Theo Lindgren·Maker's desk·Updated 2026-10-07·3 min read

Transcribe recordings into handwritten notes, free and local

A two-hour meeting recording, a lecture, a voice memo full of ideas — turning them into a shareable handwritten note used to mean four steps: transcription app, copy, layout, export.

We compressed it into one: upload an audio file, get a handwritten page. Free, with the audio never leaving your device. Here's how it works, our real test results, and the honest limitations.

The three routes, compared

Route	Cost	Privacy	Best for
In-browser Whisper (ours)	0	Audio stays on device	Daily notes, privacy-sensitive content
Cloud ASR (iFlytek, Doubao, Groq…)	Per minute or limited free tier	Audio uploaded	Very long files, speed-critical
Phone dictation + copy-paste	0	Depends	One-off short snippets

The local route's only cost: a one-time model download, and transcription speed depends on your hardware.

How we implemented it (the technical bits)

Files are decoded via the browser's decodeAudioData and resampled to 16kHz mono;
The Whisper model (~80MB fast / ~250MB high-quality, q8 quantized) is self-hosted on our own domain, chunked into 30-second windows;
Desktop Chrome with WebGPU gets GPU acceleration; everything else falls back to WASM (CPU);
The transcript lands in the same editor — fonts, paper, export all behave identically to typing.

We self-host the model instead of calling the HuggingFace CDN to avoid cross-origin blocks and keep downloads working in China.

Real test: a TTS sample

Environment: macOS + Chrome, no discrete GPU (WASM path), fast model.

Input: synthesized Chinese speech saying "你好,这是一段测试语音,用来验证音频转写功能。" (~5s)
Output: "你好,這是一段測試語音,用來驗證音頻轉寫功能。"

Content fully correct, but rendered in traditional characters — a known tendency of base-tier models; switch to the high-quality tier or fix by hand. First model load took about a minute; afterwards the browser caches it.

The import-audio entry in the toolbar

Three steps

On the tool page, click "Import audio file" and pick mp3 / wav / m4a;
Wait for the progress hint (includes first-time model download);
The transcript lands in the editor — fix typos, pick a style, export.

Limitations, stated plainly

Base-tier accuracy degrades with heavy accents and noise; use the high-quality tier and proofread for anything important;
30+ minute files take noticeably long — split them;
Safari's decoding of some audio codecs is incomplete; desktop Chrome / Edge recommended;
A file that is quiet all the way through stays quiet in the transcript. Lift the level first: fix a quiet voice memo.

Tested, with honest limitations

Every step was tested on desktop Chrome (macOS, October 2026); screenshots are from real runs. Speech recognition is provided by the browser vendor and quality varies — proofread important text. We collect no text or audio — see the Privacy Policy. Technical reference: MDN · Web Speech API.

Try it yourself

Templates