Bell Labs designed the first speech-recognition system in the early 1950s. The technology was amazingly accurate at determining single-digit spoken numbers. By the early 1970s, natural-language speech understanding was demonstrated by Terry Winograd's SHRDLU system, a robot that understood commands such as "move the red block on top of the smaller green one." By the mid-1990s, speech-recognition technology was being applied to the Internet. In 1996, the California State University at Northridge was able to demonstrate an experimental voice-controlled Web interface. In the summer of 1999, the first commercial voice portal was launched. By 2000, many major technology developers were introducing voice-driven portals.

Until recently, a number of issues made it nearly impossible to develop a speech-recognition engine that would recognize fluent and natural speech. The basic challenges faced by voice application developers included the following:

1. Variability of Speech Patterns: Different people speak the same language differently and even speak the same word in many different ways. Interpreting speech variability has led to the development of complex pattern analysis. Understanding natural pauses, speaking rates, and changes in volume has been a complex and difficult task.
2. Processing Power: In the mid-1980s, a new technique known as Hidden Markov Models improved the ability to recognize word relationships. This computation-intensive technology eventually led to powerful speech-recognition applications. Achieving the real-time voice recognition found in voice portals requires processing power that was not commercially viable until recently.
3. Extracting Meaning: Very few speech-recognition applications are able to accurately determine the meaning of words. The quality of speech interpretation depends on the ability of the speech-recognition engine to properly choose the best match for spoken sounds from its list of expected text phrases. A more advanced process was required to extract meaning from those words. Because of the many possible ways that people speak and the many words that are used to communicate the same concept, a full understanding of human factors was critical to properly interpret meaning.
4. Background Noise: Because mobile phone users often access voice portals, background has been difficult for voice-recognition developers to filter out. The development of better microphones has helped, but issues such as wind, murmurs, and music have made it a challenge to properly isolate voice from noise.
5. Continuous Speech Recognition: Designing systems that are powerful enough to understand and respond to continuous speech requires a large amount of processing power that was not available at a reasonable cost in the past. When a person speaks at a natural rate, it has been difficult to distinguish which sounds were associated with specific words. For example, the phrase "to recognize speech" could have been misunderstood to be "to wreck a nice beach." Users do not naturally speak – with – pauses – between – their – words. As a result, processing phrases in real time as they are naturally spoken has been a major challenge.

Many of the solutions to problems associated with speech recognition are still being fine-tuned. Powerful computers have provided the processing power to overcome many of the limitations of the past. However, the most common speech-recognition systems of today are still very different from the way in which people naturally interact.