Training details of jasper10x5_LibriSpeech_nvgrad

[GabrielLin]

Could you please show me more details of jasper10x5_LibriSpeech_nvgrad pre-train model? Such as hardware, training time, whether to use external data or not, whether to use synthetic data or not. Thanks.

[okuchaiev]

@blisc could you please comment on this

[blisc]

The pre-trained model can be achieved as is with the jasper10x5_LibriSpeech_nvgrad.py example config.

• No data external to LibriSpeech is used (no synthetic data was used).
• We used 3-fold speed perturbation (0.9, 1.0, 1.1) as data augmentation
• It was done on 8xV100

In additional, compared to older Jasper models:

• We switched from python_speech_features to librosa for pre-processing
• We switched from SGD to Novograd
• We added dithering
• Input spectrogram is now normalized per frequency bin as opposed to over the whole spectrogram

[GabrielLin]

Thank you for your valuable information. @blisc