Today we're releasing Klang Pianissimo, an open-weight speech recognition model for Swedish. Pianissimo cuts the error rate of comparable models by 75% and transcribes at 2,500× realtime.
The KB-Whisper series previously dominated the trade-off between speed and quality among open models. Pianissimo achieves a lower word error rate (WER) than KB-Whisper medium and is comparable to KB-Whisper large, a model more than twice as large and 64× slower than Pianissimo. This was measured on the open datasets Common Voice and FLEURS as well as our internal datasets. A more technical writeup of the benchmarks is available here.
Transcription at incredible speed
When processing audio in parallel, Pianissimo can transcribe more than forty hours of audio in one minute on a single NVIDIA A100. A single 30-second clip takes about 95 milliseconds to transcribe.
Pianissimo is fine-tuned from NVIDIA's Parakeet v3, a 0.6B-parameter FastConformer-TDT model designed for fast transcription. Local attention keeps the attention cost linear in recording length, allowing the model to handle long recordings efficiently.
Specialized for Swedish
On Klang Dialects, our crowdsourced benchmark of regional Swedish speech, Pianissimo reduces word error rate by 81% compared with the original Parakeet v3. The improvements extend to 76% on Common Voice Swedish test, 57% on Swedish FLEURS and over 83% on our internal dataset of read sentences.
How we trained Pianissimo
We trained Pianissimo on over 50,000 hours of Swedish speech, then refined it on a 4,000-hour subset with more reliable transcripts. Much of the work went into checking that the accompanying text actually corresponded to what was said.
In How we trained Pianissimo, we explain how we filtered the data, built the training curriculum and fixed an unexpected problem that caused the model to skip whole sentences in long recordings.
Open weights and digital sovereignty
We could only start Klang because open models were available. Pianissimo builds on NVIDIA’s work, and we’re releasing it openly so others can build on ours. We believe Europe needs a larger ecosystem of open models. With Pianissimo, you can run the model locally, on infrastructure you choose, and build your own products and services around it.
Future models
Pianissimo is our first model, but it won’t be our last. We see it as the first step towards a collection of models in different sizes, specialized for European languages. Stay tuned for more news!
Get Pianissimo
Download the weights at KlangAI/pianissimo-sv on Hugging Face, released under CC BY 4.0. The article How we trained Pianissimo contains details and the full benchmark methodology for our numbers.