Train Deep speech with Indian english

(Lissyx) #3

For context, using a french dataset produced from audiobooks on top of english model did provide interesting results with around 100 hours of data.

@bhadwal.abhishek Can you provide more context on your case? Amount and source of data, any pre-processing happening, your language model, etc.

0 Likes

(Bhadwal Abhishek) #4

Thanks buddy. But what was the source of your data. I mean from where did you take the data from and what was the size of your data set .

0 Likes

(Bhadwal Abhishek) #5

Hi buddy,
data and all i still have to collect.
Can i use the same language model which have been used for Deep speech 2 pre trained model ?

0 Likes

(Murugan R) #6

i recorded my own audios 1500. and augment different pitch,distance etc…get 20000 more audio files.

0 Likes

(Bhadwal Abhishek) #7

okay,
But other than that did you do anything else I mean did you create your own language model or any changes any where in deep speech ?

0 Likes

(Murugan R) #8

yes i am using DeepSpeech v0.2.0 pretrained model sir. and i am finetuning this with my audio files.

0 Likes

(Bhadwal Abhishek) #9

And how much time and what hardware you used for the same .

0 Likes

(Bhadwal Abhishek) #10

are you done with it .?
what was the accuracy you got in the same ?

0 Likes

(Murugan R) #11

4 GB GTX 1500 series GPU. it is taking 35 hours continues training sir.

are you done with it .?
what was the accuracy you got in the same ?

yes i did this model with my own audio, it gives good results for indian english accent.

0 Likes

(Bhadwal Abhishek) #12

And when you said you have done some fine tuning …
does it mean just adding the new data set on top the already pre trained model or something else also ?

0 Likes

(Murugan R) #13

just adding the new data set on top the already pre trained model

yes pretrained + own audiofiles( fine tune) generating a new model.

0 Likes

(Bhadwal Abhishek) #14

thanks a lot buddy you are really a life saver…
pls let me know if you do social service to donate the data set : )

0 Likes

(Murugan R) #15

my datasets only a particular purpose only recorded audio. that is not covered all. this is one disadvantage for us. if you want more data, you tried to record different kind of speaker and do data augmentation sir.

i have a code for data augmentation UI. it is very easy to handle augmentation process.

if you want data please do this. it will help to you lot.

come for ready to help page.

0 Likes

(Bhadwal Abhishek) #16

okay …
thanks a ton brother :slight_smile:

0 Likes

(Murugan R) #17

:slightly_smiling_face:

0 Likes

(Bhadwal Abhishek) #18

hey buddy … i was wondering if you could share your pre trained model …
as it requires alot of hardware … and it ll take time…
i just wanted to see how it 'll work …
if it is okay with you .

Thanks

0 Likes

(Sudarshan Gurav14) #19

@bhadwal.abhishek hello sir how to train our dataset can you please help me?
I am fresher I want to learn how to retrain the deep speech model for Indian accent? @lissyx when I am doing that I was facing an error which is key_layer/_1 bais not found in checkpoint 0.4.1 I am continuing the training from release model

0 Likes

(Lissyx) #20

Please continue training using the proper git branch v0.4.1 and not master

0 Likes

(Sudarshan Gurav14) #21

will try sir thanks you for immediate reply

0 Likes

(Sudarshan Gurav14) #22

@lissyx I am creating a pbmm model after that we want testing that model I am getting the error which is RuntimeError: CreateModel failed with error code 5

Command is: deepspeech --model modell/output_graph.pbmm --alphabet modell/alphabet.txt --lm modell/lm.binary --trie modell/trie --audio how_are_you.wav

0 Likes