Introduction
Krisp provides a JS SDK with an out-of-the-box application architecture that allows developers and teams to integrate AI-powered speech clarity features in real-time communication applications.
Krisp SDK is the only robust Javascript-based Noise Cancellation SDK available for browsers. Running audio AI inside the browsers is challenging for various reasons.
Audio Pipeline and Processing
Krisp Javascript SDK is integrated with WebRTC and Web Audio as a noise audio filter.
The noise-canceling neural network is built on WASM for high performance.
The SDK receives audio buffers in chunks and processes them in a dedicated worker to offload the main thread.
Supported Audio Characteristics
Audio frame duration
10ms *
* While an audio stream of any frame size is supported, Krisp SDK processing is performed using 10ms frames.
Audio sample rates
8KHz, 16KHz, 32KHz, 24KHz, 44.1KHz, 48KHz, 88.2KHz, 96KHz
Performance
- Package Size: 12MB
- Memory Footprint: ~100 MB
- Frame Processing: 1.5-2ms
Outbound (Mic) Noise Cancellation
Benchmarks conducted for JS SDK v1.3.0 using Chrome v120 on the same reference computers used for Desktop. On Mac M1 Krisp SDK runs ~85% on efficiency cores resulting in relatively higher CPU% but ensuring better battery performance.
Inbound (Speaker) Noise Cancellation
Benchmarks conducted for JS SDK v1.4.4 using Chrome v120 on the same reference computers used for Desktop.
Background Voice Cancellation
Benchmarks conducted for JS SDK v1.4.3 using Chrome v120 on the same reference computers used for Desktop. The memory footprint is measured for the small model exclusively.
Important Notes
- CPU% is generally higher in JS SDK compared to native SDK. This is expected as the code runs in the browser in different worker threads, which provides improved stability and eliminates memory leaks.
- 8K audio stream is not supported by theSafari browser due to the Apple WebKit limitations.
Updated 5 months ago