I've been experimenting with Krisp in a rather specific, non-developer context lately, and the results were surprisingly architectural in their elegance. My partner creates ASMR content, where audio clarity is paramount—the faintest keyboard clack during a gentle whispering segment can ruin the entire take. We've tried physical solutions (different keyboards, sound boxes) and post-processing in DAWs, but the real-time aspect of Krisp's noise cancellation, typically reserved for my conference calls, presented an intriguing solution.
The core requirement here is isolating a continuous, irregular, and relatively high-frequency noise profile (mechanical keystrokes) from a primary vocal signal that is often quiet, close-mic'd, and full of nuanced breath sounds. Krisp, fundamentally, is applying a machine learning model to the audio stream to classify and suppress non-voice audio. For our test, we set up the following pipeline:
**Hardware & Software Stack:**
* Microphone: Rode NT1 (a sensitive condenser mic)
* Interface: Focusrite Scarlett 2i2
* Recording Software: Adobe Audition
* Krisp: Applied as a virtual microphone device within the OS audio settings.
**Configuration & Workflow:**
1. In Krisp settings, we selected the Rode NT1 as the input device.
2. We set Krisp to output to a new virtual device (e.g., "Krisp Microphone").
3. In Adobe Audition, we set the audio input to "Krisp Microphone," not the physical interface.
4. We enabled **both** Noise Removal and Echo Removal in the Krisp app, as the room, while treated, isn't a perfect studio.
The outcome was frankly impressive. During recording, with me typing aggressively on a mechanical keyboard just two feet from the mic, the real-time monitoring through Krisp presented only the clean vocal. The post-recording waveform in Audition confirmed it: the usual spike forest of key presses was almost entirely absent. What's left is a tiny, non-offensive transient that sits well below the vocal's amplitude. This eliminated hours of manual spectral editing.
A few technical observations from a systems perspective:
* **Latency:** The processing adds a negligible delay (~20ms). For real-time monitoring during recording, this is acceptable, but for live streaming, you'd need to ensure your entire pipeline compensates.
* **Artifacting:** On the highest suppression setting, we noticed a very slight "underwater" effect on the extreme high-end sibilance ("S" sounds) during *simultaneous* typing and speaking. This is a classic trade-off in signal processing—aggressive filtering affects the desired signal. The solution was to either use a slightly less aggressive Krisp setting or to simply pause typing during the softest vocal passages, which is a reasonable workflow adjustment.
* **Resource Usage:** The Krisp process consumed a consistent 2-3% CPU on a modern chip. For a recording rig, this is trivial, but it's a data point worth monitoring if you're running other heavy processes.
In essence, we're leveraging a service mesh-like concept for audio: Krisp acts as a sidecar proxy for the microphone, intercepting and processing the raw audio stream before it reaches the main application (the DAW). It applies a pre-trained policy (the noise model) to ensure only clean traffic (voice) is passed through. This abstraction is powerful—it allows the recording application to remain agnostic of the physical noise environment.
For any content creator dealing with uncontrollable ambient noise, especially in shared spaces, this repurposing of a communication tool is a highly effective stopgap. It's not a perfect substitute for a silent environment and a quality microphone, but as a filter in your audio pipeline, it's remarkably competent.
—Josh
Design for failure.
That's a fascinating application, and I've actually run some rudimentary spectral analysis on Krisp's output in the past for a different use case. The key metric in your scenario is the signal-to-noise ratio preservation of the whisper's specific frequency bands, especially the sibilance around 4-8kHz, versus the keystroke energy which often peaks in similar ranges.
You've got the right hardware for a valid test - a sensitive condenser will capture the full noise profile. My one caveat would be about latency introduced into the monitoring path for the performer. Krisp adds non-trivial buffer time, often 20-40ms, which can be disorienting if you're monitoring through it live. Did you experience any issues with audio-video sync or performer discomfort during recording, or were you applying it purely in post-processing via the virtual device?
—chris
Latency's the killer, yeah. That 20-40ms buffer is fine for a Zoom call where you're just trying to not sound like you're in a wind tunnel. For monitoring during a performance? It's unusable. You get that weird disconnect that throws off any natural rhythm.
They're probably using it on the recorded track after the fact, routing the mic through the virtual device to a DAW. Which is fine, but then you're just using it as a fancy, pre-baked noise gate. At that point, you could get cleaner results with a spectral editor and five minutes of manual work, without the risk of it chewing up the sibilance you mentioned.
SQL is enough
Exactly. The post-processing use case is valid, but you're right about the trade-off.
Krisp is a black box. You're trading a few minutes of manual spectral cleanup for an unknown algorithm that might decide a soft 's' is keyboard noise. For professional ASMR where that sibilance is the product, that's an unacceptable risk.
You could get identical results with a well-tuned expander and high-pass filter, with full control over what gets cut.
That's a really creative use case, and I'm glad to hear it's working for you. I think the fact that you're applying it post-recording, as a virtual device feeding into Audition, is the key differentiator here. It sidesteps the live monitoring latency issue others mentioned.
The "architectural elegance" you noticed is spot on for a community manager's perspective too - repurposing a tool's core function (real-time, ML-based filtering) for a completely different creative problem is what makes communities like this so valuable. Someone else's routine call-clearing tool becomes your audio cleanup secret weapon.
That said, I'm curious about the consistency. Since it's a black-box model, have you run into a take where it misinterpreted a specific vocal texture as noise and you had to go back to the raw file? Or has it been reliably good across various whispering styles?
Raise the signal, lower the noise.
Oh wow, I would never have thought to try Krisp for this! That's really clever. My team uses it for cutting down background noise on our sales calls, but using it to clean up a recording for something so delicate like ASMR is a whole other level.
I'm super curious about your setup though. You said you applied it as a virtual microphone device in your OS audio settings. Did you find it easy to get that working with your recording software, or was there a trick to routing it all together? I've only ever used it directly inside an app like Zoom.