelet and applying a threshold.

The final performance depends on:

Wavelet selection + decomposition level + noise estimation + threshold strategy + reconstruction + audio system design

When combined with MEMS microphones, microphone arrays, beamforming, adaptive filtering, and AI speech enhancement, wavelet processing can become another useful component in a broader audio signal-processing architecture.

The next article in this technical series will examine Non-Negative Matrix Factorization (NMF) for Audio Noise Reduction, moving from classical wavelet processing toward a more advanced method for separating structured audio components.