Uploaded image for project: 'Zapped: AcousticBrainz'
  1. Zapped: AcousticBrainz
  2. AB-449

Problems with "gender" model / measurement

XMLWordPrintable

    • Icon: Improvement Improvement
    • Resolution: Unresolved
    • Icon: Normal Normal
    • None
    • None
    • Model generation
    • None

      I wanted to start a conversation about a measurement that I think is problematic: gender.

      Here are some of the issues I see with it:

      • Potential confusion between sex and gender: although the definitions vary vastly between regions, countries, languages and cultures, a more modern, respectful and accurate language (as I understand it) would be to use "male" and "female" as units of sex, and "man" and "woman" as units of gender (even if the reality is that gender and sex are very much interconnected).
      • Strong binary enforced: only two answers are possible here. If we are talking about sex, this is ignoring a very significant part of the population who is intersex. If we are talking about gender, enforcing this binary goes against the reality of gender diversity, and risks being insulting to artists by misgendering them. Given the diversity of gender identities, classifying gender in AcousticBrainz is not feasible.

      I am not exactly sure what the way forward should be, but in my opinion, what is looked at here could be named better (probably "sex" rather than "gender"), come with a strong disclaimer, and use looser categories (for example "male-sounding" and "female-sounding"). However, this is far from ideal, in my opinion. Consider for example the diversity in sex and gender in the transgender community.

      This model is the only model that really qualifies the person rather than the song, which is why it has to be handled carefully and thoughtfully.

      We might also ask ourselves: depending on the accuracy of the modelling, how much value is there in the data, and does the value justify the problems that inevitably come with it? If it is not possible to design a model and use language that are respectful to all, should we even have it?

          [AB-449] Problems with "gender" model / measurement

          Stéphane Guillou added a comment -

          Thank you Alastair for the reply.

          I agree that an easy first step could be to re-label the data.

          Looking forward to reading more about what the projects decides to do in that space.

          Stéphane Guillou added a comment - Thank you Alastair for the reply. I agree that an easy first step could be to re-label the data. Looking forward to reading more about what the projects decides to do in that space.

          Alastair Porter added a comment -

          This is a great question and definitely something that we should look into more detail about.

          Our basic models use known examples of men and women singing. For the most part, the formants of speech follow general patterns and can be very roughly grouped depending on who is speaking. However, we know that singing can be much more expressive than speech, and has a chance of being misclassified (from what we might consider obvious cases like falsetto or screaming all the way up to more nuanced cases). It's true that much of the data that we present here is simply wrong.

          I like your idea of the labels male-sounding and female-sounding. This is an easy fix that we can do in the presentation of the data, and goes some way towards being more explicit about what we're trying to estimate.  You have a good point that this is the only model that characterises the person rather than the song. I think that we can be a lot clearer about what this specific model represents in order to be more inclusive.

          In fact, we probably need a general disclaimer like you propose on all of the data that we generate - some of it contains errors that we know about, some of it generalises incorrectly, and some of it requires knowledge of the the training process to be able to interpret the results.

          Alastair Porter added a comment - This is a great question and definitely something that we should look into more detail about. Our basic models use known examples of men and women singing. For the most part, the formants  of speech follow general patterns and can be very roughly grouped depending on who is speaking. However, we know that singing can be much more expressive than speech, and has a chance of being misclassified (from what we might consider obvious cases like falsetto or screaming all the way up to more nuanced cases). It's true that much of the data that we present here is simply wrong. I like your idea of the labels male-sounding and female-sounding. This is an easy fix that we can do in the presentation of the data, and goes some way towards being more explicit about what we're trying to estimate.  You have a good point that this is the only model that characterises the person rather than the song. I think that we can be a lot clearer about what this specific model represents in order to be more inclusive. In fact, we probably need a general disclaimer like you propose on all of the data that we generate - some of it contains errors that we know about, some of it generalises incorrectly, and some of it requires knowledge of the the training process to be able to interpret the results.

            Unassigned Unassigned
            chtfn Stéphane Guillou
            Votes:
            0 Vote for this issue
            Watchers:
            3 Start watching this issue

              Created:
              Updated:

                Version Package