Instructions to use nvidia/stt_en_fastconformer_ctc_xlarge with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- NeMo
How to use nvidia/stt_en_fastconformer_ctc_xlarge with NeMo:
import nemo.collections.asr as nemo_asr asr_model = nemo_asr.models.ASRModel.from_pretrained("nvidia/stt_en_fastconformer_ctc_xlarge") transcriptions = asr_model.transcribe(["file.wav"]) - Notebooks
- Google Colab
- Kaggle
GGUF + pure-C++ runtime in CrispASR — FastConformer-CTC XL
We've added the FastConformer-CTC XL size to CrispASR's fastconformer-ctc backend. C++ binary, GGUF — no NeMo. (Same runtime supports all three NeMo sizes — large, xlarge, xxlarge.)
Same FastConformer encoder as our parakeet / canary backends (shared core/fastconformer.h — conv subsampling + MHA with rel-PE), just a different decode head: greedy CTC. Lowest-latency English path in CrispASR — no autoregression.
CTC means no native punctuation, so the recipe is to add --punc-model fireredpunc-q8_0.gguf for English+Chinese caps/punc (or the XLM-R-based fullstop-punc-q4_k.gguf for DE/FR/IT). Word-level timing via -am canary-ctc-aligner.gguf.
Pre-quantised GGUFs (CC-BY-4.0): cstr/stt-en-fastconformer-ctc-xlarge-GGUF
./build/bin/crispasr --backend fastconformer-ctc \
-m stt-en-fastconformer-ctc-xlarge-q4_k.gguf \
-f audio.wav --punc-model fireredpunc-q8_0.gguf -osrt