Learning Personalized Pronunciations for Contact Name Recognition

Automatic speech recognition that involves people’s names is difficult because names follow a long-tail distribution and they have no commonly accepted spelling or pronunciation. This poses significant challenges to contact dialing by voice. We propose using personalized pronunciation learning: people can use their own pronunciations for their contact names. We achieve this by implicitly learning from users’ corrections and within minutes making that pronunciation available for the next voice dialing. We show that personalized pronunciations significantly reduce word error for difficult contact names by 15% relatively.