论文部分内容阅读
为了更有效地提取英语句子重音,提出了一种基于听感知特征的方法。根据音素特点,改进段长的归一化方法;根据听感知特性,引入半音程和响度特征,并以归一化的音节最高值代替其平均值,系统正确率达到78.7%,漏检率为9.37%。在此基础上,还提出了基于掩蔽效应的突显度模型,系统正确率提高到83.4%,漏检率下降到5.72%。实验表明,突显度模型更符合人的听感知,其性能接近人工标注的一致率(约为86%)。系统还具有文本无关和说话人无关的优点。
In order to extract English sentence stress more effectively, a method based on auditory perception is proposed. According to the features of phonemes, the normalization method of segment length was improved. According to the auditory perception characteristics, the semi-range and loudness characteristics were introduced. The normalized syllable maximum value was replaced by the average value, the system accuracy rate was 78.7%, the missed detection rate was 9.37%. On this basis, the model of highlighting based on masking effect is also proposed. The accuracy of the system is improved to 83.4% and the missed detection rate is reduced to 5.72%. Experiments show that the prominence model is more in line with human auditory perception, and its performance is close to that of manual annotation (about 86%). The system also has the advantage of being text independent and speaker independent.