Title: hushscribe: Confidential Audio Transcription in the Browser

URL Source: https://blog.hem.si/blog/2026-09-09-hushscribe/

Markdown Content:
---
title: hushscribe: Confidential Audio Transcription in the Browser
description: Static page transcribing via Privatemode.ai confidential computing — no backend, encrypted end-to-end, ~€0.014/min.
image: https://blog.hem.si/blog-placeholder-1.jpg
---

 09.09.2026 

# hushscribe: Confidential Audio Transcription in the Browser

[Get Site as Markdown ](https://markdown.new/https://blog.hem.si/blog/2026-09-09-hushscribe/) 

---

Repo: [dhcgn/hushscribe](https://github.com/dhcgn/hushscribe) · Live: [dhcgn.github.io/hushscribe](https://dhcgn.github.io/hushscribe/)

## Problem

I needed a way to transcribe confidential dictation. With browser-based encryption and a TEE in the backend, this archives the highest privacy currently possible — beside local transcription. As a PWA it installs as an app on Android with share-target support, so I can for example share WhatsApp audio messages for confidential transcription.

## Tech Overview

| Aspect    | Choice                                                                                           |
| --------- | ------------------------------------------------------------------------------------------------ |
| Frontend  | Vite + TypeScript static page, no backend                                                        |
| Inference | [Privatemode.ai](https://docs.privatemode.ai/) confidential VMs (attested, end-to-end encrypted) |
| Models    | whisper-large-v3 (€0.014/min), voxtral-mini-3b (€0.004/min), prices Sep 2026 + VAT               |
| Storage   | localStorage only (key, prompts, last 20 transcripts); media never stored                        |
| Install   | PWA manifest, Android share target via WebAPK (Chrome)                                           |

hushscribe is a frontend only — attested VMs, verifier, and models are Privatemode (Edgeless Systems). It contributes no cryptography of its own.

## Solution

Bring a Privatemode API key, drop in a file, get a timestamped transcript. Cost is estimated from audio length before the request. Supported input: `flac mp3 mp4 mpeg mpga m4a ogg wav webm` up to 50 MB / 1 h; other containers are re-encoded in-browser first. Set a language for timestamps and `.vtt` / `.srt` export; auto-detect returns plain text.



On Android (installed from Chrome) hushscribe appears in the system share sheet: record, tap Share, pick hushscribe — the file is passed in-memory via service worker, never written to disk. Installed mode changes nothing about encryption or storage.



## Outlook

I believe this is the first browser-based app combining confidential-computing transcription with PWA install + share. Limits: needs network (no offline), malicious extensions or a compromised endpoint still break browser-side guarantees. Verify via the attested measurement shown in the page and the reproducible SDK build.