OBSERVATIONS AND PLANS

• Other Baselines:

  
  

  • Enhanced feature vector: mfc + formants

    
    

  • Integration of HDM features

    
    

  • Comparison of parameter counts and complexity

    
    

  • Training data requirements

    
    

• Trends:

  
  

  • HDMs perform slightly better than chance

    
    

  • Single speaker results are encouraging

    
    

  • Some form of vocal tract length normalizaion, speaker adaption, or multiple targets per unit are needed