cstr/crispasr-imatrix-calib

CrispASR imatrix calibration set — Common Voice EN + DE

A tiny, CC0, multilingual read-speech sample used to compute importance matrices (imatrix) for GGUF quantisation of ASR models with CrispASR.

• en/ — 24 English clips
• de/ — 24 German clips

Provenance

Clips are drawn from the dev split of Mozilla Common Voice 17.0 (via the fsicoli/common_voice_17_0 mirror), which is released under CC0 1.0 (public domain). Re-distributed here unchanged, same licence.

Why this exists

llama.cpp-style imatrix quantisation improves low-bit quality by weighting per-tensor quantisation error by the activation energy the model actually uses. For audio models that means running audio through the model, not a text corpus (which is what the common calibration_datav3 text file does — it only calibrates the text decoder). There is no off-the-shelf audio imatrix corpus, so this is a clean-licence starting point.

Language coverage matters. In CrispASR's A/B harness (tools/imatrix_ab.py), calibrating qwen3-asr-0.6b q4_k on this EN+DE set improved prefill-logit cosine vs the f16 gold from 0.890 → 0.941 (+0.051), every held-out clip up — whereas an English-only corpus regressed it. Calibrate on the languages/domains you actually target, and scale this up (more clips, more languages) for production.

Use

export CRISPASR_IMATRIX_OUT=model.imatrix.gguf
for f in en/*.mp3 de/*.mp3; do
    crispasr -m model-f16.gguf -f "$f"   # merges into the imatrix each run
done
crispasr-quantize model-f16.gguf model-q4_k.gguf q4_k --imatrix model.imatrix.gguf

See docs/quantize.md.