Krisp builds machine-learning software that removes background noise and isolates the human voice in real time across platforms. The product runs locally with low overhead on Intel and ARM architectures and includes features like noise cancellation for high-fidelity voice channels and acoustic correction to remove room echoes. The company has added enterprise-oriented controls (provisioning and central administration) and plans features such as cross-service call recording, analysis, and voice cancellation of other people’s voices. Krisp experienced massive growth in 2020 — roughly 20x active users, 23x enterprise accounts, a 13x improvement in ARR — and now counts more than 1,200 enterprise customers with B2B revenue eclipsing B2C. To speed model experimentation and reduce cloud costs the team invested in an on-premises data center with Nvidia A100 GPUs instead of relying on AWS/Azure/Google Cloud. Hiring is focused in Armenia: the research team was about 25 strong and the company planned to grow it to roughly 45 by the end of 2021; management aims to reach $15M ARR before pursuing a larger round. Krisp applies a machine-learning system to audio in real time to remove non-voice sounds and isolate the speaker, claiming about 15 milliseconds of latency and modest computational overhead so it can run on most devices. The product launched as free standalone software with a paid tier and integrations such as Discord, but the company has shifted toward enterprise customers. In the past year Krisp grew from 0 to 600 paying enterprises and from $0 to $4 million in annual recurring revenue, with customers including a call center of ~40,000 employees and a large tech buyer that purchased 2,000 licenses. The team is split between the U.S. and Armenia, and the company emerged from UC Berkeley’s SkyDeck in 2018. Krisp says it will continue improving its noise-suppression tech while building features to surface conversation metadata and provide real-time feedback on speaking (a "Grammarly for voice" concept). The company emphasizes on-device processing and privacy, stating audio never leaves the device. 2Hz builds Krisp, a desktop application that uses on-device machine learning to subtract background noise (crowds, traffic, crying kids) while preserving the human voice. The app modifies both outgoing and incoming audio at the OS level, producing a clean signal with very short latency (around 15 milliseconds) and no cloud processing or storage. The team pivoted from a carrier/platform approach to a consumer-facing product and launched first on Mac, where early-adopter traction was stronger. A Windows release is expected soon (the article cites either late this month or early January) and will include a gaming-focused variant with potential power-user features. Technically, the company has ported its DNN to NVIDIA GPUs, Intel CPU/GNA, and ARM, with Qualcomm in the pipeline. The team demoed the prototype at UC Berkeley’s SkyDeck accelerator.

Total raised

$16M

Funding rounds

3

Latest round

Series A

Latest activity

Feb 2021