On the design of an automatic speaker independent digits recognition system for Romanian language

Abstract

For decades, engineers and scientists have studied the phenomenon and production of speech, with an eye on creating more effective and efficient systems for human-computer interaction. This paper presents a large number of experiments made to create an automatic speech recognition system (ASR) for spoken Romanian connected digits. State-of-the-art hidden Markov acoustic models (HMMs) and a finite state grammar language model are used, in order to build and optimize a fully-functional digit recognizer system in Romanian language. The applications of speech recognition in daily life are multiple, and truly there are no limits to the use cases of this technology: from niche applications like medical interfaces and industrial command and control systems to consumer applications, where modern phone operating systems offer speech interfaces to interact with the system. The rich mathematical framework of HMMs makes statistical approaches very feasible for this task, and one of the goals of this paper is to confirm the validity and reproducibility of this method. Another objective is the integration of the components and toolkits necessary to build a recognition system, briefly describe the processes involved in speech representation, the mathematics behind it and the analysis for improving and optimizing the primary evaluation metrics. The results show the advantage of training with a larger speaker database, in order to obtain an independent speech recognizer, with more than 60% WER improvements compared to a dependent model, for a 90 speaker database used for evaluation. The implementation of the system and the experiments, along with the evaluation results for decoding and optimization are provided.