What is the training set of the pre trained model

[JRMeyer]

>>> jackhuang
[December 19, 2017, 7:40am]

I used the TIMIT dataset to test the pre-trained model, and the wer is
about 27%. I want to know the training set of the pre-trained model, so
I can try to improve the selection strategy of choosing training set.
Can anyone help me?

[This is an archived TTS discussion thread from discourse.mozilla.org/t/what-is-the-training-set-of-the-pre-trained-model]

[JRMeyer]

>>> kdavis
[December 19, 2017, 11:16am]

The pre-trained model was trained on Fisher, Switchboard, and Librivox
training data sets.

[Archived Post]

[JRMeyer]

>>> jackhuang
[December 19, 2017, 11:34am]

Combining these three sets to be the training set and validation set？

[Archived Post]

[JRMeyer]

>>> kdavis
[December 19, 2017, 12:50pm]

No.

The training set was the Fisher, Switchboard, and Librivox training data
sets. slash
The validation set was the clean Librivox validation data set. slash
The test set was the clean Librivox test data set.

[Archived Post]

[JRMeyer]

>>> jackhuang
[December 20, 2017, 3:57am]

Thank you, and is the language model the 4-gram language model with a
30,000 word vocabulary trained on the Fisher and Switchboard
transcriptions as the paper says?

[Archived Post]

[JRMeyer]

>>> kdavis
[December 20, 2017, 9:37am]

No. We didn't try to exactly reproduce the paper's results.

We created a KenLM language model based off of Fisher, Switchboard,
andLibrivox training data sets as well as part of Wikipedia.

[Archived Post]

[JRMeyer]

>>> rao_jjn
[December 20, 2017, 12:02pm]

Is it possible to train this model further using other voice datasets ?

[Archived Post]

[JRMeyer]

>>> yv001
[December 20, 2017, 12:27pm]

I don't think that's possible until tensorflow checkpoints are also
published - frozen out_graph.pb cannot be used for further training
AFAIK

[Archived Post]

[JRMeyer]

>>> jackhuang
[December 22, 2017, 6:32am]

Does the common Voice data set include other data sets like Librivox,
Switchboard, Fisher?

[Archived Post]

[JRMeyer]

>>> kdavis
[December 22, 2017, 7:28am]

No. Common Voice, Librivox, Switchboard, and Fisher are separate,
distinct data sets.

[Archived Post]

[JRMeyer]

>>> jackhuang
[December 25, 2017, 3:45am]

Thank you, and may I ask the WER value of the clean Librivox test data
set?

[Archived Post]

[JRMeyer]

>>> jackhuang
[December 27, 2017, 3:40am]

And did you use the 4-gram model to train the language model?

[Archived Post]

[JRMeyer]

>>> kdavis
[December 27, 2017, 8:12am]

The WER is 6.0 percent for the Librivox clean, test data set

[Archived Post]

[JRMeyer]

>>> kdavis
[December 27, 2017, 8:13am]

Not sure what you're asking here.

Do you mean 'Is the language model a 4-gram language model?'

[Archived Post]

[JRMeyer]

>>> jackhuang
[December 27, 2017, 8:20am]

Yes, that't what I want to express.

[Archived Post]

[JRMeyer]

>>> panybj
[December 29, 2017, 12:50pm]



> Fisher, Switchboard,

Where can I download Fisher and Switchboard datasets?

[Archived Post]

[JRMeyer]

>>> kdavis
[January 2, 2018, 9:34am]

You have to purchase Fisher and Switchboad from
LDC.

[Archived Post]

[JRMeyer]

>>> srikar
[September 11, 2018, 10:46am]

So, the Common Voice data is not factored into the pre-trained model?

[Archived Post]