How Krisp noise cancellation works: setup and an A/B check

Krisp sits between your physical audio device and a compatible communication app. Turning it on does not help a call that still records directly from the physical microphone. First verify the audio route, then compare what processing changes.

This article explains the official setup and test guidance reviewed on September 11, 2026. It includes a repeatable listening procedure, not measured noise reduction, latency or a published desk-noise recording. For choosing a physical input first, use the dictation microphone comparison guide.

Microphone and Speaker process opposite directions

According to Krisp's virtual-device explanation, installation creates Krisp Microphone and Krisp Speaker on your computer. They are software devices rather than additional hardware.

For outgoing speech, the route is:

Physical microphone
        ↓
Krisp Microphone
        ↓
Communication app

Krisp says that processing happens on your device before the audio reaches the communication app. For incoming sound, the direction is different:

Communication app
        ↓
Krisp Speaker
        ↓
Physical speaker or headphones

That separation matters when diagnosing a complaint. If another participant hears your keyboard, investigate the outgoing microphone path. If you hear noise from their room, that is the incoming path. Enabling the wrong direction does not establish that the relevant audio was processed.

The support pages describe routing and local processing; they do not provide enough detail to reconstruct Krisp's model architecture or calculate an expected decibel reduction for your room. This guide does not infer those details from the product's AI label.

Configure both sides of the connection

Krisp's application setup guide requires a target app that supports virtual devices and provides suitable audio settings. For the direct setup described there:

1. In Krisp, choose the physical microphone you want to record and the physical speaker or headphones you want to hear.
2. Enable the relevant Noise Cancellation control in Krisp.
3. In the calling app, choose Krisp Microphone as the input and Krisp Speaker as the output for the directions you want processed.
4. Run the app's own audio check and confirm that the selected devices remain correct after reconnecting a headset.

The virtual-device guide says Krisp's input should be a physical device, not another virtual microphone. Avoid building an unexplained chain of virtual devices and then attributing its behavior to Krisp alone.

Browser apps may have separate browser and site device choices. If the expected Krisp device is absent, consult the app-specific setup linked from Krisp's guide. Do not assume every text-entry or recording app can use a virtual microphone merely because a calling app can.

Compare the same recording with Krisp's built-in test

The official Test Noise Cancellation instructions describe opening the profile menu, choosing Test Noise Cancellation, recording with the microphone button and comparing the original with the Krisp toggle enabled. They recommend at least 12–15 seconds with speech present.

Using the same recording helps avoid comparing two different performances. Keep speaking volume and playback volume consistent while listening. Include clear speech with the ordinary sound you want to investigate, such as your keyboard or a fan.

Krisp says this test recording is local and is not sent to or stored on its cloud servers. That statement describes the test feature; it should not be extended to meeting recording, transcription or another app's storage without checking those features separately. The local versus cloud guide explains those boundaries.

Listen for speech damage as well as quieter noise

Download the Krisp A/B listening worksheet. Record the app version, OS, physical mic, selected virtual device, noise condition and which processing controls were active.

Listen for four things:

• Noise remaining: is the keyboard, fan or other target sound still noticeable?
• Words preserved: are sentence openings, quiet syllables and names still understandable?
• Voice changes: does speech become choppy, thin or inconsistent in level?
• Usefulness: would the processed version make the intended conversation easier to follow?

Try a quiet baseline, speech while typing, and a steady background sound as separate recordings. These are proposed conditions, not tests already performed for this article. Keep the raw and processed versions together where the tool allows saving them, and label anything you could only audition inside the app.

If asking someone else to compare recordings, use files you have permission to share and label them A and B without announcing which you prefer. Do not turn a subjective preference into a percentage of noise removed.

Check the result inside the destination app

The built-in test tells you about that test path. A calling app may also process noise, adjust levels or encode the audio. Keep those settings unchanged while comparing Krisp off and on; if you alter them, record a new condition.

For dictation, assess the transcript as well as the recording. Cleaner-sounding audio does not by itself prove fewer recognition errors. Use the same text, language and model, then count meaningful corrections. First confirm that the dictation app receives usable audio through the selected device.

If Flow produces no text after an input change, use the recording and insertion troubleshooting guide. A visible device name is not evidence that the intended speech reached the app.

Keep only the processing that helps the task

If noise is less distracting and speech remains clear in your actual app, keep that configuration and record it. If important syllables disappear or the transcript needs more correction, compare the unprocessed physical input and another mic position before adding more processing.

The useful outcome is an understandable call or usable transcript in your environment. Without saved observations, the worksheet is a way to reach that decision, not evidence that Krisp outperforms another tool.