Open-vocabulary spoken document retrieval based on new subword models and subword phonetic similarity