Advanced algorithms and architectures for speech understanding

From MaRDI portal





The articles in this volume will not be reviewed individually. The book presents an overview of the major results in the field of natural speech understanding inside the European Strategic Program for Research and Development in Information Technology, ESPRIT. It is concerned with Project 26 ``Advanced Algorithms and Architectures for Speech and Image Processing. The project objectives are to develop the algorithmic and architectural techniques for recognizing and understanding spoken or visual signals, and to demonstrate these techniques in suitable applications. This ambitious aim results in a thorough and valuable realization. The system goes from a continuously natural-spoken language sentence to its meaning: it is trainable, exploits a limited vocabulary, and has nearly real-time performance. The book provides a valuable source of information for all professionals in computer science and artificial intelligence. The chapters in the book are organized due to the basic philosophy of the project: to separate the front-end of the process as a recognition module and an understanding module. Both modules are completed with suitable hardware and parallel structures. The presentation contains four main chapters: recognition algorithms, real-time hardware implementation of the recognition stage, understanding algorithms and parallel architectures. Ideas for future developments are subject to the last chapter. Chapter 1 deals with some notes and gives an overview of the book. Chapter 2 is concerned with recognition algorithms. They are developed with the main objective of designing an efficient technique to extract a lattice of word hypotheses from a continuously spoken utterance. Attention is paid to the accuracy of the recognition process - it must be high enough to obtain a good understanding rate. A description of the system is supplied, followed by the lexicon structure, the representation of the words of the lexicon and the verification module. The final result of this first task is the high-performance speaker-dependent continuous- speech word hypothesizer for large vocabularies. Most of the algorithms and architectures are applicable to high-performance, large and very large isolated word recognition. The main results are: use of state-of- the art signal processing techniques for speech analysis; novel techniques for phonetic classification, phonetic matching, and lexical access to large vocabularies; sub-word speech units optimization; control strategies for interaction between lexical access and detailed Hidden Markov Model verification; vocabulary flexibility; automatic training; adaptability to different languages; capability of real-time performance with a multi-DSP architecture and parallel implementability of the algorithms; continuous speech word hypothesization. Chapter 3 is concerned with all the choices relevant to the hardware architecture of the recognition stage. It includes a detailed description of the system, hardware, firmware blocks, and details on other system functions. Different experiments concerning single- and two-step isolated word recognition, and single-step continuous speech recognition, are described. The understanding algorithms are subject to Chapter 4. The objectives of this task are to parse the word lattice in order to find the most likely word sequence, and to provide a formal representation of its meaning. The chapter covers the essentials about language modeling, parsing and control. The basic principles of the design of the parallel architecture are discussed in Chapter 5. That is in connection with the aim to achieve real-time performance in the understanding stage. The last chapter outlines the future development of the main points of the recognition algorithms, hardware and understanding algorithms. Experimentals are included in any chapter, which ends with an extensive bibliography.





Describes a project that uses

Uses Software






This page was built for publication: Advanced algorithms and architectures for speech understanding

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3997485)