Pianissimo: The best Danish speech-to-text model

After releasing Pianissimo in Swedish last month, we decided to train a Danish model too. Not only is it one of the fastest models, it also achieved the highest quality of all the models we tested, both open source and proprietary!

Released on Hugging Face today, download it yourself or read more about how we trained it below.

Benchmarks

Quality can be difficult to measure accurately for speech-to-text models. We therefore created three completely new datasets to evaluate Pianissimo alongside existing Danish ASR models. We created test sets based on speeches at the EU Parliament, conversations on TV, and DanPASS, a speech dataset developed at the University of Copenhagen. Together they contain over 10,000 segments of text from more than 10 hours of speech. We compared Pianissimo to both the latest open source models and the best closed language models from Google, OpenAI, Mistral and ElevenLabs. Pianissimo performed better than all of them.

We’ll discuss the datasets and why benchmarks can be misleading in our upcoming article on benchmark overfitting.

How we trained it

Many of the lessons from training Pianissimo for Swedish went into the creation of this model. Starting from over 80,000 hours of publicly available speech data, we filtered, aligned, segmented and augmented the audio in a similar way to how we prepared the Swedish Pianissimo model. We based the model on NVIDIA’s multilingual Parakeet v3 model and trained it for almost 500,000 steps over several days.

Quality at speed

We selected this model architecture because it's really fast. It can transcribe over 1 hour of speech in less than 1 second on an H100 GPU. This makes it practical for large-scale use, as well as local inference on laptops and phones. Normally, this speed comes at a cost in recognition quality, but Pianissimo is able to outperform even the much larger models. To make it possible to run Pianissimo on as many devices as possible, we release it in several different quantized formats:

• ONNX: 4 versions down to 4-bit. For CPUs and browser use
• GGUF: 5 versions down to 4-bit. For CPU and GPU
• MLX: 4 versions down to 4-bit. For Mac M-series GPUs.
• CoreML: 8-bit and 16-bit for Neural Engine on Mac and iPhone

Size: not just about speed and cost

While being able to run the models fast on cheaper hardware is a great benefit, it’s not all about costs. Being able to run a model locally on your laptop or phone is also about control. With small open source models like Pianissimo, you can decide who to trust with your data. Some data might be better kept local, whereas much can be processed by trusted providers. By releasing Pianissimo openly, we make it possible for as many providers as possible to give you the highest possible quality, while letting you decide who can be trusted with your data!

Try it out or read more

You can try our demo or check out the model yourself on Hugging Face.