Back
Back
Voice Isolation 2.5 separates the person speaking from everything around them, so your STT gets one clean voice.
Voice Isolation 2.5
Same STT engine with and without Voice Isolation
Raw microphone
As you see, those words Kumquat. keep on popping up in the Juxtapose. background. And we don't see them over here. Zucchini.
3 interfering words
After Krisp Voice Isolation
As you see, those words keep on popping up in the background. And we don't see them over here.
0 interfering words
3 interfering words removed
Why agents break
Three things go wrong, and all three start in the audio, not in your LLM.
Modern speech to text handles noise well. Another person talking is a different problem. The engine transcribes both voices and hands your agent the mixture.
Agents wait for silence to decide the turn is over. People pause mid thought. The result is dead air on one side and talk over on the other.
To an agent, a listener saying mhm sounds like an interruption. So it stops talking when nobody asked it to.
Integration
VIVA runs first, in front of your speech-to-text. Everything after it inherits a clean signal.
Runs where you run
What's in VIVA SDK
Each model does one job well. Run one, or run them together.
Separates the person speaking from everything around them: competing voices, background chatter, room echo. Your STT gets one clean voice, shaped so the engine reads it the way it was trained to.
Predicts when a speaker is done, straight from raw audio. No transcription needed. Kills the awkward pause and the talk-over. Multilingual out of the box.
Classifies speech that lands mid-response: real interruption or a "mhm," takeover or backchannel. Tells your agent when to stop, and when to keep going.
Speech or silence, accurately, and robust against background noise and secondary voices, so you get fewer false triggers.
Accent, gender and synthetic-speech detection in real time. The read on a caller that humans take for granted, as a signal you can route on.
New · Voice Isolation 2.5
We measured this the way a skeptic would. 1,685 real-world recordings, three conditions, ten STT engines from seven vendors, nothing tuned to any single engine.
Voice Isolation has one job: keep speech-to-text working when the audio is bad.
The main model
Strongest at removing noise and other speakers. Goes fully silent when only a bystander is talking.
Lighter processing
Removes noise and background voices with lighter processing while the main speaker is talking, so the voice sounds more natural.
A third of the compute
Comparable quality on average at about a third of the compute. For edge, embedded and high-concurrency deployments.
32 kHz
Built for 32 kHz high-fidelity streams such as WebRTC or WebSocket, rather than narrowband telephony codecs.
Two sizes
Most teams cap isolation to the calls they think need it. Now both models are light enough to leave on.
New: 70% less compute
Matches the default on typical calls, with a gap only in extreme conditions. Built for high volume server fleets, edge, mobile and embedded.
Choose it when
CPU is the constraint
With FP16*
209 → 358 live calls
The default
Strongest in the hardest conditions: extreme noise, dense background voices, badly degraded streams.
Choose it when
Worst-case quality matters most
With FP16*
76 → 138 live calls
New: CPU optimization
One SDK flag runs Voice Isolation in 16-bit math on CPUs built for it. Same model, same quality, about twice the live calls per server.
On Ice Lake (c6i, m6i, r6i)? Move to Graviton3 (c7g) or Sapphire Rapids (c7i) and turn the flag on.
See the full benchmarkLive calls per server
*FP16 figures are for these CPUs, with the SDK flag on. Same model and quality.
Integration
One call in front of your speech to text. Nothing downstream changes.
Tell us your use case and we'll set you up with the SDK and a trial to start.
Customers
“At scale, the biggest challenge in voice AI isn’t the model. It’s the quality of the signal going into it. Krisp addresses that at the source, which improves everything downstream from transcription to response.”
“The Krisp VIVA noise reduction, voice isolation, and turn taking models are a big unlock for production voice agents. Highly, highly recommended.”
“Our users expect world-class quality from all of their communication channels within Discord, and especially with their voice communications. Integrating Krisp’s AI-based Noise Cancellation technology into our desktop and mobile applications was seamless.”
“We are thrilled to partner with Krisp, a leader in Voice AI, to provide exceptional audio quality to the billions of conversations on the Twilio Video platform, allowing customers to build engaging virtual experiences across telehealth, education, remote work, and more.”
“Integrating Krisp eliminates one of the biggest challenges—unwanted interruptions—ensuring seamless, uninterrupted conversations.”
“By integrating Krisp’s advanced AI Noise Cancellation technology, we’re enhancing physician dictation accuracy and transforming telemedicine communication, reinforcing Aarista’s position as a leader in healthcare communication and patient care.”
“Roam Inventors are working around the clock to deliver the #1 Cloud HQ meeting experience. Our new partnership with Krisp means our members will get the best available noise cancellation on the planet.”
“Krisp is a meaningful part of the equation to improve audio quality for us — and importantly, it just works. That’s exactly what we need from an infrastructure partner.”
“oVice customers can now enjoy more comfortable conversations in the business metaverse with access to the world's best AI noise-cancellation technology.”
“To ensure exceptional customer experiences as we scale, we’re integrating Krisp’s best-in-class noise cancellation technology to empower Vodex to reach new heights in delivering clear and uninterrupted voice interactions.”
Tell us your use case and we'll set you up with the SDK and a trial on your audio.
