论文部分内容阅读
为提高声调识别率,利用隐条件随机场对汉语声调进行建模,通过加入音节内动态特征、音节间动态特征以及段长特征来进一步提高声调识别性能。提出了将声调模型加入大词汇量连续语音识别系统的区分性方法,根据最小音子错误准则区分性训练模型相关的概率权重,对声学模型及声调模型概率进行加权;给出了两种权重组合策略并通过一种平滑方法来克服权重训练过拟合的问题。实验结果表明,基于隐条件随机场声调模型能够显著提高声调识别率以及大词汇量语音识别的识别率,同时与全局模型权重方法比较,区分性的模型权重训练能够在声调模型加入连续语音识别系统之后,进一步提高系统的识别性能。
In order to improve the recognition rate of tones, the Chinese tone is modeled by implicit conditional random field, and the performance of tone recognition is further improved by adding the dynamic features in syllables, the dynamic features between syllables and the features of segment length. A discriminative method of adding the tone model to a large vocabulary continuous speech recognition system is proposed. According to the probability weight of the discriminant training model, the probabilities of the acoustic model and the tone model are weighted according to the minimum phonetic error criterion. Two weight combinations Strategy and overcome the weight training over-fitting by a smooth method. Experimental results show that based on implicit conditional random field tone model, the recognition rate of tone recognition and large vocabulary speech recognition can be significantly improved. Compared with the global model weighting method, the discriminative model weight training can be applied to the continuous speech recognition system After that, to further improve the system’s recognition performance.