Customer Reviews

Statistical Methods for Speech Recognition (Language, Speech, and Communication)

5 star:		(6)
4 star:		(1)
3 star:		(0)
2 star:		(0)
1 star:		(0)

 

 

This book provides a comprehensive introduction to the statistical models and algorithms used for speech recognition. Jelinek sets up the speech recognition problem in the traditional way as the decoding half of Shannon's noisy channel model. While Jelinek glosses over signal processing, he provides an excellent overview of the symbolic stages of processing involved in speech recognition.

After a quick introduction, Jelinek digs into the statistics behind Hidden Markov Models (HMMs), the foundation of almost all of today's speech recognizers. This is followed by chapters devoted to acoustic modeling (probability of acoustics given words) and language modeling (probability of a given sequence of words), and the algorithmic search induced by this model. There are also advanced chapters on fast match (widely used heuristics for pruning search), the Expectation-Maximization (EM) algorithm for training, and the use of decision trees, maximum entropy and backoff for language models. He covers several auxiliary topics including information theory and perplexity, the spelling to phoneme mapping, and the use of triphones for cross-phoneme modeling. Each chapter is a worthy introduction to an important topic.

This book does not presuppose much in the way of mathematical, computational, or linguistic background. A simple intro to probability and some experience with search problems would be of help, but isn't necessary -- you'll learn a lot about these topics reading the book.

All in all, this is the best thorough introduction to speech recognition that you can find. Read it along with Manning and Schuetze's "Foundations of Statistical Natural Language Processing" from the same series; there's a little overlap in language modeling, but not much. You might want to start with the gentler book by Jurafsky and Martin, "Speech and Language Processing", before tackling either Jelinek or Manning and Schuetze.

This book is simply, as of 1999, the best of its kind, and I expect it will remain a core speech math text for a decade at least. It covers the construction, utilization and refinement of Markov speech models, but doesn't include any accoustic signal processing.

This book is possibly the first of its kind - exclusively devoted to Statistical Speech Recognition. The author is a pioneer in the area - one of the 'fathers' of the field,as it were. Thus one expects the text to be authoritative, and it is. The 'information density' is very high - it's a small book, but absolutely packed with information. You'll learn a lot about Hidden Markov Models and their use in Speech Recognition, but it also addresses many other issues, like language modelling and grammar, making it much more than a mere 'speech maths' book.

However, this is definitely not meant for absolute newcomers to the field of speech processing, and it does assume some background in advaced mathematics as well, especially in probability.

If you're looking for other aspects of Speech Recognition or code, you've come to the wrong place - but please don't spoil the rating of an excellent book by complaining that it doesn't have what it never promised to :-) - if you want a solid introduction to the field as a whole, i'd suggest 'Fundamentals of Speech Recognition' by Rabiner & Juang, and if it's code that you're looking for, there's lots of excellent open source stuff available on the net, notably from CMU and Cambridge, and there are some recent books in the market exclusively devoted to implementation of speech recognition systems.

To sum up, if you have some exposure to speech recognition and want to learn the maths & concepts behind the Statistical approach to Speech Recognition, this is your book.

This book provides important and interesting mathematic developpements for people who are experts in speech recognition. It's really complete and helpful but we are obliged to recognize that this is, most of time, a description of the IBM ASR system. Not as general as it could ...

I bought this book because I wanted a comprehensive introduction on the statistical approach to speech recognition. There is no doubt that this is an excellent book, that achieves this. If you are new to the field of speech recognition, be warned that this book isn't exactly the easiest to read, though.

For example, chapter 2 which discusses Hidden Markov Models, laying part of foundation for the following chapters, is full of mathematical formulas that won't be easy to follow unless you already have some background on the topic. I would recommend that instead you read L. Rabiner's paper "A Tutorial on Hidden Markov Models and Selected Applications in Speech Recognition". Rabiner not only shows the formulas, he describes their meaning, and the tutorial makes it easy for you to follow the text and actually understand what is going on.

That said, every chapter includes a section on additional reading (the above paper is mentioned in chapter 2) so you can always look up the references to help you understand the material, if you need to.

To summarize, this is an excellent text, that I would recommend to experts in the field, but beginners may need additional reading to get a better understanding of the book.

This book provides an excellent overview of speech recognition technology using Hidden Markov Models. Although Jelinek is clearly speaking with respect to his experience at IBM - he might as well be describing any other commercial speech recognition framework in the world. As a researcher and programmer in the area of speech recognition I regard this book as an excellent reference. It is concise, and I would say that anyone with a reasonable grasp of mathematics should have no trouble understanding most of the topics. In some of the more advanced areas some readers might need to refer to one of reference papers described in the book. I agree with the first reader - destined to be a classic!

This is an excellent book for people with speech recognition knowledge. The algorithms are very well described in a sound and comprehensive mathematical framework.