|
|
Abstract The Krisp audio plugin represents a paradigm shift from traditional spectral subtraction and Wiener filtering to data-driven deep learning for real-time noise suppression. This paper dissects Krisp’s operational pipeline, from its dual-microphone or single-channel input to its output in VoIP and streaming contexts. We examine the model architecture (likely a convolutional recurrent neural network, CRNN), the training data ecosystem, latency constraints, and the trade-off between noise removal and speech distortion. Comparative analysis against classic methods (e.g., RNNoise, WebRTC’s NS) highlights Krisp’s advantages in non-stationary noise suppression and its limitations in musical and transient preservation. 1. Introduction Krisp, introduced around 2017, gained prominence as a virtual audio device (macOS/Windows) that removes background noise from both sides of a conversation in real time. Unlike hardware-based solutions, Krisp operates entirely in software, intercepting audio streams at the OS audio driver level.