When will you release deepspeech pretrained model v0.2.0?


(Murugan R) #1

Indian accent for English with technical terms have a lot of troubles. Our v0.1.1 training dataset is really focused on American English with mostly no accent.

So @lissyx sir said

v0.2.0 pretrained model may be resolve this issue but not sure. because this is focused on Common Voice audio datasets.

@lissyx sir don’t clearly known about when that pretrained model release?

@reuben sir Do you know that pretrained model release date?

thank you so much sir


(Bacon Ator) #2

The 0.2.0 update is now available, check the github page.


(Murugan R) #3

yeah, yes sir. thank you so mush sir


(David0861) #4

You can download the model here: https://github.com/mozilla/DeepSpeech/releases


(Murugan R) #5

sir @david0861 thank you sir. i was downloaded and run that pretrained model. it gives good results. but it is not detecting Indian Accent pronunciation sentence. how can i improve Indian accent accuracy sir.?


(David0861) #6

You have to re-train the model with Indian Accent data. So you have to create the .csv files with the new data, download checkpoints and follow the instruction to re-train.

Pd: if you don’t have enough data you can “create” more with “data augmentation” technique. You can follow this parts on this post.


(Murugan R) #7

@david0861 sir awesome wonderful hint sir. Thank you so much sir.

if you don’t have enough data you can “create” more with “data augmentation” technique. You can follow this parts on this post

i have one more doubt in data_augmentation,

DATA AUGMENTATION :

record same sentence with different distances from microphone, and in different locations in room/house

sir if i record my own audio with different locations, distances with the same sentence, it is particularly working in a command control based application only satisfy.

but i need a common voice or speech recognition model. is it satisfy common speech application


(David0861) #8

About data augmentation you have just follow the second method:

2/ **Other way** : A mozilla tool to modify wav’s with echos, pitch, adding noise…


(Vincent Foucault) #9

Hi, @muruganrajenthirean

A good way is add the max samples as possible, with the max variations possibilities :

  • record your voice with all possibilities (distance, tone, speed, ambiant noise)

  • augment data with the mozilla tool :
    – add echo
    – change pitch (stay in correct values)
    – change speed (stay in correct values)
    – add noise (Your noise -> your environment)
    You should really see improvment in accuracy !
    hope it will help you

PS : this discussion isn’t in the correct topic ! If you need more explications, open a specific topic


(Murugan R) #10

definitely sir. now i am open a new topic sir


(Vincent Foucault) #11

Yes i saw that, sorry !


(Murugan R) #12

:slightly_smiling_face:


(Lissyx) #13

I know @Tilman_Kamp was working on that :slight_smile: