QuartzNet

Implementation of QuartzNet ASR model in PyTorch

Usage

Setup

To launch and inference in nvidia-docker container follow these instructions:

0. Install nvidia-docker
1. Run ./docker-build.sh

Training

To launch training follow these instructions:

1. Set preferred configurations in config/config.yaml. In particular you might want to set dataset: it can be either numbers or librispeech
2. In docker-run.sh change memory, memory-swap, shm-size, cpuset-cpus, gpus, and data volume to desired values
3. Set WANDB_API_KEY environment variable to your wandb key
4. Run ./docker-train.sh

All outputs including models will be saved to outputs dir.

Inference

To launch inference run the following command:

./docker-inference.sh model_path device bpe_path input_path

Where:

• model_path is a path to .pth model file
• device is the device to inference on: either 'cpu', 'cuda' or cuda device number
• bpe_path is a path to yttm bpe model .model file
• input_path is a path to input audio file to parse text from

Predicted output will be printed to stdout and saved into a file in inferenced folder

Pretrained models

My currently best model trained on librispeech and the respective config can be downloaded here.

It is not very good however because I only trained it to ~59 WER on librispeech