speech-class

Code used in the research of English dialect classification.

Directory of Files

Data Wrangling

These files have the goal of taking .wav audio files, generating force-aligned phoneme transcriptions, and outputting sound metrics into a .csv file. This data can then be analyzed using a series of Python classes and methods.

• note - before the process begins labeled audio files must be put into the montreal-input and montreal-output directories
  
• wav_conversion.py - a Python script with a single function convert_mp3_to_wav() used to convert a directory of .mp3 files into .wav files (only necessary if this is not already the case)
  
• textgrid_creation.py - a Python script with a single function create_textgrids() used to generate a series of textgrids in the specified directory that have the specified text as a single interval
  
• custom_mfa_align.py - a Python script with a single function align() that runs a series of Montreal Forced Aligner terminal operations to align the textgrids in a specified directory
  
• formant-extraction.py - a Praat script which extracts formant measurements and other sound metrics from each phoneme in the textgrids of the specified directory, outputting results into a spreadsheet named formants.csv
  

speech_modeling.py

A Python script with two classes.

• h_input
  • is initialized with the arguments raw_data which should be a pandas dataframe produced from the output of formant_extraction.py and id_vars which are non-value features of the output produced by formant_extraction.py
    
  • the instance variable self.input_df represents the processed dataframe
    
  • process method takes arguments location_specificity="country" and vowels_only=True
    
  • select_features method takes the argument selected_features=['F1'] and drops other features from self.input_df
    
  • normalize method takes the argument method='z'
    
  • if method=='z' then features will be normalized using standardization, if method=='minmax'then features will be normalized using min-max normalization
    
  • select_places method takes the argument places=["uk", "usa"] and drops samples that aren't labeled with these locations
    
  • revert method removes previous processing and brings the input dataframe back to its original state
    
  • output_input_df method takes the argument filename="input_df.csv"
    
  
  
• h_model
  • takes the dataframe from the processed self.input_df to be used in the implementation of different statistical models
    
  • is initialized with the arguments data which should be a copy of self.input_df, model_features which should be a list of features in this dataframe to be modeled, y_feature which should be the feature to be predicted (likely location), and y_main which should be the particular value of y_feature which will be separated from the rest (these models perform binary classification, predicting whether a sample is labeled with y_main or any other label)
    
  • the fit method takes arguments model_type="rforest", cv_method="LOO", test_size=0.30 if the "train-test" cross validation method is being used, and var_imp_type="mdi" which specifies the type of variable importance metric to be calculated

Additional Files and Directories

• gmu_scraping.py - this file scrapes from the GMU Archive all of the English language files and their demographic information as .wav files as seen in the specified directory
  
• LICENSE - basic MIT License
  
• README - this file
  
• gmu-modeling.ipynb - jupyter notebook used for modeling the GMU Archive data
  
• data - all of the audio tracks, aligned textgrids, and output .csv files from all of the English language samples in the two example datasets
  

Praat Scripts

These files correspond to Python files but with less functionality and running in Praat.

• wav-conversion.praat
  
• textgrid-creation.praat
  
• formant-extraction-praat
  

Notes

• Text alignment can be done without the custom_mfa_align.py file using the Montreal Forced Aligner directly in the terminal.
  
• The following lines of code download the models necessary for forced alignment. Other models may be used.
  mfa model download acoustic english_us_arpa
mfa model download dictionary english_us_arpa
mfa align --clean CORPUS_DIRECTORY DICTIONARY_PATH ACOUSTIC_MODEL_PATH OUTPUT_DIRECTORY implements the forced aligner
  mfa align --clean /home/user/montreal-input english_us_arpa english_us_arpa /home/user/montreal-output (example)
  
• adapted from this tutorial by Eleanor Chodroff