Experimental Determination of the Intelligibility of Two Different Speech Synthesizers in Noise

摘要

The main goal of this thesis is to provide some criteria for classifying the effectiveness of the two major inexpensive synthetic speech technologies in a realistic communications environment. Synthetic speech has become commonplace within society. One cause of this proliferation is the availability of varied inexpensive synthetic speech systems to meet almost any application from those in industry to those in communicative disorders. The ability to choose the most effective communication system is an increasingly important consideration as the role of synthetic speech in society grows. This study examined two predominant inexpensive methods of synthesizing speech: formant synthesis(FS) and linear prediction coding(LPC). A pilot study indicated that upon first presentation of noncontextual material that FS was significantly more understandable than LPC. The results fall into two categories: the training data and the subsequent test data. Whereas the training data indicate the eventual equality of mean percent correct word scores for the two synthesizers without noise, the test data indicate the superior performance of LPC with interfacing noise. The effect of the interfering noise is studied as a cause and the significance of this study to current research in synthetic speech is discussed.