Flexible Decision Trees for Grapheme Based Speech Recognition

Authors: Borislava Mimer, Sebastian Stüker, Tanja Schultz

Abstract:

Over the last decades research in the field of automatic speech recognition (ASR) has seen enormous progress. Speech recognition systems are now deployed in real world applications, such as commercial software systems on PCs or Workstations, embedded in consumer devices such as cell phones or car navigation systems, or as part of specialized appliances. With this increasing economic relevance it has become more and more important to be able to rapidly extend speech recognition systems by new words, or to port them to new, previously unseen languages or domains. Hereby the speed and cost of development are of great importance. One of the most labor and cost intensive components of a speech recognition system is the pronunciation dictionary. Its creation often requires the application of linguistic knowledge. Even though automatic procedures for the creation of phoneme based dictionaries exist, they often require manual postprocessing by experts. This manual postprocessing is expensive and time intensiv. Therefore, the use of grapheme based speech recognizers has seen increased research lately [1, 2, 3]. Hereby words are segmented into graphemes instead of phonemes. The use of graphemes as modeling units has the advantage over the use of phonemes that it makes the creation of the pronunciation dictionary a trivial task, saving time and money. While a phoneme sequence is designed to describe the pronunciation of a word, the relation between the grapheme sequence of a word and its pronunciation is highly dependent on the writing system of the language in question and can be rather loosely coupled. Therefore, the context depending modeling of the units und the sharing of parameters are of central importance. Also, one does not obtain pronunciation variants when using grapheme based pronunciation dictionaries. However, recent experiments have shown that graphemes are equally well suited as phonemes for specific languages [2, 3]. In this article we investigate the potential of a flexible decision tree clustering scheme for context dependent modeling as proposed by Hua et al. for grapheme based speech recognition. To do so we trained grapheme based speech recognizers in two languages — English and German — and compared the word error rates when using the regular clustering procedure to when using the flexible clustering. Through the use of the enhanced clustering procedure we were able to reduce the word error rate of the grapheme beased recognizer by up to 9.3% relative, showing that for German and English graphemes are suited as units for implicit pronunciation modeling.