Converting a PyTorch QuartzNet Model

NeMo project provides the QuartzNet model.

Downloading the Pre-trained QuartzNet Model

To download the pre-trained model, refer to the NeMo Speech Models Catalog. Here are the instructions on how to obtain QuartzNet in ONNX format.

  1. Install the NeMo toolkit, using the instructions.

  2. Run the following code:

    import nemo
    import nemo.collections.asr as nemo_asr
    quartznet = nemo_asr.models.EncDecCTCModel.from_pretrained(model_name="QuartzNet15x5Base-En")
    # Export QuartzNet model to ONNX format

    This code produces 3 ONNX model files: encoder_qn.onnx, decoder_qn.onnx, qn.onnx. They are decoder, encoder, and a combined decoder(encoder(x)) models, respectively.

Converting an ONNX QuartzNet model to IR

If using a combined model:

mo --input_model <MODEL_DIR>/qt.onnx --input_shape [B,64,X]

If using separate models:

mo --input_model <MODEL_DIR>/encoder_qt.onnx --input_shape [B,64,X]
mo --input_model <MODEL_DIR>/decoder_qt.onnx --input_shape [B,1024,Y]

Where shape is determined by the audio file Mel-Spectrogram length: B - batch dimension, X - dimension based on the input length, Y - determined by encoder output, usually X / 2.