Urgent.News

What's breaking now, across thousands of outlets.

Tech

I built a free Whisper transcription site that runs in the browser: 5 things that broke

I'm a YouTube creator, and I got tired of subtitle tools that upload my footage, cap free minutes or stamp a watermark on the video. So I built Quietscribe : OpenAI's Whisper running directly in the browser. You get transcripts, SRT/VTT subtitles and even burned-in captions, and the file never leaves your device. It's a static site on Cloudflare Pages: no backend for the default mode and zero…

A YouTube creator created a free, browser-based Whisper transcription site called Quietscribe. The site runs OpenAI's Whisper directly in the browser, providing transcripts, SRT/VTT subtitles, and burned-in captions without uploading any files. Using Cloudflare Pages for a static site, the creator faced five main issues in building the site.

Firstly, auto-detecting the language was set to English by default. However, when testing Urdu, it produced gibberish English text. The solution was to use the first 30 seconds of speech to run a decoding step and pick the language with the highest score.

Secondly, iPhone Safari would reload the entire page when the model download reached 100%. This was caused by the WASM thread duplicating memory. To fix this, the site determined the chosen model size from sessionStorage and stepped down if the page had crashed before.

Thirdly, small models could get stuck repeating words, particularly in lower-resource languages. To address this, the site detected loops (4+ identical words or a low unique-word ratio) and re-ran the piece with no_repeat_ngram_size set to 4 and a repetition penalty of 1.15.

Fourthly, some languages required larger models than phones could handle. In such cases, an opt-in Cloud mode was added where the browser decodes the audio to 16 kHz mono and sends each 60-second piece to a Cloudflare Worker running whisper-large-v3-turbo. The file is never stored, and whisper-large-v3-turbo cannot translate, so the Worker transcribed first and then translated each segment using @cf/meta/m2m100-1.2b.

Lastly, burned-in captions and MP4 produced VP9-in-MP4 in some Chromium builds, which many players rejected. Explicit avc1 codec strings were requested, and if failing, the site fell back to WebM. Long captions needed to be shrunk before wrapping, or single orphan words would appear on a second line. The site is private by default, but the creator added a share target in the web manifest, allowing Android users to tap Share → Quietscribe from their gallery or WhatsApp.

The file still never leaves the device. Quietscribe is free and available at quietscribe.pages.dev. Feedback is welcomed regarding speed and accuracy in various languages.

Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.

Read the original at dev.to →

More in Tech

TechCrunch Mobility: A new, bigger AV boom is here

Welcome back to TechCrunch Mobility, your hub for the future of transportation and now, more than ever, the role AI is playing in it.

  • 81 autonomous vehicle startups pitched to Trucks VC in 2026, a 2/3 increase from last year
  • Foundation models enable startups to create end-to-end AV architectures within months
  • Pure-play robotaxi businesses emerge as direct competitors to Waymo

Giving an AI Assistant Access to Your Accounts: What to Check

The most useful AI assistants are the ones connected to your real tools, because they can see the invoice, the calendar entry and the customer thread instead of asking you to paste them in.

  • AI assistants with tool access pose significant risk
  • Read access risks privacy, write access risks operations & finances
  • Start with read-only access, human approve write actions

More from Sunday 1 November →