Hello . I downloaded the following file from https://github.com/mozilla/DeepSpeech-examples/tree/r0.8/net_framework GateHub and ran it in Visual Studio software on Windows, but when I say a word through a microphone, for example:
HI
you are
two
on
Output text :
I
you are
to
o
Going back to this, I want to know how I can improve its accuracy for testing. I came across scorer in the documents, but I do not know how to use this in Windows. Thank you for your help.
Please read the docs and search before you post, for .NET you can use:
sttClient.EnableExternalScorer(scorer ?? “kenlm.scorer”);
Hello . I said words like close , you are , Hi . through the microphone, but the output showed me these. What can I do to improve the DeepSpeak model through the microphone? I just want to use deep speech , which means I do not want to teach datasets.
You don’t provide any context, and poor detection can be related to a lot of factors: speed of speech, accent, etc.
Not to mention you are using example code, which is provided as-is.
I said a word like hello, close, etc. through the microphone in English with a delay of about 5 seconds, but I did not get a good output. My Native language is Persian
Another test I said in a sentence like How are you without pausing but still did not get a good result.
Sorry, but I have already explained to you:
- example code, we can’t guarantee there is no bug
- model is mostly trained on american english
- you are speaking with persian accent
thank you . Now, if I test words with an American English accent through the Logman Dictionary, for example, will I get a good result?
And can I teach Mozilla on a 50 GB English database to improve Deep speech mockery? Do I have to do fine tune or transfer learning ?
As said previously, we can’t guarantee, I don’t know what the logman dictionary does here.
Again, it depends on what data that is … If you have 50GB of English in Persian accent, you can fine-tune the released english checkpoints. Please check the docs.
thank you . By 50GB I mean the following database
You are mixing a lot of different subjects here.
-
Test some audio that you record with a mic directly with command line DeepSpeech to see how it does with your accent.
-
Try the 0.9.1 model.
-
Understand how it works then use it in .NET.
-
Common voice or a dictionary won’t do much good if you are trying to recognize a Persian accent. You will need hundreds of hours of Persian accent to train.
1 Like
We are documenting that Common Voice English is used for training: Release DeepSpeech 0.8.2 · mozilla/DeepSpeech · GitHub