About the MOCKINGJAY model about s3prl

[andi611]

Hi,

Do you have any results showing the difference between linear spectrogram and melspectrogram?

Empirically, I find reconstruction target has little effect on phone classification or speaker classification.

Also, is it possible that you provide some pre-trained models for the linear spectrogram?

I am currently training a new model with linear as output reconstruction target.
I'm using the following setting:

win_ms: 25
hop_ms: 10
n_freq: 201
n_mels: 80

input:
  feat_type: mel
  log: True
  delta: 1
  cmvn: True

target:
  feat_type: linear
  log: True
  delta: 0
  cmvn: True

Hope you will find it useful. I'll comment here once its done.

[andi611]

The model is now available in the Mockingjay directory under the S3PRL Google Drive.
The name of the ckpt is: on-the-fly-melLinearBase960-b12-d03-T-libri
A simple tutorial on how to use this model can be found here.