Krisp

Intercepts audio streams at the OS level (kernel audio drivers on Windows/Mac, PulseAudio on Linux) before they reach communication applications, applies neural network-based noise classification to isolate speech frequencies, and reconstructs clean audio in real-time with <50ms latency. Uses spectral subtraction combined with deep learning models trained on 10,000+ hours of noise samples to distinguish speech from environmental noise without requiring application-level integration.

Captures audio from the communication application, streams it to Krisp's cloud transcription service using WebRTC or HTTP chunking, applies automatic speech recognition (ASR) with speaker identification to tag which participant said what, and returns real-time captions with 2-3 second latency. Supports 99 languages via multilingual ASR models and handles code-switching (mixing languages mid-sentence) through language detection per utterance.

Post-processes completed meeting transcripts using a two-stage summarization pipeline: first, extractive summarization identifies key sentences via TF-IDF and topic modeling; second, abstractive summarization uses a fine-tuned T5 or BART model to generate concise summaries (2-5 sentences) that capture decisions and context. Operates on Krisp's backend after meeting ends, with results available within 30 seconds of call termination.

Analyzes meeting transcripts using named entity recognition (NER) and dependency parsing to identify action items (tasks with implied or explicit ownership), extracts deadline signals from temporal expressions, and maps action items to participants using pronoun resolution and speaker context. Outputs structured JSON with task description, assigned owner, deadline, and confidence score, enabling direct integration with project management tools via Zapier or native API.

Creates a virtual audio input/output device at the OS level (using WaveRT on Windows, CoreAudio on macOS, PulseAudio on Linux) that intercepts all audio flowing through the system. Applications select 'Krisp Microphone' as their input device, and Krisp processes the audio stream before passing it to the application, enabling noise cancellation and transcription without requiring native plugins or SDKs for each platform.

Uses voice biometrics and speaker embedding models (similar to speaker verification systems) to identify and track individual participants across multiple meetings. Builds a speaker profile from the first few utterances of each participant, then matches subsequent speakers against this profile using cosine similarity on mel-frequency cepstral coefficient (MFCC) embeddings. Enables consistent speaker labeling even if participants don't explicitly introduce themselves.

Aggregates data from multiple meetings (transcripts, summaries, action items, speaker participation) and generates analytics visualizations including speaking time per participant, meeting frequency, action item completion rates, and topic trends over time. Data is stored in Krisp's backend and accessible via web dashboard or API, enabling team leads to understand meeting patterns and team dynamics without manual analysis.

Provides optional offline noise cancellation mode that runs the neural network model locally on the user's device without sending audio to Krisp's cloud servers. Uses quantized (INT8) versions of the noise suppression model (~50MB) to reduce memory footprint, enabling inference on devices with limited resources. Trades off slightly lower accuracy (2-3% degradation) for complete privacy and elimination of cloud latency.