汉语耳语音库的建立与听觉实验研究

The Establishment of a Chinese Whisper Database and Perceptual Experiment

耳语音的识别和转换是个全新的课题,可应用于公共场合下的通讯和公安司法工作的某些特殊需要等方面.首先建立了一个包含1172个字和98个近音词的单人女声的汉语耳语音库.通过对两个听觉测试实验数据的统计分析,研究了人耳对汉语耳语音字和近音词声调的辨认率特征,得出人耳对孤立字四个声调的辨认率由高到低的排序为三声>四声>二声>一声.同时也得出人耳对词声调的辨认能力比字要强得多.幅值包络和音长这两个特征参量能够反映出汉语耳语音声调的特性,基于此参数对汉语耳语音字进行声调识别实验,其声调识别率已达到了人耳的平均辨认率,为连续耳语音声调识别研究打下了基础.

<Abstrcat>Whispering is a special way of speaking to communicate message lowly or privately. The whispered speech recognition and the reconstruction of normal speech from whisper are needed for some specific purposes, such as the private speech communication by mobile phone in public or the speech processing for police or of military use. However, few research has been conducted in these fields and many problems remain unsolved. In this paper, a Chinese whisper database in preparation for future work on whispered speech processing is introduced. The database consists of 1 172 characters and 98 closed-tone words from a female. Based on this database two auditory perceptual tests were conducted to investigate the tonal identification features of single characters and closed-tone words in Chinese whispered speech. From the experimental results the following conclusions can be drawn: (1) Tone 3 has the highest human perception accuracy, followed by tone 4 and tone 2, and tone 1 is the hardest to identify. (2) Women have better auditory perception than men as far as whispered speech is concerned. (3) The accuracy of identifying the tones is decided largely by the duration of sound. For tone 2 and tone 3, the longer the duration is, the more accurate the identification will be. On the contrary, for the fourth tone, the shorter the duration is, the more accurate the identification will be. (4) For human ears, it is easier to distinguish Chinese words of similar pronunciation with variation of the first character than to distinguish those with variation of the last character. (5) The human perception of words in whispered speech is more efficient than that of single characters. Because there is no fundamental frequency in the whispered speech, other features should be found to represent the tones. It is conjectured that the amplitude contour and duration can be used as the tone feature parameters in the Chinese whispered tone recognition. Another experiment is designed to validate this. The results show that the average tone recognition rates of characters by computer are at the same level as that of human perception. This forms the foundation for future tone study of continuous whispered speech.