Concurrent Constraint Programming and Tree-Based Acoustic Modelling

The design of acoustic models is key to a reliable connection between acoustic waveform and linguistic message in terms of individual speech units. We present an original application of concurrent constraint programming in this important area of spoken language processing. The application presented here employs concurrent constraint programming – represented by Mozart/Oz [1] – to overcome the problem of sparse training data in order to provide context-dependent acoustic models for automatic speech recognition. State–of–the–art automatic speech recognition relies on standard pattern recognition algorithms, i.e. Hidden–Markov models (HMMs), to estimate representative yet robust acoustic models from a limited amount of speech data. In tree-based acoustic modelling, phonetic decision trees are employed to predict models for phonetic contexts which are not contained in the training data [2]. To this end, phoneme contexts are classified using phonetic categories of speech sounds. The assumption behind this is that context phonemes which belong to the same class have similar acoustic effects on the phoneme in question. This results in clustered phonetic contexts and a reduced number of acoustic models. Thus, tree-based acoustic modelling represents one approach to maintaining the balance between model complexity and available training data when building a HMM-based speech recogniser. We employ concurrent constraint programming in Mozart/Oz to implement this crucial component of statistical training for large vocabulary speech recognition. The principal idea which led us to apply this paradigm to speech recognition based on decision trees was to perform deterministic inferences among the trained models first, in order to avoid superfluous case distinctions. This method is known as propagate and distribute where propagation performs simple inferences and distribution performs case distinctions. The procedure presented here is comprised of four steps as described below. Characteristic acoustic parameters of a specific speech unit are trained in terms of context-dependent phonemes (allophones) which take into consideration the left and right phoneme context, respectively. Then, a phonetic specification system such as the International Phonetic Alphabet which divides the