基于多空间概率分布的汉语连续语音声调识别研究

Research about Tone Recognition of Mandarin Continuous Speech Based on Multi-space Probability Distribution

汉语是一种带声调的语言,声调信息在汉语语音识别中具有非常重要的意义。提出了embedded声调模型与explicit声调模型相结合的方法用以识别汉语连续语音的声调。该方法能够将逐帧的基频信息和较强时长的基频信息相结合来识别声调。在"863-Test"和"TestCorpus 98"测试集上的实验表明,该方法分别能够达到96.12%和93.78%的声调识别正确率。

Chinese Mandarin is the tonal language.Tone is important to Mandarin speech recognition.We proposed a method to recognize the tone of Mandarin continuous speech,which is the combination of embedded tone model and explicit tone model.This method can fuse the fundamental frequency information of short time and long time.The experiments in ＂863-Test＂ and ＂TestCorpus98＂ test show that our proposed method can achieve 96.12% and 93.78% tone recognition correct rate separatively.