Krisp VIVA is an SDK of audio models for voice AI agents. Its core component, voice isolation, strips background noise, cross-talk, and secondary speakers from the incoming audio stream before it reaches voice activity detection or speech-to-text, so the downstream models hear only the primary speaker.

The models operate on the raw audio signal rather than on a transcript, which means they work across languages and accents without per-language configuration. Companion models handle conversation dynamics: turn prediction estimates when the speaker has finished, interruption prediction classifies mid-response speech as an interruption, a question, or a backchannel, and a voice activity detector holds up against background voices.

The SDK is built to run on server CPUs and ships as native libraries with bindings for common backend languages, and it plugs into the pipelines that voice agents are typically built on.

Features

Voice isolation: removes noise, cross-talk, and secondary speakers ahead of VAD and STT
Turn prediction: signals end of speech from audio alone so the agent can respond promptly
Interruption prediction: distinguishes interruptions from questions and backchannels while the agent is speaking
Voice activity detection: a CPU-oriented detector that ignores background voices
Language-agnostic: operates on the audio signal, with no transcription or language setting required
Platforms: Windows, macOS, Linux, web (JavaScript/WASM), iOS, and Android
Bindings: C++, Node.js, Python, Go, and Rust
Pipeline integrations: LiveKit, WebRTC, and Pipecat
Related SDK: a separate RTC SDK covers human-to-human calls with noise cancellation, accent conversion, and voice translation

Claim Krisp VIVA