Acoustic Modeling of Subword Units for Large Vocabulary Speaker Independent Speech Recognition

The field of large vocabulary, continuous speech recognition has advanced to the point where there are several systems capable of attaining between 90 and 95% word accuracy for speaker independent recognition of a 1000 word vocabulary, spoken fluently for a task with a perplexity (average word branching factor) of about 60. There are several factors which account for the high performance achieved by these systems, including the use of hidden Markov models (HMM) for acoustic modeling, the use of context dependent sub-word units, the representation of between-word phonemic variation, and the use of corrective training techniques to emphasize differences between acoustically similar words in the vocabulary. In this paper we describe one of the large vocabulary speech recognition systems which is being developed at AT&T Bell Laboratories, and discuss the methods used to provide high word recognition accuracy. In particular, we focus on the techniques used to obtain acoustic models of the sub-word units (both context independent and context dependent units), and discuss the resulting system performance as a function of the type of acoustic modeling used. INTRODUCTION In the past few years there have been proposed a number of systems for large vocabulary speech recognition which have achieved high word recognition accuracy [1-6]. Although a couple of the systems have concentrated on either isolated word input [6], or have been trained to individual speakers [5, 6], most current large vocabulary recognition systems have the goal of performing speech recognition on fluent input (continuous speech) by any talker (speaker independent systems). The approach to large vocabulary speech recognition we adopt in this study is a pattern recognition based approach. For a detailed description of the system we have developed, the reader is referred to [7]. The basic speech units in the system are modeled acoustically based on a lexical description of words in the vocabulary. No assumption is made, a priori, about the mapping between acoustic measurements and phonemes; such a mapping is entirely learned via a finite training set of utterances. The resulting speech units, which we call phone-like units (PLU's) are essentially acoustic descriptions of linguistically-based units as represented in the words occurring in the given training set. The focus of this paper is a discussion of various methods used to create a set of acoustic models for characterizing the PLU's used in large vocabulary recognition (LVR). The set of context independent (CO units we used in this study is a fixed set of 47 phone-like units (PLU's), in which each PLU is associated with a linguistically defined phoneme symbol. We model each CI PLU using a continuous density hidden Markov model (CDHMM) with a Gaussian mixture state observation density. Each word model is defined as the concatenation of the PLU models according to a fixed lexicon defined by the set of 47 associated phoneme symbols. We also consider a set of context dependent (CD) units which includes PLUs' defined by left, right and both left and right context. t On leave from CSELT, Torino, Italy.