0
$\begingroup$

I'm working on a DL project to recognize (10 - 15) Arabic speech commands from a continuous stream of audio, and I want to create a dataset similar to Google's Speech Commands dataset.

Fortunately, I found a helpful answer, but I still have some questions before I start to collect the data from contributors for several days.

Q1: Can I use all the words from the Google's dataset as unknown classes in Arabic?

I'll record 16-bit mono PCM .wav audio files, sampled at 16000 Hz using Audacity as the following:

Record 10 audio files by the same person, such that each file will contain all different speech commands separated by 2 - 3 seconds of silence, and for each file:

  • Record with different distances from the microphone
  • And/or using different microphones (but the same computer)
  • And/or with different emotions (this way is a little bit hard)
  • And/or with different loudness of his/her voice ... etc.

Q2: Is there any useless way among these four ways?

Split .wav files into smaller files with only one word each, using this Voice Activity Detector for python.

Now, if there are 50 - 100 contributors, then the result is 500 - 1000 .wav for each speech command.

Extract the loudest section of each resulting audio file to make its length one second at most, and then I'll apply zero-padding on each audio file has a length less than one second.

Q3: Is it better if each resulting one-word audio file has different alignment rather than applying zero-padding such that all audio files will left-aligned?

kindly write down your suggested edits and recommendations if any.


Update: Maybe the reason that there is no answer or any comment here after two days is that my question needs some edits or adding some missing details. Therefore, if someone tells me how can I have better luck, I would appreciate it.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.