GGUF + pure-C++ runtime in CrispASR — FastConformer-CTC XXL

#2
by cstr - opened

We've added the FastConformer-CTC XXL size to CrispASR. C++ binary, GGUF — no NeMo.

Same fastconformer-ctc backend as the large/xlarge siblings — shared core/fastconformer.h encoder (conv subsampling + MHA with rel-PE), greedy CTC head. CTC = no autoregression, no native punctuation.

Punc recipe for any CTC backend in CrispASR: --punc-model fireredpunc-q8_0.gguf (BERT-base, EN+CN) or fullstop-punc-q4_k.gguf (XLM-R-large, EN/DE/FR/IT). Both ship as GGUF.

Pre-quantised GGUFs (CC-BY-4.0): cstr/stt-en-fastconformer-ctc-xxlarge-GGUF

./build/bin/crispasr --backend fastconformer-ctc \
    -m stt-en-fastconformer-ctc-xxlarge-q4_k.gguf \
    -f audio.wav --punc-model fireredpunc-q8_0.gguf -osrt

Sibling sizes: large, xlarge.

Sign up or log in to comment