GGUF + pure-C++ runtime in CrispASR (German XLSR-53)
#4
by cstr - opened
We've added the Jonatas Grosman XLSR-53 German fine-tune to CrispASR's wav2vec2 backend. C++ binary, GGUF — no Python.
CrispASR has a small fleet of German wav2vec2 GGUFs that all run through the same wav2vec2-ggml.cpp (we dispatch on GGUF metadata, not filename — pre-norm vs post-norm and the pos_conv variants are detected automatically):
- XLSR-53 German (this one) — Jonatas Grosman fine-tune
- XLSR-53 German CV13 — community CV13 fine-tune
- Base German CV9
- XLS-R-1B German, XLS-R-1B TEVR — larger 1B variants
Implementation note: our CNN frontend is a ggml graph (im2col + mul_mat, F32 to avoid F16 precision collapse through 7 conv layers) — 10.8× faster than the naïve scalar-loop C++ baseline.
CTC means no native punctuation/casing — the standard recipe is --punc-model fullstop-punc-q4_k.gguf (XLM-R-large, DE/EN/FR/IT, ~254 MB Q4_K) for restoration.
./build/bin/crispasr --backend wav2vec2 -m wav2vec2-xlsr-de-q4_k.gguf -l de \
-f audio.wav --punc-model fullstop-punc-q4_k.gguf