Pianissimo

Pianissimo (model id pianissimo-sv) is an open-weight Swedish speech-recognition model from Klang AI, a Swedish company, released on 23 September 2026 with a newsroom post, a research write-up by Mattias Fält and a Hugging Face repository; Computer Sweden reported it the same day. It is a fine-tune of NVIDIA’s Parakeet v3 architecture on about 50,000 hours of Swedish speech, with 0.6 billion parameters, taking audio and returning text. Klang reports word error rates of 4.46 per cent on Common Voice, 6.51 per cent on FLEURS Swedish and 4.85 per cent on its own Swedish dialect test set, and says it matches KB-Whisper large at 64 times the speed. Those are the developer’s own measurements and are not stored as scores, and its speed claims do not agree with each other: 2,500 times real time on one GPU, 3,600 times on the product page and more than 4,000 times on another, so the headline speed should be tested rather than quoted. The licence is CC BY 4.0, which permits commercial use with attribution. Klang also offers it through Berget AI’s API; no price was found. For Nordic organisations transcribing meetings, calls or archives in Swedish, a small open model that can run on their own hardware is the relevant part.