CoolFace
Modelpublic

Reza2kn/nvidia_stt_fa_fastconformer_hybrid_large-NVFP4

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face
0likes16downloads
Model Card

nvidiasttfafastconformerhybrid_large-NVFP4

NVFP4 (W4A4) post-training quantization of `nvidia/stt_fa_fastconformer_hybrid_large` via NVIDIA modelopt.

  • —Base architecture: EncDecHybridRNNTCTCBPEModel (NeMo)
  • —Calibration: 32 Persian clips from Reza2kn/persian-asr-eval-v0 (held out from eval).
  • —Hardware target: NVIDIA Blackwell tensor cores.

Eval — Reza2kn/persian-asr-eval-v0 (FLEURS-fa, 200 clips)

VariantWER ↓CER ↓per-clip latencypeak VRAM
NVFP4 (this repo)33.06%10.91%37 ms603 MiB

Usage

python
import nemo.collections.asr as nemo_asr
m = nemo_asr.models.ASRModel.restore_from("nvidia_stt_fa_fastconformer_hybrid_large-NVFP4.nemo").cuda().eval()
transcripts = m.transcribe(["clip.wav"])
print(transcripts[0])

License

Inherits the base model's license.