Pre-trained model - how are they created?

[etlweather]

Hello,

The docs https://stt.readthedocs.io/en/latest/AUGMENTATION.html mention various augmentation methods. I was wondering if the pre-trained models for download were created using these methods or they are straight from Librispeech and Common Voice unmodified?

I get OK results on my audio when I have very good recording but it quickly degrade when I use real-life recordings of meetings / phone calls as the audio has noise (like AC) and other "quality" (like roomy sounding, not full tonality as not close enough to mic, etc.) to it.

[reuben]

@JRMeyer ping

[JRMeyer]

Hi @etlweather,

there was quite a bit of augmentation used in training the v1.0 English model. The kind of augmentation was a simple variation on SpecAugment [1].

First 50 Epochs [1]

  "augment": [
    "frequency_mask[p=0.9,n=2:5,size=2]",
    "time_mask[p=0.9,n=3:4,size=25:100,domain=spectrogram]"
  ]

Second 50 Epochs [2]

  "augment": [
    "frequency_mask[p=0.9,n=5:7,size=2]",
    "time_mask[p=0.9,n=4:5,size=100:125,domain=spectrogram]"
  ]

Final 50 Epochs [3]

  "augment": [
    "frequency_mask[p=0.9,n=5:7,size=2]",
    "time_mask[p=0.9,n=4:5,size=100:125,domain=spectrogram]"
  ]