A Scalable System for Embedded Large Vocabulary Continuous Speech Recognition

This paper presents a system for large vocabulary continuous speech recognition in condition of constrained hardware resources. We investigate efficient pruning and caching strategy aiming to handle extensive acoustic and linguistic modeling. Software components are analyzed in terms of resource consuming. Then, we evaluate the system performance in extreme configuration where acoustic and linguistic models are dramatically pruned. Results show that the system design we proposed allows to use large HMM-based acoustic models and trigram language models while performing very fast decoding, under 0.6 real-time on a standard desktop computer while remaining the transcript relevance.