CtrlK
Krisp AI (formerly 2Hz) is an artificial intelligence company that develops real-time voice AI products, including noise cancellation, accent conversion, voice translation, meeting transcription, and AI-powered meeting notes. Founded in 2017 in Yerevan, Armenia, by Davit Baghdasaryan and Artavazd Minasyan, the company is headquartered in Berkeley, California. Krisp's core technology uses deep neural networks to separate human speech from background noise in real time, processing audio entirely on-device without sending voice data to external servers. As of 2025, Krisp's noise cancellation technology has been deployed on over 200 million devices and processes more than 75 billion minutes of voice conversations per month. The company employs approximately 322 people across four continents and serves enterprise customers including Siemens, Autodesk, Sony, Cisco, and Okta.
Krisp was founded in 2017 under the name 2Hz by Davit Baghdasaryan and Artavazd Minasyan. Baghdasaryan had previously spent nine years working at Twilio in San Francisco, where he experienced firsthand the problems caused by background noise during conference calls. Minasyan holds a PhD in mathematics and had been conducting research in machine learning. The two co-founders, drawing on their backgrounds in physics and mathematics, set out to build a software solution that could remove background noise from voice calls using deep learning.
After approximately six months of research and development, Minasyan delivered the first working noise cancellation prototype to Baghdasaryan. The initial team included Stepan Sargsyan as Chief Scientist. Despite strong foundations in physics and mathematics, the founding team had limited prior experience in machine learning and digital signal processing (DSP), which required significant learning and experimentation during the early stages of product development.
To train their deep neural network, the team collected and processed 20,000 unique noise samples and 10,000 clear voice recordings from individuals of diverse ages, genders, and ethnic backgrounds, totaling over 2,500 hours of audio data. This dataset formed the foundation for krispNet, the company's proprietary deep neural network model for real-time noise suppression.
In 2018, the company (still operating as 2Hz) was accepted into Berkeley SkyDeck, the startup accelerator at the University of California, Berkeley. 2Hz became the first Armenian startup to participate in the SkyDeck accelerator program. The team completed the six-month program in May 2018.
During this period, the company raised its initial funding: $500,000 from SkyDeck along with friends and family in a pre-seed round, followed by a $1.5 million seed round led by Sierra Ventures and Shanda Group. The company later rebranded from 2Hz to Krisp, a name chosen to reflect the clarity and crispness of the audio its technology produced.
In December 2018, Krisp won the ProductHunt Golden Kitty Award for "Audio & Video Product of the Year," marking the product's breakout moment with the consumer tech community. The win attracted significant media attention, including coverage from TechCrunch and other major technology publications.
By 2019, Krisp had attracted 30,000 users from more than 200 countries and had cleaned over 25 million minutes of noisy audio. The product initially pursued a B2B licensing model, seeking partnerships with major conferencing platforms such as Skype, Zoom, and WebEx. However, the company eventually pivoted to offering a direct consumer application, which proved more effective for driving user adoption.
The COVID-19 pandemic in 2020 dramatically accelerated demand for Krisp's technology. With millions of workers shifting to remote work, background noise during video calls became a widespread problem. Krisp's ability to run locally on any device and work with any conferencing application positioned it as a versatile solution for individuals and enterprises alike.
In August 2020, Krisp closed a $5 million Series A funding round, with investments from Storm Ventures, Sierra Ventures, TechNexus, and Hive Ventures. The company used the funding to expand its engineering team and accelerate product development.
The year 2020 proved to be a landmark year for Krisp in terms of industry recognition:
| Award | Organization | Year |
|---|---|---|
| 100 Best Inventions | TIME Magazine | 2020 |
| AI 50: America's Most Promising AI Companies | Forbes | 2020 |
| Cloud 100 Rising Stars | Forbes | 2020 |
| Cool Vendors in Digital Workplace Programs | Gartner (Top 3) | 2020 |
| Golden Kitty Award (Audio & Video Product of the Year) | ProductHunt | 2018 |
| People's Voice Winner (Productivity & Collaboration) | Webby Awards | 2021 |
TIME Magazine named Krisp one of the 100 Best Inventions of 2020 in the Artificial Intelligence category, citing its ability to filter out background noise from calls using deep learning. Forbes recognized Krisp on both its AI 50 list of America's most promising artificial intelligence companies and its Cloud 100 list as a Rising Star. Gartner ranked Krisp in the top three Cool Vendors for Digital Workplace Programs and Applications.
By the end of 2020, Krisp's team had grown to 90 employees, having added 56 new hires (internally referred to as "Krispions") during the year.
In 2021, Krisp secured a $9 million Series A-1 round to further expand its enterprise offerings and invest in new voice AI capabilities beyond noise cancellation. Additional investors included RTP Global, Granatus Ventures, and others. By this point, the company had raised a total of approximately $15.5 million in venture capital.
Krisp also won two Webby Awards in 2021, including the People's Voice Winner in the Apps and Software category for Productivity and Collaboration.
During 2022 and 2023, the company expanded its product lineup significantly, moving beyond pure noise cancellation to become a comprehensive voice AI platform. New features included meeting transcription, AI-generated meeting summaries, action item tracking, and an AI meeting assistant that worked without requiring a bot to join the call.
In June 2024, Krisp launched an Early Access program for its AI Accent Conversion technology, designed primarily for the contact center and BPO (business process outsourcing) industry. The technology dynamically adjusts a speaker's accent to match the accent most familiar to the listener, while preserving the speaker's natural voice characteristics.
The initial release supported accent conversion for agents in India and the Philippines who serve primarily US-based customers. Krisp later expanded support for English-speaking agents in Latin America, South Africa, and other regions.
Accent Conversion offers two customizable modes:
| Mode | Description |
|---|---|
| Voice Profiles | Adapts to business needs with natural-sounding voices tailored to the target audience |
| Voice Preservation | Improves clarity and reduces accent strength while retaining the speaker's unique tone and inflection |
According to Krisp, the accent conversion technology delivered measurable business results for contact center customers: a 26% increase in sales conversions and a 57% improvement in Net Promoter Score (NPS).
In December 2025, Krisp released its Accent Conversion SDK, making the technology available for third-party developers to embed in their own applications.
In February 2026, Krisp launched its Voice Translation SDK, enabling real-time multilingual voice-to-voice translation in live customer conversations. The technology had been in production use since 2025 as part of Krisp's Call Center AI platform. Krisp also announced AI Voice Translation v2.0, featuring synchronous mode, auto-scoring, custom prompts, and automatic language detection.
In June 2025, Krisp announced the launch of its integrated real-time Voice AI Platform for call centers at CCW (Customer Contact Week) Vegas. The platform combined AI Noise Cancellation, AI Accent Conversion, AI Live Interpreter for speech-to-speech translation, and AI Agent Assist tools into a single solution.
As of 2025, Krisp's technology was trusted by some of the world's largest global call centers and service providers. Company data showed that Krisp helped contact centers achieve up to a 78% drop in noise complaints, an 8% increase in customer satisfaction (CSAT), and as much as a 10% decrease in average handle time (AHT).
The company grew to approximately 322 employees across four continents and had processed over 4 trillion minutes of voice conversations since its founding.
At the core of Krisp's noise cancellation technology is krispNet, a proprietary deep neural network designed for real-time noise suppression. The model was trained on more than 10,000 hours of human voice recordings and background noise samples, collected from tens of thousands of different noise types and voice recordings. The training dataset included 20,000 unique noise samples and 10,000 speaker recordings.
krispNet uses a combination of neural network architectures suited for audio processing:
Krisp's noise cancellation follows a multi-stage processing pipeline:
| Stage | Process |
|---|---|
| Input | Noisy audio is captured from the microphone |
| Pre-processing | Feature extraction converts raw audio into spectrograms or Mel-frequency cepstral coefficients (MFCCs), a time-frequency representation that reflects the human auditory system's response |
| Neural Network Inference | krispNet processes the extracted features, separating speech from noise |
| Post-processing | Clean speech features are converted back into an audio signal using DSP algorithms |
| Output | Clean audio is delivered to the conferencing application |
The entire pipeline runs with less than 20 milliseconds of algorithmic latency, well below the recommended maximum of 200 milliseconds for real-time audio applications. The model operates on small chunks of the audio signal, processing each chunk without introducing perceptible delay.
A key architectural decision in Krisp's design is that all noise cancellation processing happens entirely on the user's device. No audio data is sent to external servers, and no voice data is stored after processing. This on-device approach provides several advantages:
For non-English transcription, depending on the language and settings, audio may be sent to servers for processing and then immediately deleted once the transcript is generated.
Krisp offers multiple model variants optimized for different hardware capabilities:
| Model Size | Speed | Quality | Use Case |
|---|---|---|---|
| Small | 7x faster than Big models | Good noise cancellation quality | Lower-end devices, resource-constrained environments |
| Big | Standard processing speed | Highest speech and noise quality | Devices with sufficient CPU headroom |
To deploy the DNN inside web browsers, Krisp's engineering team achieved a 30x compression of the original model. They developed a smaller DNN that was 10x smaller than the main model while maintaining comparable quality. The team used WebAssembly (WASM) with the XNNPACK optimization library to achieve near-native C++ performance in the browser.
A particular challenge for the Chrome extension was that Chrome's audio filter architecture provides audio in 3-millisecond frames, while the DNN was designed to process 30-millisecond frames. Krisp engineer Artak developed a distributed processing algorithm that split the DNN computation across multiple 3-millisecond windows, introducing negligible additional latency. The entire process from concept to production deployment took three months.
Beyond noise cancellation, Krisp has expanded its voice AI technology stack to include:
| Capability | Description |
|---|---|
| Voice Isolation | Separates the primary speaker's voice from other human voices in the environment |
| Accent Conversion | Dynamically adjusts speaker accent to match the listener's native accent in real time |
| Voice Translation | Real-time voice-to-voice translation for multilingual conversations |
| Transcription | Speech-to-text conversion, supporting English on-device and 15 additional languages via server-based processing |
| Meeting Summarization | AI-generated meeting summaries and action items using natural language processing |
| Turn-Taking Detection | Identifies when speakers switch during conversations for better transcription accuracy |
Krisp's AI Meeting Assistant is a bot-free note-taking tool for professionals. Unlike competitor products that require a virtual bot to join conference calls, Krisp processes audio locally on the user's device, meaning no third-party bot appears as a meeting participant.
Key features of the AI Meeting Assistant include:
| Feature | Description |
|---|---|
| Noise Cancellation | Removes background noise from both incoming and outgoing audio |
| Meeting Transcription | Real-time transcription in 16 languages (English on-device, 15 additional server-based) |
| AI Meeting Notes | Automated generation of meeting summaries and action items |
| Recording Modes | Transcript-only, audio recording, or video recording options |
| Custom Vocabulary | Support for up to 750 custom terms to improve transcription accuracy for industry-specific terminology |
| AI-Suggested Agendas | Pre-meeting agenda suggestions based on calendar context |
| Timeline Navigation | Post-meeting timeline view for navigating long meeting recordings |
| Meeting Notifications | Alerts for important meeting events and follow-ups |
The AI Meeting Assistant works across all major conferencing platforms including Zoom, Microsoft Teams, Google Meet, Discord, Webex, and over 30 additional integrations.
Krisp's Call Center AI product is designed for contact centers, BPO operations, and customer experience (CX) platforms. The product combines multiple voice AI capabilities into a unified solution for improving call quality and agent productivity.
Call Center AI features include:
| Feature | Description |
|---|---|
| Agent and Customer Noise Cancellation | Removes background noise on both sides of the conversation |
| Voice Isolation | Separates the agent's voice from neighboring agents in open-plan environments |
| AI Accent Conversion | Adjusts agent accent to match the customer's native accent |
| Voice Translation | Real-time multilingual voice interpretation (add-on) |
| After-Call Summaries | AI-generated call summaries for documentation and CRM updates |
| Knowledge Chat | AI-powered agent assistance with real-time information retrieval |
| Voice Macros | Automated voice responses for common phrases and disclosures |
| Live Captions | Real-time text display of the conversation for agent reference |
| Usage Analytics | Dashboard for monitoring agent performance and platform metrics |
| Real-Time Monitoring | Supervisory tools for live call quality oversight |
Krisp's Call Center AI integrates with major CX platforms, and the company has established partnerships with providers including Five9, TTEC, Concentrix, and Cognizant.
Krisp offers a software development kit (SDK) that allows third-party developers to embed Krisp's voice AI capabilities into their own applications. The SDK is available for Windows, macOS, Linux, iOS, Android, and web browsers (via JavaScript and WebAssembly).
The SDK provides access to:
As of 2025, Krisp's AI Voice SDK powers over 150 million devices and is integrated into applications including Discord, RingCentral, Synthflow, Vapi, and more than 100 additional products.
Krisp is available across multiple platforms and integrates with a wide range of applications:
| Category | Supported Platforms / Integrations |
|---|---|
| Desktop | Windows, macOS, Linux |
| Mobile | iOS, Android |
| Browser | Chrome extension (WebAssembly-based) |
| Conferencing | Zoom, Microsoft Teams, Google Meet, Discord, Webex, Slack |
| CRM | Salesforce, HubSpot |
| Productivity | Slack, Zapier, Connectwise |
| Contact Center | Five9, Twilio, Vonage, RingCentral |
| Partner | Integration |
|---|---|
| Discord | Krisp noise cancellation integrated into Discord's desktop and mobile voice chat |
| Twilio | Krisp serves as Twilio's launch partner for the Audio Processor API, providing noise cancellation for Twilio Programmable Voice |
| RingCentral | Noise cancellation integrated into RingCentral's video conferencing platform |
As of 2026, Krisp offers tiered pricing across its three product lines:
| Plan | Monthly Price (per user) | Annual Price (per user/month) | Key Features |
|---|---|---|---|
| Free Trial | $0 (7 days) | N/A | Unlimited transcription, noise cancellation, AI notes, limited accent conversion |
| Core | $16 | $8 | Unlimited AI note-taker, noise cancellation, integrations (Slack, HubSpot, Zapier, Teams), mobile app, 1 hr/day accent conversion, 5 GB storage |
| Advanced | $30 | $15 | Everything in Core plus 4 hr/day speaker-side accent conversion, unlimited listener-side conversion, Salesforce and Connectwise integrations, manager view, 30 GB storage |
| Enterprise | Custom | Custom | SSO/SCIM, advanced security (SOC 2, BAA), on-device transcription and recordings, super admin role, dedicated account manager, unlimited storage, HIPAA compliance |
| Plan | Price | Key Features |
|---|---|---|
| CC Core | Starting at $10/agent/month (annual) | Agent and customer noise cancellation, voice isolation, after-call summaries, SSO/SCIM, usage analytics, real-time monitoring |
| CC Advanced | Custom (14-day trial available) | Everything in CC Core plus accent conversion, knowledge chat, voice macros, live captions, optional voice translation add-on |
| Plan | Availability | Key Features |
|---|---|---|
| Early Stage | Application required | Voice isolation, noise/voice cancellation, accent conversion SDKs |
| Enterprise | Application required | Higher-quality models, laughter-robust turn-taking, server-side batch processing, model customization |
Krisp has obtained several security certifications relevant to enterprise and regulated industry deployments:
| Certification | Description |
|---|---|
| SOC 2 Type II | Demonstrates strong security and privacy practices across security, availability, processing integrity, confidentiality, and privacy |
| HIPAA | Business Associate Agreements available for healthcare organizations handling Protected Health Information (PHI) |
| GDPR | Data collection and processing based on user consent, compliant with European data protection regulations |
| PCI DSS | Payment Card Industry Data Security Standard certification for handling potential cardholder data in customer environments |
Data stored in the cloud is encrypted both at rest and in transit within Amazon Web Services (AWS) infrastructure. All cloud-stored data resides on secure servers located in the United States, and access is restricted to authorized Krisp personnel.
Krisp has raised a total of approximately $15.5 million in venture capital across multiple rounds:
| Round | Date | Amount | Key Investors |
|---|---|---|---|
| Pre-Seed | 2018 | $500,000 | Berkeley SkyDeck, friends and family |
| Seed | 2018 | $1.5 million | Sierra Ventures, Shanda Group |
| Series A | August 2020 | $5 million | Storm Ventures, Sierra Ventures, TechNexus, Hive Ventures |
| Series A-1 | 2021 | $9 million | RTP Global, Granatus Ventures, and others |
Granatus Ventures, an Armenian venture capital firm, has been an early and continuing investor in Krisp. Other investors include Berkeley SkyDeck, Storm Ventures, Sierra Ventures, RTP Global, and additional venture capital firms.
Krisp operates at the intersection of AI noise cancellation and AI meeting assistants, competing with products in both categories.
| Product | Developer | Hardware Requirement | Platform Support | Key Difference |
|---|---|---|---|---|
| Krisp | Krisp Technologies | None (software-only) | Windows, macOS, Linux, iOS, Android, Web | Works with any app, any device |
| NVIDIA Broadcast (RTX Voice) | NVIDIA | NVIDIA RTX GPU required | Windows only | Free for RTX GPU owners, higher CPU and memory usage |
| Apple Voice Isolation | Apple | Apple Silicon devices | macOS, iOS | Built into FaceTime and supported apps |
| Native noise suppression | Zoom, Teams, Meet | None | Within respective apps only | Limited to specific conferencing platform |
In head-to-head comparisons, Krisp HD has shown approximately 10% higher average noise cancellation quality than NVIDIA RTX Voice, while using roughly 2x less CPU and 5x less memory. However, RTX Voice has been noted as producing fewer voice cut-off artifacts in some scenarios. Krisp's main advantage over built-in platform solutions is its ability to work across any conferencing application rather than being locked to a single platform.
| Product | Bot-Free | Noise Cancellation | Languages | Notable Feature |
|---|---|---|---|---|
| Krisp | Yes | Yes | 16 | On-device processing, accent conversion |
| Otter.ai | No (OtterPilot bot) | No | 3 | Meeting GenAI cross-conversation search |
| Fireflies.ai | No (bot joins calls) | No | 100+ | HIPAA-compliant, strong CRM integrations |
| Fathom | No (bot joins calls) | No | English primary | Free AI notetaker, highest G2 rating |
Krisp's technology is used by organizations across multiple industries:
| Industry | Notable Customers |
|---|---|
| Technology | Cisco, Autodesk, Okta, Discord |
| Media & Entertainment | Sony |
| Manufacturing | Siemens |
| Publishing | Medium |
| BPO / Contact Center | TTEC, Concentrix, Cognizant |
| Communications | Twilio, RingCentral, Vonage |
| Detail | Value |
|---|---|
| Founded | 2017 |
| Original Name | 2Hz |
| Founders | Davit Baghdasaryan (CEO), Artavazd Minasyan (CTO) |
| Chief Scientist | Stepan Sargsyan |
| Headquarters | 2150 Shattuck Ave, Penthouse 1300, Berkeley, California 94704 |
| Employees | ~322 (2025) |
| Total Funding | ~$15.5 million |
| Key Investors | Granatus Ventures, Storm Ventures, Sierra Ventures, RTP Global, Berkeley SkyDeck |
| Conversations Processed | 4+ trillion minutes (cumulative) |
| Devices Deployed | 200+ million |
| Monthly Processing Volume | 75+ billion minutes |