Crisp - Automatic Silence & Filler Remover
A native SwiftUI macOS app that automatically cuts long pauses and filler words (um, uh, hmm) from screen recordings — audio and video together — producing tight jump-cuts. Runs 100% locally, never touches the original, and includes experimental on-device ML models for filler detection.
Project Overview
Crisp makes your recordings crisp. It is a native SwiftUI app that automatically removes long pauses, silence, and filler words — um, uh, hmm, aww — from screen recordings, cutting them out of the audio and the video together so you skip straight to real editing. Everything runs 100% locally; nothing is uploaded. The original is never touched: it is backed up first, and the result is saved as a new …_cleaned.mp4 next to it.
How It Works
The pipeline is four stages. Backup copies the original to _originals/. Detect pauses uses ffmpeg silencedetect to find silence from real audio energy, which stays accurate even when speech timing is irregular. Find fillers runs whisper.cpp for word-level timestamps and matches hesitation words. Re-render keeps only the wanted segments and joins them with ffmpeg, audio and video locked together at the same resolution and frame rate — no downscaling.
The user picks how aggressive the cut is (Gentle to Very aggressive) and whether to strip filler words, drags in one or several videos, and watches a live progress log.
Research — Custom On-Device Models
Crisp is experimenting with its own small on-device ML models for filler detection as a fast, opt-in alternative to full speech-to-text. The first, Wren, is a tiny CNN trained on PodcastFillers that spots “um”/“uh” at roughly 600× real-time with 0.94 precision on held-out speakers. All training code, benchmarks, and a native eval dashboard live in a separate research/ tree, kept out of the shipped app; trained models are published to Hugging Face.
Channels & Distribution
Crisp builds in three side-by-side channels — Stable, Nightly, and Dev — each a separate app with its own icon and settings. Stable and Nightly auto-update from GitHub Releases; Dev never does. The roadmap and design notes are tracked in-repo as plain Markdown via Backlog.md, so shipped, in-progress, and planned work all live in git with no external service.
License
GPL-3.0. © Syntax Lab Technology.