汉语语音合成中基于语境特征的清浊音时长调整

摘要

In Mandarin TTS, the duration of unvoiced and voiced phonemes in a syllable is a very important factor related to the naturalness of synthesized speech. We propose an unvoiced/voiced duration adjustment algorithm based on context features for HMM-based Mandarin TTS. In the algorithm, the relative duration of the unvoiced part in a syllable is clustered with context features. During the synthesis, a reference relative duration of the unvoiced part is generated from the decision tree, and the duration of the unvoiced part and voiced part in the synthesized speech is adjusted accordingly. Experiments show that this algorithm can improve the accuracyofdurationpredictioninHMM-basedMandarinTTS,and caneffectivelyimprovethenaturalnessofsynthesizedspeech.%　　汉语语音合成中音节内清音和浊音的时长是影响合成语音自然度的重要因素。在HMM汉语语音合成中，提出了一种基于语境特征的清浊音时长调整算法。在算法中，首先对清音相对音节的时长根据语境特征进行决策树聚类。合成时，从该决策树得到对应音节的清音相对时长参考值，合成语音的清音和浊音时长按照参考值进行调整。试验表明该算法可以提高HMM汉语语音合成的时长预测准确度，有效地提高合成语音的自然度。