Adacel is a world leader in the development of accurate, noise tolerant speech recognition applications. Adacel’s proprietary intelligent post processing, grammar development toolkits and automated speech engine optimization, Adacel has earned an enviable reputation for delivery speech recognition systems for the most demanding applications such as the F35 Joint Strike Fighter.

In applications such as simulation or training environments, voice systems face less than ideal conditions. Humans tend to deviate in different ways from basic language, injecting their own personal and regional speech patterns. Exclamations and expletives often pepper normal speech. Disfluencies (e.g., “uh” and “um”) find their way into longer utterances, especially when the speaker is performing other tasks while conversing. Other complications include environments, in which normal speech is affected by emotion, stress, and other physiological effects. All these deviations from normal, doctrinal speech serve to complicate the process of automatic speech recognition (ASR).

Expert application of today’s speech engine technology, coupled with domain expertise and cross-application of communication and artificial intelligence disciplines, enables experts at companies such as Adacel to deliver speech recognition applications that can achieve consistently high recognition rates under the messy, variable, noisy, jerky, and very human real world conditions that people routinely communicate within. An example of this can be found in Adacel’s Joint Strike Fighter (JSF) solution, where noise and stress levels in the cockpit are incredibly high and pilots will be of many nationalities with various levels of competence in the English language. In another example application of Adacel’s speech recognition expertise – our Air Traffic Control (ATC) simulation system – user competence varies from the new entrant with no experience and a very limited knowledge of the ATC phraseology to a controller with 20 years plus experience. Again, all such classes of users must be accommodated with greater than 98% recognition results, day in, day out, at hundreds of sites around the world and all of this without the need for the users to carry out phonetic training of the system.

Synthetic Speech

Speech recognition is typically coupled with voice generation, or text-to-speech (TTS), which enables the computer-generated outcome to be spoken back to the user. Its most important element is quality, or recognizability, of output. Recent advances in phoneme-based generation have improved voice output to where the voices sound recognizably human. The more human the voice, the less likely the human participants will be distracted by a “computer-sounding” voice on the other end of the radio, and can focus more on their tasks at hand. Voice interfaces, both speech recognition and generation, can only provide so much capability to a system. The benefits of these technologies are only truly seen when they are used as the interface to knowledge-rich systems that can assign meaning to and react to the recognized speech, and that can generate speech that has valid contextual meaning to the listener. Much of the work in developing voice-enabled technology is in the underlying behavior system, and how well that system can deal with natural language.