论文部分内容阅读
为在少量数据情况下显著提高方言普通话的识别率,针对标准普通话和方言普通话之间发音差异是连续变化的特点,在少量方言普通话的基础上,提出了基于距离度量的识别基元扩展方法,并将扩展基元与状态相关的基于基元的模型归并方法相结合。采用1 h的上海普通话数据作为开发集,用本方法,使音节错误率降低了17.3%。另外与自适应方法的结合使用,还可以将音节错误率再降低6.6%,这比单纯应用自适应方法错误率多降低了5.4%。
In order to significantly improve the recognition rate of dialect mandarin in the case of a small amount of data, the pronunciation difference between standard mandarin and dialect mandarin is a continuous change. Based on a small amount of dialect mandarin, a recognition unit expansion method based on distance measure is proposed, And combines the expansion primitive with the state-dependent primitive-based model aggregation method. Using 1 h of Shanghai mandarin data as a development set, this method reduced the syllable error rate by 17.3%. In addition, in combination with the adaptive method, the syllable error rate can be reduced by 6.6%, which is 5.4% lower than that of the simple adaptive method.