3

I'm going to train my neural network in recognizing if a person is male or female from audio speech samples. Unfortunately the only datasets I could find, were dedicated for natural language processing.

Does anyone know where to find wav files labeled with this kind of information (male or female)? I'm interested in rather short samples of a person's voice (from 1s to 20s).

|improve this question
3

Mozilla just released the Common Voice dataset

https://voice.mozilla.org/en/datasets

Each entry in the dataset consists of a unique MP3 and corresponding text file. Many of the 1,368 recorded hours in the dataset also include demographic metadata like age, sex, and accent that can help train the accuracy of speech recognition engines.

License is CC-0

(HackerNews thread)

|improve this answer
3

This is not really possible.

You may be able to do something like "train a neural network to classify a person's gender similar to the way that most people within a culture would classify a person's gender based upon their voice."

It may be useful to know what gender a person is likely to be perceived to be, but this "social perception" is not the same thing as a person's gender.

If something like "social perception" is acceptable, you could use the LibriVox project and use something like https://github.com/ropensci/gender to classify the reader names. Here's an example of the types of audio and readers available from LibriVox. https://librivox.org/tea-cup-reading-and-fortune-telling-by-tea-leaves-by-a-highland-seer/

|improve this answer
  • If you something, probably should be if to you... Too small to suggest as edit – nmtoken Jan 18 '18 at 8:00
  • Thank you for your answer. Unfortunately, it does not solve my problem at all and I still believe, that such a dataset is possible to be found somewhere. Here is a link to a similar one kaggle.com/primaryobjects/voicegender/data. Instead of all those numeric values, there should be a wav file. But as you can notice, it contains male/female classification. – Szymon Wolarz Jan 18 '18 at 23:03
  • It's hard to tell, but it seems like the person who put together that kaggle dataset labeled the wav files themself. You could do the same and then you could train a model to classify a person's gender roughly the way that you would. This would still not be the same thing as a person's gender. – fgregg Jan 19 '18 at 3:37
  • Well, so could you explain, why this isn't really possible? We have a recording of someone speaking and the one who collected those recordings is likely to know, if it was a man or woman speaking. Voice frequencies differ in case of man or woman. So where is the problem? – Szymon Wolarz Jan 19 '18 at 14:34
  • 1
    Sorry, then, I understood gender in a different way (like, more casual way). I'm still looking for a dataset, similar to the one I mentioned above. Nevertheless, thank you again for your answers. – Szymon Wolarz Jan 19 '18 at 18:02

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.