Klang is releasing its first open speech-to-text model for Swedish. The model, called Pianissimo, is trained for Swedish and Swedish dialects and is efficient enough to run locally on an ordinary laptop. It is released as open source, freely available for others to use and build on.
Pianissimo builds on Nvidia Parakeet, an open speech-to-text model developed for fast, efficient transcription, which Klang has optimised for Swedish and Swedish dialects. In Klang's tests the model has accuracy comparable to KB Whisper Large, the best open Swedish model from the National Library of Sweden, but is roughly 60 times faster.
Klang's transcription model reaches around 3,600x real time, which means it can process the equivalent of just over an hour of audio in one second. That speed makes it possible to transcribe speech continuously during a conversation with very little delay, which makes the model particularly well suited to real-time transcription and other services where large volumes of audio need to be processed quickly.
We're building more and more technology that needs real-time transcription ourselves, but we've lacked an open model adapted for Swedish that combines high quality with the speed we need. The fact that Pianissimo runs on ordinary consumer hardware also opens up entirely new kinds of services. Now we're looking forward to seeing what others build on the model, says Mattias Fält, Head of AI, co-founder and lead for Pianissimo at Klang AI.
Wants more Swedish companies to build on each other's technology
Klang is releasing the model openly and freely to everyone, and sees this as an important part of building a stronger Swedish and European AI ecosystem, where companies can do more to build on technology that has already been developed.
If Swedish and European companies are going to build competitive AI, we need to get better at building on top of each other's work, with interoperability at the core. By opening up the model, others can spend their time on the next layer of innovation. For us, it's also about reducing our dependence on both external model providers and non-European infrastructure, Mattias Fält continues.
Pianissimo can run locally and offline straight away, and is also offered through the API of Swedish company Berget AI, on infrastructure in Sweden. Companies that want to use the model as a hosted service can therefore do so through a Swedish provider.
We believe strongly in the Swedish ecosystem, and this shows what becomes possible when Swedish players work together openly and closely. We're really pleased to take the next step in our partnership with Klang and to offer new possibilities for live transcription in Swedish. Just as before, we can guarantee that data is processed entirely within Sweden and that nothing is stored, which is essential for many organisations with strict data security requirements, says Andreas Lundmark, CEO, Berget AI.
The first model in a larger initiative
Pianissimo is the first of several AI models that Klang plans to release openly. More models are planned in the near future, some optimised for speed and others for even higher accuracy. Klang is also planning versions for more languages, starting with Danish and Norwegian, with the ambition of then developing more specialised models for the European market.
More about Klang Pianissimo
- The model is built for use cases where Swedish speech recognition has to work in real time, such as live transcription, accessibility solutions and other services that need to follow a conversation while it is happening.
- The model can run locally and offline on ordinary consumer hardware, without audio or data having to be sent to an external AI service.
- In Klang's tests, Pianissimo reaches around 3,600x real time, the equivalent of roughly 40 hours of audio processed in under a minute.
- The model is released as an open model, which means it may be used, modified and built on.
- The accuracy and speed comparisons were made on standard datasets such as Common Voice and FLEURS, and on Klang's own dataset of Swedish dialects.
Klang Pianissimo is available from 23 September on Hugging Face and as an API through Berget AI.
Try Pianissimo yourself