A path-based layered architecture using HMM for automatic speech recognition

Generally, speech recognition systems are based on one layer of acoustic HMM states where the recognition process consists on selecting a sequence of those states providing the best match with the speech utterance. In this paper we propose a new approach based on two layers. The first layer implements a standard acoustic modeling. The second layer models the path followed by the speech signal along the activated states of the acoustic models, defining a set of state-probability based HMMs. This method presents two main advantages in front of conventional recognizers: a consistent pruning of the possible paths preceding and following each state in the recognition process, and the possibility of modeling high-level information in the second layer in a somewhat independent fashion from the acoustic training. A testing database from a real voice recognition application has been used to study the performance of the system in a changeable environment.