KlangAI/pianissimo-da-gguf

Klang Pianissimo for Danish: GGUF

GGUF versions of Klang Pianissimo for Danish, our Danish speech recognition model, for transcribe.cpp. They run on the CPU or GPU of an ordinary computer, without Python, PyTorch or NeMo.

Versions

File	Download	WER	Same transcript as the original
Original fp32 (NeMo)	2.51 GB	6.5
pianissimo-da-Q8_0.gguf	740 MB	6.4	183 of 200
pianissimo-da-Q5_K_M.gguf	549 MB	6.5	159 of 200
pianissimo-da-Q4_K_M.gguf	485 MB	6.5	142 of 200
pianissimo-da-F16.gguf	1.26 GB	6.6	187 of 200
pianissimo-da-F32.gguf	2.51 GB	6.6	188 of 200

Word error rate (WER, %) on 200 Danish FLEURS test clips; lower is better. The last column counts clips with transcripts identical to the original model. We recommend Q8_0: about a third of the original's download size with comparable accuracy on this test.

Usage

Build transcribe.cpp (see its README), download a file and transcribe a 16 kHz mono WAV file:

hf download KlangAI/pianissimo-da-gguf pianissimo-da-Q8_0.gguf --local-dir models
build/bin/transcribe-cli -m models/pianissimo-da-Q8_0.gguf -l da audio.wav

Always pass -l da. Recordings of any length can be passed in one call: transcribe.cpp runs the model with the same local attention as the original (256 encoder frames, about 20 seconds, on each side).

For Python, use the original model with NeMo or the ONNX versions; on a Mac, the MLX versions run on the GPU.

How they were made

Converted from the released checkpoint with transcribe.cpp's own convert-parakeet.py and quantized with transcribe-quantize (transcribe.cpp commit 5700f12). The converter uses its standard Parakeet v3 settings: Pianissimo retains v3's architecture and tokenizer, so only the weights differ.

License and citation

Released under CC BY 4.0, like the original model.

@misc{klang2026pianissimoda,
  title = {Klang Pianissimo for Danish},
  author = {{Klang}},
  year = {2026},
  howpublished = {Hugging Face model repository},
  url = {https://huggingface.co/KlangAI/pianissimo-da}
}