Audio & Acoustics · Comparison
Short answer: Classical noise suppression (spectral subtraction, Wiener filtering, adaptive filters) estimates the noise from the signal itself and removes it with very little code, memory and delay. It behaves predictably, but it assumes the noise changes slowly, so keyboard clicks, babble and other fast-changing sounds get through. Neural noise suppression learns what speech and noise sound like from large datasets and handles changing noise much better, at the cost of more compute, a training pipeline and less predictable behaviour on sounds unlike its training data. Algorithmic delay is set mainly by frame length and look-ahead, not by which family you pick. Many shipping designs are hybrids: classical processing where there is a physical model or a reference signal, and a small network for the decisions a rule can’t make.
A guide for audio and embedded teams choosing a single-channel noise suppressor for headsets, conferencing devices, intercoms or industrial communication. We compare the two families on delay, cost, robustness and how they fail.
Side by side
Both families usually work in the same place: a short-time Fourier transform, a gain for each time-frequency bin, and an inverse transform. They differ in where the gain comes from.
| Criterion | Classical DSP | Deep learning |
|---|---|---|
| How it works | Tracks a noise spectrum during pauses or with minimum statistics, then applies spectral subtraction or a Wiener-type gain. Adaptive filters such as NLMS instead cancel noise that is correlated with a separate reference signal. | A network trained on pairs of noisy and clean speech predicts a mask, a set of band gains or the clean signal itself, in the time-frequency domain or directly on the waveform. |
| Latency | One analysis frame plus overlap, typically 10 to 30 ms of algorithmic delay. Time-domain adaptive filters add almost none. | The same frame delay for a causal streaming model. Look-ahead frames improve quality and add delay one for one. Inference must also finish within each hop. |
| Compute & memory | An FFT and a few vector operations per frame, with a few kilobytes of state. Fits a small fixed-point DSP or microcontroller. | Typically from tens of thousands of parameters for hybrid band-gain models to several million for full-band models. On a device it usually needs quantization and a streaming architecture. |
| Data needs | None for training. Tuning needs recordings from the device and structured listening. | Large sets of clean speech, noise and room responses for training, plus recordings from the target device to check for mismatch. |
| Steady noise | Strong. Fans, hum and hiss match the slowly-changing-noise assumption. | Strong, though for steady noise alone the extra cost buys little. |
| Fast-changing noise | Weak. The noise estimate lags behind typing, clatter, babble and music, so they leak through or pull speech down with them. | Much better, because the model recognizes speech itself instead of waiting for the noise to hold still. |
| Unseen conditions | Makes few assumptions about the noise type, so it degrades gradually and in ways an engineer can predict. | Can fail abruptly on noise, languages, levels or microphones unlike the training data. Wide training data and augmentation reduce the risk; they don’t remove it. |
| Typical artefacts | Musical noise, a residual that pumps with the speech, and muffled speech when suppression is pushed hard. | Missing consonants or whole syllables at low signal-to-noise ratio, and speech from a nearby talker or a TV kept as if it were the user. |
This compares general properties of the two families. Figures are typical orders of magnitude, not measurements; how much either one helps on a given product has to be measured on the device, with its microphones and its noise.
When each one wins
The deciding questions are how fast the noise changes, how much compute each frame can spend, and how predictable the product must be.
The noise is mostly steady, the processor is a small DSP or microcontroller, or the product needs behaviour that can be explained and bounded, as in hearing, safety or intercom devices. An adaptive filter is the right tool whenever a clean reference of the noise or the echo is available.
The noise changes quickly (typing, babble, machinery that starts and stops), the platform can run a quantized network on every hop, and the team can maintain a training set and an evaluation suite.
Compute is limited but classical quality isn’t enough. Common patterns: a small network predicts band gains or the noise estimate and a classical filter applies them; a linear echo canceller or a beamformer runs first and a network cleans the residual; a classical gain floor limits how much the network may remove.
How to choose
A decision list we’d work through before committing to an architecture.
Record the product’s real conditions and sort the noise into steady and fast-changing. If nearly all of it is steady, a well-tuned classical suppressor may be enough.
A call, a hearing device and a speech recognizer tolerate different delay. The budget sets frame length and look-ahead for both families.
Know the cycles and memory available per hop on the target processor, after the rest of the audio path has taken its share.
If the device has a loudspeaker signal, a second microphone or a vibration sensor, use adaptive filtering or beamforming for what it can remove before any single-channel suppressor.
Run a classical baseline and a neural candidate through one evaluation: objective quality and intelligibility scores, word error rate if a recognizer listens, and listening tests on noise held out from training.
Before we start
No. On fast-changing noise a well-trained network usually sounds clearly better. On steady noise the difference is small, and a classical suppressor costs far less and behaves more predictably. The answer depends on the noise, the processor and how much unpredictability the product can accept.
Not necessarily. In both, algorithmic delay comes mainly from the analysis frame and any look-ahead. A causal neural model with the same frame has the same algorithmic delay, provided inference finishes within each hop. Models that use future frames trade delay for quality.
Small ones can. Hybrid designs that predict a few dozen band gains need little compute and have run on embedded processors for years. Larger full-band models usually need a faster core or a neural accelerator. Our on-device audio AI page covers the constraints.
It is the artefact most associated with spectral subtraction: isolated time-frequency bins survive the subtraction at random and are heard as short, watery tones. Smoothing the gain over time, a spectral floor and decision-directed estimation of the signal-to-noise ratio all reduce it. Our spectral subtraction demo lets you hear it.
Keep whole noise types, rooms, talkers and devices out of training and report results on those separately. Add recordings from the real product, including conditions with no speech at all, and listen to the worst cases instead of only averaging scores.
Keep exploring
Audio & Acoustics · Deep Learning vs Classical DSP Noise Suppression
Share the problem, the data you have and what success would look like. We’ll discuss a practical next step.
Please don’t include confidential datasets, credentials or patient information.
yochai@mlaia.com
WhatsApp: +972 52 484 6282