What is Noise Suppression?

Noisy calls hurt trust. Customers ask us to repeat. Agents feel stress. We can fix this with the right suppression and gear. The steps are clear and practical.

Noise suppression in audio processing ¹ removes unwanted background sounds so speech is clearer. It works in real time, on the device or in the cloud, and it is not the same as active noise cancellation for headphones ².

In this guide, the key ideas come first, then tools and settings you can use today. Each section ends with simple tables you can act on. The aim is clarity, not hype.

What is noise suppression?

Open-plan offices add chatter, keyboards, HVAC hum, and traffic. The customer hears all of it. Calls feel messy. We need the mic to pass voice and reduce the rest.

Noise suppression estimates non-speech sounds and turns them down. It keeps speech as natural as possible. It runs before encoding so the network carries a cleaner signal.

Core concept in one line

Noise suppression separates “speech” from “not speech.” It reduces the latter and preserves the former. It is a front-end speech enhancement step, not a comfort feature like ANC.

How it differs from active noise cancellation

Active noise cancellation uses microphones and speakers to cancel ambient sound for the listener. It protects the ear. Noise suppression filters the microphone signal for the remote party. It protects the call. These are different paths in the chain, and they can work together without conflict.

Where suppression runs

Suppression can sit in several places. It can run inside a headset’s DSP. It can live in the OS audio stack. It can run in the browser via WebRTC audio processing ³. It can run in the softphone or UC client. It can also run server-side in your media platform. The earlier it runs, the less the encoder wastes bits on noise.

Typical noises and how suppression treats them

Constant hum from HVAC or fans is easy to estimate. Key clicks, door slams, or barking dogs are harder. Speech-like noise from nearby talkers is hardest. Good systems adapt to changing rooms. They track energy and spectra frame by frame and apply gain that spares vowels and consonants.

Quick reference

How do DSP and AI noise removal compare?

Classic DSP is fast and predictable. AI sounds more “magical.” Both have trade-offs. The right choice depends on latency, CPU, noise type, and privacy rules.

DSP methods rely on rules like spectral subtraction and gating. AI methods learn patterns to separate speech and noise. AI often sounds cleaner, but it needs more compute and careful tuning.

What “DSP” usually means here

Traditional approaches include spectral subtraction and Wiener filtering ⁴, and noise gating. They estimate a noise floor, then subtract or attenuate it in the frequency domain. They are small, fast, and stable. They introduce “musical noise” when estimates are wrong. They shine on steady noise.

What “AI” means in practice

Modern systems use small neural networks for speech enhancement ⁵ trained on mixed audio. They predict a mask that keeps speech and suppresses noise. They handle non-stationary sounds better, like dishes, typing, or kids. They can retain the timbre of the voice. They may add slight latency and need CPU or a small NPU. Lightweight models can run in real time on laptops and phones.

Latency and compute budget

Some DSP pipelines add under 5 ms of algorithmic delay. Neural models often add 10–20 ms or more, depending on frame size and lookahead. On contact center desktops, this is fine. On hard real-time intercoms or radio links, keep budgets tight. Measure end-to-end, not just the filter.

Artifacts and speech quality

Over-aggressive DSP can leave metallic “chirps.” Over-aggressive AI can smear consonants or pump ambience. Both can reduce ASR accuracy when they distort sibilants and plosives. A balanced setting often beats “max” noise removal.

Privacy and deployment

On-device DSP keeps audio local. Cloud AI may send media to a service. Many teams prefer browser or client-side AI to avoid data transfer. Check your policy and your customer contracts.

Side-by-side summary

Which headset and mic choices help my agents?

The best algorithm cannot fix a bad microphone. Source quality wins. A good headset makes every downstream step easier and cheaper.

Choose a noise-rejecting mic with proper placement, stable USB or DECT transport, and fit that agents keep on all day. Favor cardioid or dual-mic beamforming near the mouth.

Mic type matters more than brand

A boom mic that sits two finger-widths from the corner of the mouth beats a pretty earbud from any brand. Cardioid dynamic capsules reject room sound. Dual-microphone arrays with beamforming also work well when tuned for speech. Avoid far-field mics for busy floors. They hear the room first, the voice second.

Wired vs. wireless

USB headsets remove analog noise from the PC jack and are easy to standardize. DECT wireless headsets for offices ⁶ have longer range and stable links in offices. Bluetooth is fine when you lock codecs and disable auto EQ. Latency varies. Test call-control features with your softphone.

Fit, wear, and hygiene

Agents keep what feels good. Light clamp force and soft pads reduce fatigue. Replace cushions on a schedule. Keep spare windshields for booms. Train agents to place the mic off-center to avoid breath pops.

Practical selection table

Room and desk tweaks that cost little

Soft desk mats cut keyboard impact noise. Small desk shields reduce direct path from neighbors. A short rug under the chair reduces wheel noise. These small steps lower the load on any algorithm and raise call quality fast.

How do I tune suppression without clipping speech?

Too much suppression sounds clean but wrong. Words clip. Callers say “you faded out there.” We want quiet calls and natural voice at the same time.

Start with moderate strength, confirm levels and AGC, then adjust gate, attack, and release. Use test phrases with sibilants and plosives. Track MOS and handle rates, not just “sounds good.”

A simple, repeatable tuning plan

1. Set levels first. Aim for −18 dBFS RMS with peaks near −6 dBFS on speech. Disable double AGC. One gain control is enough.
2. Choose a moderate profile. Many tools have Low/Medium/High. Start at Medium.
3. Protect consonants. Use test lines like “six sharp swings” and “please pack the box.” If consonants blur, lower strength or shorten lookahead.
4. Tame only the floor. Raise the noise gate threshold only until room tone dips. Keep hold/decay gentle so soft words survive.
5. Watch post-processing. A compressor after suppression can re-raise noise between words. Use a slow release and modest ratios.
6. Measure, do not guess. Log double-talk loss, clipping count, and customer “couldn’t hear you” tags.

Key parameters and what they do

Test once, then codify

Create a 60-second test script that includes quiet speech, fast speech, sibilants, and plosives. Record in a quiet room and a busy room. Save waveforms and settings. Compare objective metrics like SNR improvement and short-term PESQ, then listen. Roll the winning preset to all agents through device management so settings stay consistent.

Does network jitter affect perceived noise?

Yes, the network can “sound” like noise. Packet loss and jitter fill the gaps with PLC. Listeners hear warbles, mutes, and high-frequency grit.

Jitter does not add room noise, but it harms clarity. A jitter buffer trades delay for continuity. Good suppression cannot fix missing packets, so fix transport first.

What changes in the ear

When packets arrive late or drop, the decoder uses packet loss concealment in VoIP codecs ⁷. It reuses old frames or predicts new ones. This sounds like robotic warble or a sandpaper hiss. Customers call it “static” or “noise,” even though it comes from the network, not the room.

Jitter buffer tuning

Most UC stacks adapt buffers between 20–120 ms. Too small, and audio gaps appear. Too large, and agents talk over each other. If your calls span regions, allow a slightly larger target jitter buffer. Keep end-to-end mouth-to-ear under ~200 ms for natural turn-taking. For local contact center traffic, aim lower.

Transport basics that pay off

Use wired Ethernet for agent PCs when possible. Prefer QoS with DSCP for real-time media. Keep Wi-Fi on 5 GHz or 6 GHz with strong RSSI and low channel load. Close heavy background sync tasks. Split media and screen share where the platform supports it. These steps reduce concealment events that people hear as “noise.”

How suppression interacts with network issues

Suppression cannot restore lost packets. It only shapes the signal that makes it to the encoder. In some cases, strong suppression reduces codec complexity and helps PLC a bit because the input is cleaner. But if loss is over a few percent, transport wins over any extra suppression.

Quick checklist

Conclusion

Noise suppression helps most when source audio is solid and the network is steady. Pick good mics, set moderate strength, and fix transport first. Then fine-tune.

Footnotes