Self-adaptation using eigenvoices for large-vocabulary continuous speech recognition