Nordic AI company Klang.ai is releasing its first open speech-to-text model for Danish. The model, called Pianissimo, outperforms the leading alternatives on speed and accuracy, both open local models and commercial offerings from tech giants such as Google, OpenAI and ElevenLabs.
Pianissimo is built on Nvidia Parakeet, an open speech-to-text model designed for fast, efficient transcription, which Klang has optimised for Danish. In the tests, Pianissimo has a lower error rate than every alternative it was compared with, and is faster than most of them.
We're building more and more technology that depends on real-time transcription, but until fairly recently the Scandinavian languages have been an afterthought. Now we've built the model we believe we need, on both speed and accuracy, says Mattias Fält, Head of AI, co-founder and lead for Pianissimo at Klang AI.
The launch comes at a time of rapid progress in Danish speech technology, with several new open models and a growing ecosystem of researchers, companies and developers. By releasing Pianissimo as an open model, Klang wants to help that progress continue, with more high-performing models.
If Nordic companies are going to build competitive AI, we need to get better at building on each other's work. With Pianissimo we want to contribute to that, and at the same time create more alternatives to the large international model providers, to strengthen the technological independence of the Nordics and Europe, continues Mattias Fält.
Tested on authentic Danish speech, at 3,600x real time
Klang designed the tests itself, to measure how well the models transcribe authentic speech that has not appeared in known training data. In the tests, Pianissimo's error rate is 10 and 11 percent lower than the open Danish models Edda and Saga respectively, and 4 percent lower than the commercial model ElevenLabs Scribe v2. Results are measured in word error rate (WER), where a lower value means fewer transcription errors.
Pianissimo is also built for fast, resource-efficient transcription and reaches roughly 3,600x real time, which means an hour of recorded audio can be processed in about one second. The model can also run locally and offline on ordinary consumer hardware, without audio or data being sent to an external service.
The next step in a Nordic push
After the Swedish version of Pianissimo launched in September, Danish is the next step in Klang's investment in open speech technology for the Scandinavian languages. A Norwegian version is planned shortly, along with more models optimised for either speed or accuracy.
More about Klang Pianissimo
- Technology: Based on Nvidia Parakeet. Runs locally and offline on consumer hardware.
- Accuracy: 11.83% WER, the best result among the models in Klang's benchmark.
- Benchmark: Three separate datasets of authentic Danish speech, from the European Parliament, Danish television and language research at the University of Copenhagen (DanPASS).
- Speed: Roughly 3,600x real time, or 40 hours of audio in under a minute.
- Use cases: Built for uses that need both high quality and speed, such as live transcription, accessibility tools and other services that have to follow a conversation as it happens.
- Availability: The model is released with open weights under the CC BY 4.0 licence, which means it can be freely used, modified and built on, including commercially, with attribution of source and licence.
Klang Pianissimo is available on Hugging Face from 3 November.
Try Pianissimo yourself