Polish model

[abb128]

An initial Polish model has been trained on the 160 hours of Mozilla Common Voice using 80/10/10 train/test/dev speaker split, with 4.51% WER on unseen speakers (some of which may actually potentially not be unseen because Common Voice allows anonymous submissions and doesn't link them? not certain)

It is available here: https://april.sapples.net/april-polish-dev-2_pl.april

[phodina]

Could you link an article or lists of steps on how to train the model for different languages? I assumed you got the dataset from here.

What would be the next step if I'd like to train it for czech/slovak language?

[abb128]

@phodina From my testing, training on common voice data actually didn't work that well for Internet content with talking, because the speech is too high-quality and clear. The dataset is composed entirely of reading-speech, as the dataset collection is done by having people read written sentences out loud, and this produces a somewhat different kind of speech compared to natural talking or conversation.

I may write an article with some findings and instructions later, but for now I trained the model using this recipe with some modifications to use common voice instead of LibriSpeech, and I used this to finally export the checkpoint to a .april file.

[phodina]

Hi @abb128 , thanks for the explanation. I'll look at the recipe you suggested!

[dreamcat4]

Hello,
I could not follow what was being explained here. But you like to know if there is a good pathway to convert from Mozilla Common Voice --> LTSM --> .april model.

The desired language (for myself) is greek. However for any other languages, as a general workflow. It would be very helpful.

[EDIT]

But also: how to be informed / notified once new .april models gets added? To know to come back / check again.

[Doomsdayrs]

I may write an article with some findings and instructions later, but for now I trained the model using this recipe with some modifications to use common voice instead of LibriSpeech

What are the modifications you performed? Can you provide the patch file?