pruned NVIDIA's new 30B speech LLM from 67.5GB to 10GB with zero retraining, it still transcribes at 5.16% WER (teacher: 4.12%)
Automatic Speech Recognition model, built with NeMo. Download, explore files and run it on Hugging Face.
Automatic Speech Recognition model, built with NeMo. Download, explore files and run it on Hugging Face.