The audio used to calibrate importance matrices (imatrix) for GGUF ASR quantisation lives as a dataset on the Hub (not in git — it’s audio):
cstr/crispasr-imatrix-calib — CC0, 24 EN + 24 DE clips.
Clips are the first N mp3s of the dev split of Common Voice 17.0 (CC0),
via the fsicoli/common_voice_17_0 mirror. The Hub datasets loader no longer
runs the dataset script, so pull the raw tar shard and extract directly:
from huggingface_hub import hf_hub_download
import tarfile, os
for lang in ["en", "de"]:
p = hf_hub_download("fsicoli/common_voice_17_0",
f"audio/{lang}/dev/{lang}_dev_0.tar", repo_type="dataset")
os.makedirs(f"cv/{lang}", exist_ok=True)
with tarfile.open(p) as t:
n = 0
for m in t:
if not m.name.endswith(".mp3"):
continue
open(f"cv/{lang}/{lang}_{n:03d}.mp3", "wb").write(t.extractfile(m).read())
n += 1
if n >= 24:
break
crispasr decodes mp3 natively (miniaudio), so feed the clips straight in — see
../imatrix_ab.py and ../../docs/quantize.md.
Scale up for production: more clips, and the languages/domains you actually target — an English-only set regressed the imatrix in the A/B; EN+DE won.