This app takes in voice audio data to classify whether someone is male or female. The dataset used for training can be found here. The features used all depend on the frequency distribution of the audio sample.
To run an example with all plots, run voice-gender.R and uncomment the plots you would like to view. They are the same ones that appear in the full Shiny app.
The full shiny app can be run from voiceGenderPage.R. Open it in Chrome on localhost:3038 to view the app running locally.
Voice data features were read in from a csv file and the dataset was split into training/testing (80%/20%). Labels were converted by using "male" as 1. An SVM classifier was trained and verified with an AUC generally above 0.9.
To plot the results of classifying with the SVM, PCA was used to reduce down to two dimensions. Color was added according to the predicted label.
In the Shiny app, voice audio input was transformed into a frequency distribution with weights at each frequency. The values were converted to be between 0 and 1 to maintain consistency with the training set by subtracting the min and dividing by the range of values.
After frequency transforming and scaling, the following features are calculated:
- meanFreq: Mean frequency
- sd: Standard deviation of frequency
- median: Median frequency
- Q25: Frequency at 25th percentile
- Q75: Frequency at 75th percentile
- IQR: Interquartile range
These features, which were also used to classify the training data, were then sent to the svm, which predicted the gender based on a empirically determined threshold.
The following R packages are required to run.
e1071kernlabggplot2ggfortifyROCRshinyshinyearr- Found at https://github.com/nstrayer/shinyearr and installed using
devtools::install_github("nstrayer/shinyearr").
- Found at https://github.com/nstrayer/shinyearr and installed using
tidyverseforcatsHmisc