K

Klang Pianissimo

Open Swedish ASR model from Klang AI - ~64x faster than KB-Whisper large at slightly lower accuracy

Excellent Swedish Support

Features

Real-time Transcription Yes
Speaker Identification No
AI Summaries No
Action Items No
Meeting Bot (auto-joins calls) Not verified

Core Features

  • 0.6B-param FastConformer encoder + Token-and-Duration Transducer (TDT) decoder; finetune of nvidia/parakeet-tdt-0.6b-v3 (exact repo ID verbatim in the model-card YAML)
  • 24-layer encoder, 8x subsampling, local attention (~20.5s context each side), 2.51 GB checkpoint
  • Trained on ~50,000 hours of Swedish (RixVox, Common Voice, FLEURS + internal)
  • Punctuation, capitalisation, word-level timestamps, NeMo phrase boosting for custom vocabulary
  • Released 2026-09-23 under CC BY 4.0; API via Berget AI (Sweden)
  • ON-DEVICE variants shipped ~2026-10-01: ONNX int8/int4/fp16/fp32 (any CPU, 515 MB-2.56 GB, 12-74x RTFx) and MLX bf16/8-bit/4-bit/fp32 (Apple silicon, up to 173x RTFx) - Swedish ASR now runs on a laptop
  • Danish/Norwegian: UPGRADED 2026-10-23 from soft roadmap to publicly stated - a co-founder told the Pianissimo hackathon that Danish is the second model and Norwegian follows, and that the Klang Dialects benchmark is already extended to both. Still no dates and nothing written on the model card. The room Q&A leaned toward one model per language rather than a multilingual one (UNCERTAIN - damaged recording)
  • Third-party ecosystem: klang.ai/pianissimo listed 15 tools on 2026-10-24 - 13 by named outsiders, 2 by Klang (the early-October e-mail counted six DICTATION tools and gave no total, so the directory's growth rate is not established; dictation went six to eight), across Mac, Windows, Linux, Chrome, iPhone and web; two do regulated work (BBIC/IBIC/ASI drafting; de-identification). Berget AI named on the page as the launch inference partner (Sweden, zero retention) - this resolves the "Valiant"/"palliates" in both hackathon transcripts

Languages

Swedish

Pricing

🎁
Free Tier Available

Open weights, CC BY 4.0

Pricing model: Free

Integrations

Berget AI Hugging Face ONNX MLX CoreML

Latest News

New Feature 2026-09-23

Klang Releases Pianissimo, an Open Swedish Speech Model 64x Faster Than KB-Whisper

Klang published Pianissimo, a 0.6B-parameter Swedish speech recognition model under a CC BY 4.0 licence, fine-tuned from NVIDIA Parakeet and trained on roughly 50,000 hours of Swedish. It transcribes at about 2,500 times realtime against KB-Whisper large at 39 times, so roughly 64 times faster. On accuracy, Klang's own published comparison puts KB-Whisper large ahead on all three Swedish datasets tested, including Klang's own dialect benchmark, at 3.91, 5.08 and 2.30 percent word error rate against Pianissimo at 4.46, 6.51 and 4.85. On-device variants for ONNX and Apple silicon followed in October, so Swedish speech recognition now runs on a laptop.

User Reviews

No user reviews yet. Be the first to write a review!

Quick feedback

Hi! Do you use Klang Pianissimo?