Enterprise Bot - Blog

Noise cancellation for voice AI: comparing options

Written by Enterprise Bot | Nov 19, 2026, 12:52:47 PM

Background noise is one of the quietest ways a voice bot fails. Here is how the main options compare, and why the cleanest-sounding audio is not always the most accurate. 

If your voice AI mishears people, noise is often the reason. Noise cancellation removes background sound before speech recognition runs, and your options range from open-source models like RNNoise and DeepFilterNet, to commercial SDKs like Krisp and NVIDIA, to denoisers built straight into the big speech APIs.

Here is the part most teams miss. The right option is not the one that sounds cleanest. It is the one that lowers recognition errors on your own calls. Those are not always the same thing, and the gap between them is what this piece is about.

Why noise cancellation decides more than it looks

When a voice bot gives a wrong answer, the model usually gets the blame. Often the model is innocent. It answered the wrong question because it was handed the wrong words, and it was handed the wrong words because the audio was noisy.

Noise cancellation sits right at the top of the pipeline, before speech-to-text, before intent, before the response. Everything after it inherits whatever it does. Clean the signal well and accuracy rises across the whole chain. Clean it badly and you push a hidden error into every step that follows. That is why this unglamorous stage deserves more attention than it usually gets.

The options, side by side

Here are the main options a contact centre actually weighs up.

Option

Approach

Where it runs

Latency added

Licence

Krisp

Neural noise and background-voice removal

Server SDK or on device

15 ms (vendor)

Commercial

NVIDIA Maxine

Neural denoiser, plus echo and reverb removal

GPU container or hosted

Not published

Commercial

RNNoise

Small hybrid DSP and neural network

Any CPU, on device or server

20 ms window

Open source (BSD)

DeepFilterNet 3

Two-stage deep filtering

CPU, real time

40 ms

Open source (MIT / Apache)

Picovoice Koala

Proprietary neural model

On device

Not published

Commercial, free tier

Microsoft Audio Stack

DSP: beamforming, echo, gain, noise

Client, inside the Speech SDK

Not published

Bundled

Google Cloud STT denoiser

Server-side denoise before transcription

Inside the API

Not published

Part of STT pricing

OpenAI Realtime noise reduction

Filter before voice detection and the model

Inside the API

Not published

Part of the API

Agora noise suppression

Neural, selectable modes

Client SDKs

Configurable

Commercial

Twilio audio processor

Plug-in point for Krisp or RNNoise

Browser SDK

Per engine

Per engine

Three groups, roughly. The open-source models, RNNoise and DeepFilterNet, are free and run on a CPU, which makes them easy to pilot. The commercial SDKs, Krisp, NVIDIA, Picovoice and Agora, add polish like background-voice removal for a licence fee. And several speech platforms, Google, OpenAI and Microsoft, now denoise inside the API, which is convenient but gives you less say over where the cleaning happens.

Two you might expect to see are missing. Dolby.io's communications APIs have been retired, and we found no publicly documented AI noise-suppression feature in AudioCodes VoiceAI Connect, so we left both out [CONFIRM].

How they get measured, and what the numbers say

Most comparisons point to Microsoft's Deep Noise Suppression Challenge, which ran five times between 2020 and 2023 and became the reference test. The 2023 edition scored systems on human-rated quality and on word accuracy from speech recognition, and the real-time tracks held them to 40 milliseconds or less of added latency. Two other measures come up a lot: PESQ and STOI, which need a clean reference recording to compare against, and DNSMOS, which estimates human ratings with a neural model.

A few published results are worth knowing:

  • DeepFilterNet 3 reports PESQ 3.17 and STOI 0.944 on the VoiceBank and DEMAND test set [CONFIRM].
  • RNNoise beats a classic DSP suppressor on babble, car and street noise, though its paper publishes the result only as a chart [CONFIRM].
  • Picovoice reports a larger STOI gain than RNNoise at 0 dB on DNS data, on its own benchmark [CONFIRM].
  • Krisp reports far fewer false voice-detection triggers with background-voice cancellation, again its own data [CONFIRM].

Keep one thing in mind through all of it. Most of these scores measure how clean the audio sounds to a person, which is not the same as how well a recogniser reads it. And NVIDIA, Agora, Microsoft and Google do not publish quality or latency figures for these features at all, so until you test them, you are taking those on trust.

The catch: cleaner audio is not always better transcription

Here is the finding that should change how you choose. Researchers have shown that cleaning up a single audio channel often barely improves recognition, and sometimes makes it worse, because the cleaning leaves small artefacts the recogniser trips over (Iwamoto and colleagues, Interspeech 2022) [CONFIRM].

Krisp's own data shows the same pattern. Background-voice cancellation helped the system work out when someone was speaking, but the word error rate roughly doubled once the recogniser transcribed the processed audio [CONFIRM]. In other words, it sounded cleaner and read worse.

So the question to ask is not how clean the audio is. It is whether the option improves turn-taking and recognition on your calls. And that depends on where you actually use the cleaned audio: to detect that someone is speaking, to transcribe what they said, or both. A model that is great for the first job can quietly hurt the second.

How to choose

Four things decide it in practice.

  1. Latency. Stay within a few tens of milliseconds. The whole reply has to fit inside a natural pause, so every millisecond the denoiser adds is taken from the time the rest of the system has to answer.
  2. Background voices. Contact-centre callers sit in open offices, in cars, and at home with the television on. Models that suppress other voices, not just steady noise, matter most for clean turn-taking.
  3. Where it runs. Server-side, inside the voice pipeline, covers every caller on every device. Client-side cleaning only helps the callers using your own app.
  4. Test on your own calls. This is the one that counts. Measure voice detection and word error rate on your recordings, with and without suppression, and do it per language.

What we do

Our voice pipeline uses a speech-focused neural noise model, chosen for two reasons. It adds very little latency, and it suppresses background voices as well as steady noise. We do not ask you to take that on faith. In a pilot we run it on your own recordings, per language, and show you the effect on turn-taking and transcription, so you can see the difference on your calls before you commit to anything.

In production, that upstream work shows up downstream. A Swiss health insurer sees 98 percent routing accuracy, confirmed by agents after each call. Generali Switzerland cut misdirected calls by 22 percent. And after a year in production, MONETA Money Bank handles 66 percent of calls with no human support [CONFIRM]. Cleaner input is not the headline in any of those numbers, but it is what makes them possible.

Common questions

Does noise cancellation improve speech recognition? Not automatically. It can, but aggressive cleaning can also add artefacts that raise the error rate. The only way to know is to test it on your own audio.

Does noise cancellation add latency? Yes, every option adds some. The good ones keep it to tens of milliseconds, which matters because the reply has to land inside a natural pause.

Should it run on the phone line or in the AI? Running it in the voice pipeline, server-side, covers every caller on every device. Cleaning on the line or in the app only helps the callers on that one path.

The one test that settles it

Every option on the list can be made to look good on someone's benchmark. None of that tells you what it will do to your calls, in your languages, with your callers' background noise. The only comparison that settles it is the one you run on your own recordings, measuring recognition and turn-taking, with the cleaning on and off.

If you want a straight answer for your own setup, send us recordings from your noisiest lines and we will show you the effect on turn-taking and transcription. No guesswork, just your calls.

Sources

All figures come from the original brief and should be confirmed against the primary source before publishing.

  • Microsoft Deep Noise Suppression (DNS) Challenge, 2020 to 2023 (benchmark, metrics, 40 ms latency bar)
  • Iwamoto and colleagues, Interspeech 2022 (single-channel enhancement can worsen recognition)
  • Krisp, background-voice cancellation data (voice detection vs. word error rate)
  • DeepFilterNet 3, Schröter and colleagues, 2023 (PESQ 3.17, STOI 0.944 on VoiceBank and DEMAND)
  • RNNoise, Valin, 2018 (noise suppression results)
  • Picovoice (STOI gain vs. RNNoise, vendor benchmark)
  • Customer results: 98 percent routing accuracy (Swiss health insurer), 22 percent fewer misdirected calls (Generali Switzerland), 66 percent of calls handled with no human after a year (MONETA Money Bank)