Automatic generation of phone sets and lexical transcriptions

Automatic Speech Recognition (ASR) systems that have even moderately large recognition vocabularies model these words as sequencesof subword units, or phonemes. The set of these phonemes, or the phoneset, forms the basic units that the ASR system is trained to classify. This set is usually small in size, consisting typically of about 40 phones for English. The ASR system uses a dictionary in which all the words in the system’s vocabulary are transcribed in terms of these phones. The phoneset and the dictionary are specific to a language and are designed manually by an expert. The performance of the ASR system is critically dependent on the accuarcy of the dictionary. In this paper we attempt to design the phoneset and the dictionary automatically, using only the training data and their transcriptions. In order to do this we jointly optimize the dictionary as well as the acoustic models for an evolving phoneset using a Maximum a posteriori (MAP) formulation for the optimization of the dictionary and a Maximum Likelihood (ML) formulation to optimize the acoustic models. Experimental results on the Resource Management (RM) corpus show that such an automatically derived phoneset results in recognition accuracies close to that obtained using a manually designed phoneset and dictionary.