KB-Whisper for Swedish Dictation on Mac — Why It Matters

OpenAI’s Whisper is a remarkable achievement. Trained on 680,000 hours of audio from across the internet, it handles dozens of languages without any language-specific configuration. For Swedish, this means it can transcribe common speech with reasonable accuracy. But “reasonable” for a general-purpose multilingual model is not the same as “built for Swedish.” The National Library of Sweden decided to close that gap.

KB-Whisper is the result: a fine-tuned Whisper model trained specifically on Swedish speech, developed by KBLab — the AI lab at Kungliga biblioteket, the Royal Library of Sweden. This article explains how it was built, what makes it better for Swedish dictation, and how it runs locally on a Mac with no audio leaving your device.

What KB-Whisper is

KB-Whisper is not a different architecture from Whisper — it uses the same transformer-based design. What sets it apart is the training data. KBLab fine-tuned the model on over 50,000 hours of specifically Swedish speech, drawn from:

• SVT (Sweden’s public broadcaster) archives
• Sveriges Radio broadcasts
• Riksdag (parliamentary) debates
• Historical and archival audio material

This speech spans domains, speakers, and registers that matter for professional dictation: formal and informal speech, legal language, political discourse, technical explanations. Generic Whisper was exposed to some of this via internet audio, but Swedish is a small fraction of its 680,000-hour training corpus. KB-Whisper inverts that ratio for Swedish specifically.

The figures come from KBLab’s own published benchmarks, not from saega’s testing. The 47% improvement in word error rate (WER) is measured against generic Whisper of the same model size on Swedish speech. It is not a marginal improvement.

Why generic Whisper struggles with Swedish

Swedish has specific properties that challenge a general-purpose multilingual model:

Compound words. Swedish compresses concepts into single words in ways English rarely does: samhällsutvecklingen (society’s development), arbetsmarknadsutbildning (labor market training), dataskyddsförordningen (the data protection regulation). A model not well-trained on Swedish will frequently split these incorrectly or substitute a near-English equivalent.

Swedish proper nouns. Place names like Östersund, Hässleholm, Ånge, Luleå; names like Björn, Åsa, Gösta. These are phonetically distant from English-language patterns, and generic Whisper approximates them inconsistently.

Professional vocabulary. Legal, medical, technical, and political terms in Swedish differ substantially from their English counterparts. A Swedish court document contains words like rättegångsbalken, förundersökning, säkerhetsställning — none of which map cleanly onto English training data.

How KB-Whisper runs locally on Mac

Whisper models require GPU acceleration to run in real time. On Intel Macs this is slow — local Whisper was impractical for live dictation. Apple Silicon changed this. The M-series chips have a dedicated Neural Engine and full Metal support for machine learning. KB-Whisper Small runs in real time on an M1 MacBook Air without significant battery drain or heat.

In sæga, the transcription flow is entirely local. When you press the hotkey and speak, the audio is captured, processed by KB-Whisper on your device via Metal acceleration, and the resulting text is inserted at your cursor. Nothing — no audio, no text — leaves your Mac in Raw mode. For professionals handling sensitive information, this is not just a preference but a compliance requirement.

KB-Whisper in practice: what actually improves

The 47% WER improvement is a statistical aggregate. In practice it means certain categories of errors nearly disappear:

• Compound word splitting. Generic Whisper: “arbetsmarknads utbildning.” KB-Whisper: “arbetsmarknadsutbildning.”
• Swedish place names. Östersund, Skövde, Trollhättan — consistently right rather than approximately right.
• Gender agreement. Swedish grammar requires agreement between nouns and adjectives (en stor bil / ett stort hus). KB-Whisper makes fewer agreement errors because it has seen far more Swedish grammatical context.
• Intonation patterns. Swedish has distinctive pitch accent (ordaccenterna). A model with broad Swedish training handles these more naturally than one that only heard English and a small sample of Swedish.

Over a working day of professional dictation, these differences add up to meaningfully fewer corrections. A legal brief, a medical note, or a government report with fewer transcription errors is worth a noticeable amount of review time.

Who builds this and why it is not a product

KBLab is the AI research lab at Kungliga biblioteket — Sweden’s national library. Their mandate is to advance AI for Swedish language and culture, not to ship commercial products. KB-Whisper is published as a research contribution: the weights are available for researchers, developers, and anyone who wants to use them.

This is important context. KB-Whisper does not compete with Whisper — it is built on top of it, with the explicit goal of serving Swedish speakers better. It is analogous to what NbAiLab has done for Norwegian with NB-Whisper, and what language labs in other smaller language communities are doing across Europe.

How to try KB-Whisper for Swedish dictation

sæga is built around KB-Whisper and NB-Whisper as its default models for Swedish and Norwegian speakers. To try it:

• Download sæga from saega.app — no account required
• In Settings → Whisper Model, select KB-Whisper Small (downloads once, around 150 MB)
• Start dictating with Option+Space

The Raw mode is free with no time limit. The difference in compound word accuracy and proper noun handling is noticeable immediately for anyone who dictates professionally in Swedish. See our full comparison of Mac dictation apps in 2026 →