About UmBreak
Why a filler-word remover that refuses to see your audio.
Every podcast, course, and voice memo has the same problem: um, uh, like, you know. The existing tools that fix this all ask for the same trade first — upload your voice to their servers and create an account.
UmBreak was built on a simple observation: modern browsers can now run real speech-recognition models locally. So there is no technical reason your audio needs to travel anywhere. We built the tool around that fact instead of around a signup funnel.
How it works, technically
- Transcription: Whisper (tiny, English) running in your browser via Transformers.js and WebAssembly.
- Editing: word-level timestamps let you cut by clicking words in the transcript.
- Export: the Web Audio API re-assembles your audio minus the cut words, as a WAV download.
- Privacy: no backend, no database, no user accounts. The site is static files.
What it is not
- Not a Descript replacement — no multitrack editing, no overdub, no collaboration.
- Not perfect transcription — the tiny model trades accuracy for speed and size. Check important words.
- Not finished — video re-export and more languages are honest roadmap items, not hidden paywalls.
Who makes it
UmBreak is a small independent project. No funding announcements, no growth team — just a tool that respects the person using it.