Recognizing Continuous Speech Remains an Elusive Goal

Article Abstract:

Continous speech recognition (CSR) requires sophisticated computer programs. Improved understanding in articulatory and acoustic phonetics has aided CSR development. Algorithms for isolated-word-recognition (IWR) have a wide performance range and cost. The matching of word templates in this system is speaker- dependent, and the system must be trained for each user. CSR research is even more limited than IWR research. While a limited CSR system was developed in the 1970s, only IBM is working on CSR in the form of a speech-recognition typewriter. The Japanese hope to build a speech-activated typewriter within ten years. Problems that must be solved include the difficulty of large vocabularies, speaker independence, graceful error recovery, and minimizing cost and maximizing speed. A block diagram of the IBM system and a table illustrating speech-recognition milestones is given.

Phonetics

---

Computer Vision: The Challenge of Imperfect Inputs

Article Abstract:

Computer vision is difficult because sensing device information is limited, factors are compounded in an image, and vision requires a large amount of memory and many computations. Most vision systems operate on three levels of processing: low-level feature extraction, intermediate level three-dimensional object deduction, and high- level scene description. The first step is to perceive the three- dimensional world from a two-dimensional projection. Progress is limited in object recognition. Model-based vision matches incoming images to simplified models of objects. This method is getting the most research, but is still far from perfect. Photographs illustrate shape determination and a graph illustrates computer vision advances along the three levels.

Pattern recognition (Computers), Computer vision, Pattern Recognition, Object Recognition, Machine Vision

---

Recognizing Confusable Words

Article Abstract:

The new Word Recognizer speech-recognition algorithm improves performance in discriminating between the nine letters with the 'e' sound to seventy-five percent. Words or letters are spoken into the computer, which digitizes the sound and converts it into a frequency spectrum. A first pass determines which class the letter or word falls into, and the second pass orders the letters in the class based on spectral features of a template in memory.

Form Overlay

Similar abstracts:

• Abstracts: Changing industry requires new vision: rapid change and future trends. Benchmarking and enhancing best practices in the engineering and construction sector
• Abstracts: Continuous production polyol plant. Coal handling plant. Polymer plant starts up smoothly
• Abstracts: British Petroleum and Amoco tie the knot. Wanted:masters of elegant frugality. BP and Ford fund carbon research
• Abstracts: Redeveloped algorithms wring more out of supercomputers. Looking Ahead in Space. The Challenge to U.S. Supercomputers

This website is not affiliated with document authors or copyright owners. This page is provided for informational purposes only. Unintentional errors are possible.