Klang is a Swedish AI speech lab helping organisations build a memory of their conversations. That requires Swedish transcription that is both accurate and fast enough to process large amounts of audio. In our Swedish benchmarks, the specialized KB-Whisper models were the most accurate, and NVIDIA Parakeet v3, part of the NVIDIA Nemotron Speech collection of models, was the fastest.
Starting from Parakeet v3, we trained Klang Pianissimo on more than 50,000 hours of Swedish speech. Across five evaluation sets, it reduces mean word error rate from 17.7% to 4.5%, a relative reduction of about 75%, while retaining the original model's throughput.
Of the 14 open models we evaluated, Pianissimo is almost as accurate as the best. In our batched benchmark on a single NVIDIA A100 Tensor Core GPU, it transcribes at approximately 2,500 times realtime, delivering 64 times the throughput of KB-Whisper large.
Swedish recognition after fine-tuning
We evaluated recognition quality on four public datasets and an internal dataset of read speech, measuring word error rate (WER). Fine-tuning using NVIDIA NeMo open libraries improved Parakeet's recognition on every evaluation set:
| Evaluation set | Parakeet v3 WER | Pianissimo WER |
|---|---|---|
| Common Voice test | 18.5% | 4.5% |
| Common Voice development | 17.5% | 4.5% |
| FLEURS | 15.2% | 6.5% |
| Klang Dialects | 25.8% | 4.9% |
| Internal read speech | 11.5% | 2.0% |
| Mean | 17.7% | 4.5% |
The largest absolute improvement was on Klang Dialects, our crowdsourced benchmark of regional Swedish speech. Across the five sets, Pianissimo outperformed both Whisper large-v3 and KB-Whisper medium, which scored 7.2% and 4.9% mean WER, respectively.
More than 40 hours of audio per minute
Pianissimo keeps Parakeet v3's architecture, so it retains the same impressive speed.
We measured throughput using batches of 30-second clips on a single NVIDIA A100 Tensor Core GPU, running each model in its default framework. For each model, we selected the fastest configuration from the batch sizes and precision settings we tested.
| Model | Mean Swedish WER | Throughput (× realtime) |
|---|---|---|
| Parakeet v3 | 17.7% | 2,500× |
| Pianissimo | 4.5% | 2,500× |
| KB-Whisper medium | 4.9% | 66× |
| KB-Whisper large | 3.5% | 39× |
At these rates, one minute of model processing covers more than 40 hours of audio with Pianissimo, compared with 39 minutes with KB-Whisper large.
Our full benchmark methodology includes timing procedures, model settings, and results for the wider comparison.
Training Pianissimo
Parakeet v3 is a 0.6-billion-parameter model already trained to recognize speech across multiple languages. Building on those weights allowed us to concentrate on Swedish.
We assembled more than 90,000 hours of speech from publicly available material, including parliamentary recordings, subtitles, and open datasets. The data needed a lot of careful filtering before we could train on it.
After filtering, more than 50,000 hours remained. We trained on NVIDIA GPUs using NeMo open libraries, first using this broad collection and then gradually shifting towards a 4,000-hour subset with more reliable transcripts.
Our training article explains the data preparation and training in detail.
Get started
Pianissimo is available on Hugging Face, with weights released under CC BY 4.0. The repository includes instructions for running Pianissimo with NVIDIA NeMo. You can run it locally or on infrastructure you choose, or try it in the browser at klang.ai/pianissimo.
Want to adapt speech recognition to your own language or domain? Explore NVIDIA's ASR fine-tuning skill and NeMo training skill to guide a coding agent through experiment planning, data preparation, fine-tuning, and evaluation. Read the full NVIDIA fine-tuning blog post.
Open models made it possible for us to start Klang. With Pianissimo, we are building on NVIDIA's work and making the result available for others to use and adapt. Swedish is our first release; we plan to extend this work to more European languages.