Klang Dictate

The problem: dictation is an afterthought

Every operating system ships with dictation. And almost nobody uses it. Not because people do not want to speak to their computers — but because the experience is broken. Poor accuracy for non-English languages. Noticeable latency. No understanding of context or domain-specific terms. Dictation today feels like a demo, not a tool.

For Nordic languages, the situation is worse. Swedish, Norwegian, and Danish are afterthoughts in the big speech models — fine for simple commands, unreliable for real work. If you are a Swedish-speaking professional who wants to dictate emails, notes, or code comments, the built-in tools will let you down.

Our approach: on-device, Nordic-first, system-wide

Klang Dictate is our attempt to answer a simple question: what happens when dictation actually works?

We are building a macOS dictation system from the ground up with three principles:

• On-device processing — All inference runs locally on Apple Silicon's Neural Engine. No cloud round-trips, no data leaving your machine. This is not just a privacy feature — it is a latency feature. Local processing means the words appear as fast as you speak them.
• Nordic-first models — Our models are trained and optimised specifically for Swedish, Norwegian, and Danish. Not fine-tuned from English — built for Nordic languages from the start, with proper handling of regional dialects and code-switching.
• System-wide integration — Klang Dictate works in any text field across macOS. Not a separate app you paste from — a replacement for the built-in dictation that activates with a hotkey and just works.

What we are learning

Building dictation that people actually use has taught us things we could not have predicted:

Accuracy is necessary but not sufficient. A system that is 95% accurate but pauses at clause boundaries feels worse than one that is 92% accurate but flows naturally. The feel of dictation matters as much as the word error rate.

Custom vocabulary is not optional. Every user has a handful of terms — project names, technical jargon, colleague names — that trip up generic models. On-device custom vocabularies are technically hard, but without them, users abandon dictation within a week.

Context matters more than language. A doctor dictating notes needs different model behaviour than a developer dictating code comments, even in the same language. We are exploring how to make the system aware of context without requiring manual mode switching.

These insights feed directly back into our speech models and inform the direction of all our other projects.