Continuous speech recognition with adaptabilty to the speaking rate of an input speech

本研究开发了一个语音识别系统，该系统使用从输入语音的语速估计的音素持续时间信息。在本研究中，说话速率被假设为反映到平均元音长度。声学处理器使用修改的LVQ 2将输入语音变换成相似性矩阵。根据初步识别结果计算平均元音长度。根据输入语音中元音的平均长度估计每个单词模板中每个音素的持续时间。在考虑音素时长估计的基础上，使用DTW进行了口语词识别实验，对5名男性说话人的212个词汇（测试集）的词识别率为97.3%。音素持续时间信息是从另外5名男性和10名女性说话者发出的212个单词词汇（训练集）中收集的。其中，先用音素相关估计和后用音素相关估计的混合组合效果最好，并将上述方法推广到音素识别中。对5名男性使用者的212个词汇的音素识别率从71.8%提高到86.3%。

This tesearch developed a spoken word recognition system which used phoneme duration information estimated from the speaking rate of an input speech. In this research, the speaking rate is assumed to be reflected to the average vowel length. The acoustic processor transforms the input speech into a similarity matrix using the modified LVQ2. The average vowel length is computed from the preliminary recognition result. The duration of each phoneme in each word template is estimated from the average length of vowels in the input speech. By taking into account the estimated phoneme duration, the spoken word recognition experiments were carried out using the DTW.The word recognition score was 97.3% for the 212 word vocabulary uttered by 5 male speakers (test set). The phoneme duration information is collected from the 212 word vocabulary uttered by another 5 male and 10 female speakers (training set). The hybrid combination of the prceiding phoneme dependent estimation and the follwoing phoneme dependent estimation gave the best performance.The above-mentioned method was extended to phoneme recognition. The phoneme accuracy increased from 71.8% to 86.3% for phonemes in the 212 word vocabulary uttered by 5 male speakers (test set).