The first attempts to build a machine that can understand human speech were made in the late 1940s at the US Department of Defense, with the obvious goal of interpreting and translating intercepted Russian transmissions. These early experiments typically used top-down approach, trying to perform a literal word-for-word dictionary lookup. However, imagine how much time and computing resources had to be used to record and store a representation of each word in a specific language. Even then, the mapping from symbols to speech is not one-to-one since different underlying symbols can result in very similar speech sounds. As it turned out, human speech recognition operates at much lower, phoneme level. Phonemes are the smallest units of speech that distinguish one utterance from another. But the greatest problem lies in the fact that individual phonemes aren't particularly "well-behaved": individual speech sounds may vary depending on the sounds preceding and following the specific phoneme. In a modern speech recognition system, the digitalized stream of amplitudes of a speech signal captured by a sound board is first converted into the dominant frequency components. Each of these components is mapped to a specific phoneme, so the system can interpret words in a dictionary from the phoneme sequences that produce them. The key process showing the probability of one phoneme combination following another is based on a technique known as a Hidden Markov Model (HMM). The vast majority of commercial speech recognition algorithms are currently based on the HMM, with slight differences in probability calculations, endpoint detection schemes for continuous dictation, etc. The Hidden Markov Model Toolkit (HTK) from Cambridge University is a portable toolkit for building and manipulating Hidden Markov models. If you are interested in more "hands-on" approach, a HTK book provides an in-depth tutorial to building such systems. The final chapter of the tutorial describes the construction of a recognizer for simple voice dialing applications (14 complex steps described on 23 pages). You'll quickly see that creating a new speech recognizer from scratch is an extremely difficult and time consuming process - not to mention that the result is entirely language dependent. OK, so you have built a recognizer for the English language, but what about hundreds of other languages?

Making computers available everywhere, to everyone, is going to mean making speech recognition available for every language. The Microsoft Universal Phonetic Set (MUPS) project starts from the idea that we can collect the data in only one language, and perform a slight, language-specific, mathematical transformation to make it correct for the others. It will initially concentrate on getting and sharing data for English, French and Spanish. The goal in the long run is to share senones, the very small units that make up a single phoneme, building models of how human mouth, tongue, and lips form sounds.

When everyday applications of SR technology are concerned, entering text quickly and easily is one of its main strengths. Two broad categories, discrete and continuos speech recognition, are used in different scenarios. In discrete dictation, the system stores a predefined vocabulary of recognizable words, and the user must pause for a few milliseconds between words. This limitation disappears in continuos dictation products, resulting in a more natural interaction. Modern packages can now distinguish between a command to the PC and continuos dictation, making separate dictation and command modes obsolete. Limited processing power - typical for the embedded environments - is now one of the rare reasons to use a discrete speech recognition. Compared to the average speaking speed of 180 words per minute, input rates of up to 130 words per minute achieved by fully trained SR systems will clearly boost productivity in a text based environments. "Fully trained" means that you'll have to spend some time on a process called enrolling that includes reading a couple of basic sentences, sufficient for the software to create a basic user profile. Failing to do that will result in a frustratingly poor performance and limited productivity. And remember that even the best SR system won't work properly without a tiny piece of hardware. Background noise will greatly reduce accuracy rate so special noise cancellation headsets are recommended for all recognition packages.

More info on speech recognition technology, including a high-level explanation of basic concepts can be found in ZDNet Speech Recognition Special. But speech recognition technology isn't restricted only to PCs. An excellent introductory article from Speech Technology Magazine titled "Speech-enabled Appliances" describes SR and TTS technology in home and office appliances.

Quick list of Speech Recognition packages and information

Dragon Systems
Producers of one of the most accurate SR packages, Dragon NaturallySpeaking, recently bought by another industry giant, Lernout & Hauspie (L&H;). Windows-based SDK can be interfaced through ActiveX controls, COM interfaces, or SAPI methods.

IBM Voice Systems
"Big Blue" offers a (very) wide range of Home and Office software (ViaVoice for Windows and Mac), as well as enterprise voice solutions and technology for Web, telephony, Linux and mobile applications. Excellent developer options, including SDK for Windows, Linux, Java, and various other kits.

Interactive Systems
Interactive Systems Speech Toolkit uses a layered architectures, that allows application developers, researchers, and programmers to each access the speech engine at an appropriate level: SAPI, JSAPI or JPYTHON interfaces, access to the core engine in Java.

Linux SR applicationsThis document describes the basics of speech recognition on Linux platform and describes some of the available software.

<Microsoft
Probably the most popular speech developer's platform for Microsoft Windows. The newly released Microsoft Speech SDK version 5.0 includes the newly rearchitected Speech API (SAPI), updated development tools, samples, documentation, and enhanced versions of the Microsoft continuous speech recognition engine (MCSR) and Microsoft concatenative speech synthesis engine (TTS). A list of compliant third party tools includes dozens of interesting links.

Nuance
Nuance 7.0, the company's flagship speech recognition and natural language understanding server for use in telecommunications, enterprise, and Internet applications. Nuance Voyager is a voice browser that makes it possible for phone users to access people, applications, content and voice-enabled Web sites via a standard voice interface that parallels that of a Web browser. Nuance Voice Web Server, incorporating a VoiceXML interpreter, enables rapid creations and modification of voice sites for enterprise companies. Nuance also offers a wide range of integrated development tools.

Philips
Philips is another industry giant that is very active in the SR field. The list of their PC Speech Applications includes several applications for home and office users, as well as professional solutions with business-specific terminology and speech SDK.

SoftSound
SoftSound has developed a complete system for accessing the information in spoken documents such as news broadcasts or archives. The system works by applying a fast speaker-independent speech recognition system to spoken audio to obtain a rough transcript of what was spoken, and indexing this material by a specially adapted search engine.

SpeechWorks
Four complete speech recognition solutions for telephony applications: SpeechWorks 6, SpeechSite, Speechify and ETI-Eloquence.

Vocalis
Various end-user products (speechMail), Web based solutions (speecHTML), telephony services (SPEECHtel), and SpeechWare, a core SR technology underlying all products. It offers a combination of a proprietary low-level API as well as higher level functions. The low-level API provides full access to the core recognizer features, while the higher-level functions enable quick application development. The Developers Toolkit is Java based toolkit that allows developers of all experience levels to create and manage files required by the recognizer.