摘要
语音识别系统一般是将安静环境下训练得到的参数用于实际环境中, 当实际环境是安静的,语音识别系统的工作是令人满意的,然而,当实际环境中有噪声存在时,识别系统的性能就会下降.文中提出将自组织特征映射神经网络与半连续隐马尔可夫模型相结合,训练出适应噪声的隐马尔可夫模型的新方法.把该模型应用于小词汇量的孤立词语音识别系统.实验表明,该模型适合于对噪声背景下的语音进行识别.同传统的HMM模型相比,该模型具有更好的抗噪鲁棒性,在信噪比较低的情况下(2～12dB),识别率比传统HMM模型有明显提高.
Speech recognition systems work in practical environments using parameters that were trained in a quiet environment. The system performance is satisfactory when the environment is also quiet, but in a noisy environment, the performance degrades quickly. A hybrid model method was developed combining self-organizing feature mapping neural network (SOFMNN) and semi-continuous hidden markov model (SCHMM) to train noise by adapting HMM. The model trained by this method was used in an independent small size vocabularies recognition system. Experiments show this model is conformable to recognize speech in a noisy environment. Compared with the traditional HMM, this model has better noisy robustness. If the signal-to-noise ratio (SNR) is low (2-12 dB), the correct recognition rate increased distinctly.
出处
《哈尔滨工程大学学报》
EI
CAS
CSCD
北大核心
2005年第1期119-122,共4页
Journal of Harbin Engineering University
基金
哈尔滨市科学研究基金资助项目(2003AFQXJ053).
关键词
语音识别
半连续隐马尔可夫模型
自组织特征映射神经网络
噪声
Acoustic noise
Estimation
Markov processes
Mathematical models
Neural networks
Robustness (control systems)
Self organizing maps
Signal to noise ratio
