该文提出了应用声学似然分作为置信度来生成可靠口音相关单元的方法。基于可靠口音相关单元构造声学模型,并通过声学模型重构的方法将它们融合到标准普通话模型中,以改善普通话语音识别器对带多方言口音语音的识别效果。另外,还提出了使用增量式决策树融合及根据支配度选择Gauss混合2种方法来减少冗余的Gauss混合,从而提高了重构后的声学模型的效率。实验表明:该方法在不降低对标准普通话的识别率的前提下,对粤、吴口音的绝对音节错误率分别下降了9.25%和9.21%。 This paper proposes a method to generate reliable accent-related units using acoustic likelihood scores as confidence. The acoustic models are constructed based on the reliable accent-related units and fused to the standard Mandarin model through acoustic model reconstruction to improve the recognition of Mandarin speech recognizers with multi-dialect accents. In addition, two methods are proposed to reduce the redundant Gauss mixture by using incremental decision tree fusion and Gauss mixture according to the degree of dominance, so as to improve the efficiency of the reconstructed acoustic model. Experimental results show that the proposed method reduces the absolute syllable error rates of Cantonese and Wukou by 9.25% and 9.21% respectively, without reducing the recognition rate of standard Mandarin.