How can i improve Indian accent accuracy for pretrained model v0.2.0.?

i was downloaded and run that pretrained model v0.2.0. it gives good results. but it is not detecting Indian Accent pronunciation sentence. how can i improve Indian accent accuracy sir.?

Either fine-tune with enough English Indian accent data, or contribute in-mass to Common Voice :), or both :slight_smile:

@lissyx sir

fine-tune with enough English Indian accent data

how much amount of data for example i have a half an hour of indian accent audio files is there. is it enough for fine tuning a pretrained model sir?

sir which version -> deepspeech, tensorflow, & native client are needed?
thank you sir

Please refer to other posts here that are documenting fine-tuning, but it’s likely you need something like tenths or 100 hours ?

Please refer to the documentation.

@lissyx sir ,

it’s likely you need something like tenths or 100 hours

we are give an 10-100 hours of audio files, what duration of individual audio files(ex. 3s -30s) we given an input?. it must match with the similar transcript.

if the individual audio files is large then it mismatching the transcript. so what duration of audio files successfully read our deepspeech.

if any possible is there to download for mozila common voice contributed audio files with an Indian accent.?

i have a my own vocab.txt language model. and then deepspeech language model. i combine these two to build a new language model. if there is any possibilities available or not?

thank you so much sir. :slightly_smiling_face::slightly_smiling_face::slightly_smiling_face:

You should ask that to Common Voice team :slight_smile:

I’m not sure to follow you, it feels like you are describing data that has no value for training: you need audio + matching transcription, and 3 secs to 30 secs should be a good deal.

You should be able to augment our vocab.txt with your own data to rebuild it, this way it will be combined.

@muruganrajenthirean https://github.com/mozilla/DeepSpeech/issues/821

@lissyx sir,

11h](How can i improve Indian accent accuracy for pretrained model v0.2.0.?)

@muruganrajenthirean https://github.com/mozilla/DeepSpeech/issues/821

i was referred srutheesh,

Please ping me on sruteeshkumar@gmail.com if you still need help with youtube dataset etc…

srutheesh sir was replied. very nice sir. i asked to srutheesh,“how can i handle youtube dataset with the indian accent. then segmented with a correct transcript”.

srutheesh sir given youtube indian accent videos links file, segment with 30s audio script, and some instruction.

it is very useful. and i try to do my next step vey confidently sir.

thank you sir for your kindness reply.:slightly_smiling_face::slightly_smiling_face::slightly_smiling_face:

@muruganrajenthirean could you please explain to me as well how you got everything (audio files, matching transcriptions, length of audio files etc.) for Indian accent English?

1 Like

I am trying to create a speech recognition dataset especially for Indian Accents.
I am taking help from colleagues to build this.
Daily i send a article link and ask them to record and upload to google drive.
I have a problem with this approach.
All audio recordings of length 5 -7 min.
I am using DeepSpeech model for this and it requires 10 sec audio sentences .
Suggest me any approach if possible to segment audio files into corresponding sentence phrases or to build a better with 5 min length audio files.
Suggestions are more than welcome on better way to create a speech to text dataset.

write a python code and cut the entire mp3 into 10 secs length using ffmpeg command and listen to the audio and create their corresponding transcripts. You can aswell use aeneas library for auto alignment but accuracy might be low