KlangAI/pianissimo-da-onnx

Klang Pianissimo for Danish: ONNX

ONNX versions of Klang Pianissimo for Danish, our Danish speech recognition model. They run on an ordinary CPU with onnxruntime, without PyTorch or NeMo.

Versions

Version	Download	FLEURS test	Speed, Intel	Speed, ARM	Speed, Mac	Memory, 30 s clip
Original fp32	2.51 GB	7.43
int8	684 MB	7.60	35×	63×	74×	1.5 GB
int4	525 MB	7.98	12×	23×	37×	1.1 GB
fp16	1.46 GB	7.43	18×	30×	55×	4.0 GB
fp32	2.56 GB	7.43	18×	30×	55×	3.0 GB

The FLEURS column shows word error rate (WER, %) on all 930 Danish test clips, scored with our evaluation normaliser; lower is better. Speed is audio duration divided by processing time (RTFx), measured on a 30-second clip. Speed and memory were measured on the Swedish model, which has the same architecture. CPU benchmarks used 4 threads on an Intel Xeon Platinum 8481C, 8 threads on a Google Axion (ARM) and 6 threads on an Apple M5 Pro. Memory was measured on the Intel Xeon.

On a Mac, the MLX versions run on the GPU and are about twice as fast.

Usage

Install onnx-asr:

pip install onnx-asr[cpu,hub]

Transcribe a 16 kHz mono WAV file:

import onnx_asr

model = onnx_asr.load_model("KlangAI/pianissimo-da-onnx", quantization="int8")
print(model.recognize("audio.wav"))

Only the files for the chosen version are downloaded. Use quantization="int4" or "fp16", or leave it out for fp32.

With timestamps:

result = model.with_timestamps().recognize("audio.wav")
print(result.text)
for token, start in zip(result.tokens, result.timestamps):
    print(f"{start:6.2f}  {token}")

Long recordings

The transcribe_long helper included in this repository splits long recordings at quiet moments; install its audio dependency with pip install soundfile:

import sys
from huggingface_hub import snapshot_download

path = snapshot_download("KlangAI/pianissimo-da-onnx", allow_patterns=["*.py", "tokenizer.model"])
sys.path.append(path)
import pianissimo_onnx

for start, end, text in pianissimo_onnx.transcribe_long(model, "long.wav"):
    print(f"[{start:7.1f}s] {text}")

onnx-asr can also split at pauses with voice activity detection (model.with_vad(onnx_asr.load_vad("silero"))). It produces shorter segments, which improved accuracy on noisy broadcast audio but reduced it on clean speech in our tests.

Vocabulary

Our implementation of NeMo's phrase boosting supports custom vocabulary and matches NeMo's results on the original model. List names and terms to help the model spell them correctly. Install sentencepiece with pip install sentencepiece.

import sys
from huggingface_hub import snapshot_download

path = snapshot_download(
    "KlangAI/pianissimo-da-onnx",
    allow_patterns=["*.py", "tokenizer.model"],
)
sys.path.append(path)
import pianissimo_onnx

vocabulary = ["Klang AI", "Pianissimo", "Folketinget", "Aarhus Universitet"]
pianissimo_onnx.boost(model, vocabulary)
print(model.recognize("audio.wav"))

See the main model card for vocabulary boosting with NeMo.

Files

File	Contents
encoder-model{.int8,.int4,.fp16,}.onnx	Encoder. The fp32 weights are in encoder-model.onnx.data.
decoder_joint-model{.int8,.int4,.fp16,}.onnx	Decoder
vocab.txt	The 8,192 tokens, plus the blank token
nemo128.onnx	Converts audio to the 128-band spectrogram the encoder expects
config.json	Settings for onnx-asr

How they were made

All versions are exported from the released checkpoint and retain its local attention (256 frames on each side).

• int8: weights and calculations in 8 bits. Before quantizing, part of the scale of each layer's input is moved into its weights (SmoothQuant), which makes the 8-bit version more accurate. The decoder's LSTM is stored in fp16.
• int4: weights in 4 bits, in groups of 32 with a scale and offset per group chosen with HQQ. Calculations are done in fp32. Plain rounding to 4 bits was about 1 percentage point worse on the Danish test sets. The int4 encoder's offsets are stored as whole numbers plus a small exact correction, so it also runs on WebGPU in the browser (onnxruntime-web).
• fp16: weights stored in fp16.

FLEURS WER was measured with onnxruntime 1.24.1 on the CPU. See the main model card for comparisons with other models.

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}
}