7.1. Introduction

Speech recognition deals mainly with the transcription of a voice signal into a sequence of words. Most systems rely on a statistic modeling of the speech generation process (see Chapter 6). Under this approach, the message results from a linguistic model which provides an estimation of Pr( W) for any sequence of words W = ( w ₁, w ₂, ) and an acoustic channel, which encodes the message W into the signal X, and which is modeled as a probability density function f (X W) . Speech recognition thus consists of maximizing the a posteriori probability of W, which is equivalent to maximizing the product pr( W) f( XW). In that way, speech recognition can be decomposed into three problems: the develoment of the linguistic model Pr( W), the development of the acoustic model f(XW) and the development of a decoder exploiting these two models.

This chapter mostly focuses on large vocabulary continuous speech recognition (LVCSR). The range of applications for this type of technology extends well beyond automatic text dictation. LVCSR can, for instance, be used for voice access to databases, content-based indexing of audiovisual documents, the transcription of conversations and spoken language translation. Progress made in speech recognition also benefits other technologies which rely on similar basics, such as speaker recognition and language identification.

The main modules involved in a...