Developing HMM-based recognizers with ESMERALDA

ESMERALDA is an integrated environment for the development of speech recognition systems. It provides a powerful selection of methods for building statistical models together with an efficient incremental recognizer. In this paper the approaches adopted for estimating mixture densities, Hidden Markov Models, and n-gram language models are described as well as the algorithms applied during recognition. Evaluation results on a speaker independent spontaneous speech recognition task demonstrate the capabilities of ESMERALDA.