双模态语音识别中乘积HMM权重系数与瞬时SNR的关系研究

在有噪声污染等复杂情况下,为了能够得到更高的语音识别率,提出了一种新的乘积隐马尔可夫模型(HMM)用于双模态语音识别,研究并确定了模型中权重系数与瞬时信噪比(SNR)之间的关系.该模型在独立训练音频和视频HMM的基础上,建立二雏训练模型,并使用重估策略保证更高的准确性.同时引入广义几率递减(GPD)算法,调整音视频特征的权重系数.实验结果表明,提出的方法在噪声环境下体现出了良好稳定的识别性能.

Research of relationship between weight coefficient of product HMM and instantaneous SNR in bimodal speech recognition

In order to better realize speech recognition in complicated noise environment, a new Hidden Markov Model (HMM) was proposed. The relationship between weight coefficient of product HMM and instantaneous Signal Noise Ratio (SNR) was researched and confirmed. In this proposed model, a two-dimension training model was built based on independently trained audio-HMM and visual-HMM, and re-estimation strategy was used to obtain higher recognition accuracy. Generalized Probabilistic Descent (GPD) algorithm was introduced to adjust weight coefficient. The experimental results show that, the proposed bimodal recognition approach with adjusting weight coefficient exhibits good performance on speech recognition in noisy environment.