一种对加性噪声和信道函数联合补偿的模型估计方法

An algorithm of Model Compensation based on the estimation of additive noise and channel function for speech recognition

语音识别系统在面对实际环境中多变的加性噪声和信道差异的影响时性能急剧下降,抑制这些噪声和差异所造成的性能下降具有重要意义。作者提出了一种模型补偿算法,使用句子中的非语音段估计加性噪声,然后利用EM算法估计信道函数,从而在倒谱域上对失配的声学模型进行联合补偿。实验表明,相比基线系统,采用该算法的系统的平均性能相对提升幅度超过50%。算法可以动态跟踪环境的变化,性能表现优于一些传统的语音识别稳健性处理算法。

Speech recognition systems declined roughly in performance when they were facing the impacts of various additive noise and channel distortions in the actual environment, so it was of great significance for the speech recognition system to alleviate these impacts of the noise and distortions. An algorithm of model compensation was proposed, which computed the additive noise from the non-speech segments of the sentence, estimated the channel function using the EM algorithm, and jointly compensated the mismatched acoustics HMM models in the cepstral domain with them. Experiments employing this algorithm showed the significant improvement more than 50 percent relatively. The algorithm tracked the changes in the environment dynamically and it provided better performance than the traditional robust speech recognition algorithms.