This is a checkpoint for QuartzNet 15x5 trained only on LibriSpeech (speed perturbed) using NeMo, and is the one mentioned in the QuartzNet paper under section 4.1 LibriSpeech. It was trained with Apex/Amp optimization level O0 for 400 epochs.
The model achieves a greedy WER of 3.83% on LibriSpeech dev-clean, 11.08% on dev-other, 3.90% on test-clean, and 11.28% on test-other.
The files included in this model are:
The source code and developer guide is available at https://github.com/NVIDIA/NeMo.
Usage example: Download the checkpoint files and place them in a checkpoint directory. Then, run jasper_eval.py (from NeMo's ASR examples).
python jasper_eval.py --model_config=$nemo_root/nemo/examples/asr/configs/quartznet15x5.yaml --eval_datasets=test.json --load_dir=$checkpoint_dir