EVALUATION PARADIGM FOR SURNAME GENERATION PROBLEM

• 18,494 names and 25,648 manually transcribed pronunciations in the database using Worldbet symbols

  
  

• Divide the database into train and test sets (3 cuts of train-test pairs)

  
  

• Context length is used as feature to describe the sound

  
  

• Build different models using different context lengths

  
  

• Evaluate these models on the test data and compare the performance of these models against the reference model (manually transcribed pronunciations)