Word error rate, not perplexity: perplexity is meaningless for ASR and file
size proves nothing. Both models saw identical audio and decoded greedily, so
the only variable is the weights.
WER
whisper-large-v3-turbo
5.20%
Exeaon1-Voice-0.8B
5.20%
token disagreement
0.12%
No measurable degradation. The original's WER is reported alongside because
a compressed model can only be judged against what the model could do in the
first place — if the original errs on a clip, the compressed one repeating that
error is not damage we caused.
Sample, same clip, both models:
ref MISTER QUILTER IS THE APOSTLE OF THE MIDDLE CLASSES AND WE ARE GLAD TO
WELCOME HIS GOSPEL
base Mr. Quilter is the apostle of the middle classes, and we are glad to
welcome his gospel.
ours Mr. Quilter is the apostle of the middle classes, and we are glad to
welcome his gospel.
Why audio compresses this well
Whisper is an encoder-decoder: 32 encoder layers carry most of the parameters
and 4 decoder layers the rest. Both stacks are compensated — compressing only
the larger one would leave an eighth of the model on plain rounding.
Calibration uses real speech. The encoder's activations are dominated by
mel-spectrogram structure that random input does not reproduce, so noise
calibration would compensate against statistics the model never sees.
Convolutional weights in the audio frontend are left dense: conv1.weight is
[1280, 128, 3], a kernel of 3 against a group size of 128, which cannot be
usefully quantized and is a negligible share of parameters.
The container bundles preprocessor_config.json, so the feature extractor
builds without fetching anything from the base repository.
Limitations
Evaluated on read English speech (LibriSpeech-style). Accented, noisy,
overlapping or non-English audio is not covered by the number above.
WER was measured on a small clip set; treat 5.20% as an indicative figure on
clean speech, not a benchmark-suite result.
Inherits every limitation and bias of the base model.
Not evaluated for safety-critical, medical or legal transcription.
Licence and attribution
Derived from
openai/whisper-large-v3-turbo,
released under the MIT licence, which permits redistribution of modified
versions. The base repository ships no LICENSE file; the licence is declared
in its model card metadata and is reproduced in this repository.
Ours: the compression method, calibration, packaging, runtime.
Not ours: the pretrained knowledge, which comes from OpenAI. This model is
not endorsed by or affiliated with OpenAI.
Citation
bibtex
1@misc{exeaon2026,
2 title = {Exeaon: compressed models that run, and train, without
3 decompressing},
4 author = {Akpalu, Elliot Elikplim},
5 year = {2026},
6 publisher = {Zenux Plimver Technologies LTD},
7 url = {https://huggingface.co/Exeaon}
8}
