Advanced Speech Production Models

Current speech processing systems address the effects of channel noise andassume that all speakers are always in the same stress state. Neutral speechoccurs when the speaker has no task obligation other than to speak.Perceptually or physiologically induced speaker stress occurs in non-neutralconditions such as G-force, vibration, environmental noise, or emotion.Depending upon the type and degree of perceptual or physiological stress,the glottal source, vocal tract frequency structure, fundamental frequency,and intensity or duration are all affected in different ways. In addition,for a given speech utterance, a speaker's production can shift from one typeof stress state to another dynamically. There are two main approaches toaddressing the problem of speaker induced stress: (i) stress robust speechproduction models and (ii) stress robust speech processing algorithms. Thisstudy will focus on the first approach. By understanding the effects ofspeaker induced stress, it is possible to create models that are lesssensitive to stress induced variability. These new features will then beapplied as an example to the stressed speech recognition problem todetermine if they result in improved performance.Current speech processing algorithms work best with neutral speech producedin quiet environments. In many settings, speech can be produced inenvironments that cause perceptual speaker stress such as emotion, taskworkload, or perceived background noise. During perceptual stress, thespeaker will modify their production of speech to assist the listener'sability to receive the intent of the speech or to simply express emotion.Alternately, the human body can be exposed to physiological stress inducedby vibration, acceleration, deceleration, or changes in the makeup of theair supply. Both of these types of stress create a significant change on thespeech signal before it leaves the vocal tract. With better speechproduction models that seek to minimize the variability introduced by stressand noise, a wide range of speech processing applications will benefit. Thisis a foundational piece of technology that is required to take speechprocessing to the next stage in performance and reliability. Commercialapplications will benefit from more robust speech recognition, for example,because speakers will be able to express emotion naturally in the real worldenvironments that are part of their every day experience. This is an oftenquoted complaint with current speech recognition systems currently availablecommercially.