Sourcerepo

Soundbite

Turn the best minute of a Space or a podcast into a captioned video for X, TikTok, Reels, Shorts or LinkedIn, in the browser: Whisper writes and times every word on your device, and nothing is uploaded. MIT, every feature free.

Status
Open source
Language
TypeScript
License
MIT
Updated
Oct 2026
GitHub
Soundbite playing the sample Space: the caption lights up word by word over drk's name and title, with a small made-with credit in the video's top right corner, and the bar above says the captions are in

A clip maker for the best minute of a Space, a podcast or a voice note. Open the recording and drag over the moment, and Whisper writes out every word and times it as it is said, on your device. Pick a look and download an MP4 for X, TikTok, Reels, Shorts or LinkedIn. Apps like this upload the audio and charge a monthly plan for automatic captions and a clean clip; Soundbite is MIT licensed with every feature in, and the copy here pays for itself with one optional ad.

Install

Clone
git clone https://github.com/drkxbt/soundbite
Run
npm install && npm run demo
Use
soundbite(element, { adUnit, x, repo })

Features

  • Captions without typing: Whisper writes out the part you picked and times every word, in a worker on your device
  • Several moments in one clip: add each as a part, drag them into order, and every new part is transcribed as it's added
  • WebGPU for the encoder and WebAssembly for the decoder where there is a GPU, WebAssembly alone where there isn't
  • Silence is caught before Whisper runs, and its stock phrases over noise, in any language, count as no speech
  • A background picture, your logo and music under your voice, then Audiogram's shapes, visuals, colours, your picture and name, and an MP4 made faster than it plays
  • Your name, handle, picture and look kept in the browser for next time

How it works

  • The recording is decoded once at Whisper's 16 kHz, and only the part you picked is cut out and averaged to one channel
  • Whisper tiny, multilingual and pinned to one revision, runs in a module worker through transformers.js: about 66 MB on the first use with a GPU, 44 MB without, then cached by the browser
  • Word timings become captions that break at sentence ends and pauses; Audiogram pages long ones to the shape of the video
  • New captions replace only the ones inside the part you transcribed, so you can do a clip in pieces

Where it shows here

drk.vc/engines/soundbite runs it for anyone. 360p and 720p download free with a small made-with credit in the corner; 1080p and 1440p come without it after one short ad and two. A copy you run yourself has no ad unit, so every size is free there.

Screenshots

  • Just after transcribing: the bar turns mint and says the captions are in, and the captions list on the right fills with whole sentences
    TranscribeThe words, written and timed
  • After the ad is watched, the bar says the next download has no credit and the corner of the video is clear
    CreditOff for the next download
  • As the clip downloads, a small card in the corner thanks the user, with keys to follow on X and star the repo on GitHub
    DownloadA card to follow and star

Stack

  • TypeScript
  • Audiogram
  • transformers.js
  • Whisper
  • ONNX Runtime Web
  • WebGPU
  • WebCodecs

Run it locally

git clone https://github.com/drkxbt/soundbitecd soundbitenpm installnpm run demo