论文部分内容阅读
针对语音识别中由于强噪声的影响而引起的Lombard和Loud效应进行研究,提出了基于训练数据的加性噪声和Lombard及Loud效应的联合补偿法。对于加性噪声是从谱减法的逆向角度对训练数据在频谱域采用谱加法;对于Lombard和Loud语音,则采用基于隐马尔可夫模型(HMM)状态标注的训练数据补偿,该方法同时考虑Lombard和Loud语音不同声学单元的不同状态在倒谱域的多种变化和多种变异情况下不同声学单元的音长及相对音长的变化。这种基于数据的多模式补偿使模型自动适应多种噪声和语音变异情况,在强噪声环境下具有很强的鲁棒性,并且不影响识别系统在正常环境或正常发音时的识别性能.同时,由于补偿是在训练过程中得到,不增加识别时的计算复杂度。
Aiming at the Lombard and Loud effects caused by the strong noise in speech recognition, a joint compensation method based on the additive noise of the training data and the Lombard and Loud effects is proposed. For additive noise, the training data is spectrum-added in spectral domain from the perspective of spectral subtraction. For Lombard and Loud speech, training data compensation based on Hidden Markov Model (HMM) state labeling is used. The proposed method also considers Lombard And Loud speech different acoustic units of different states in the cepstral domain of a variety of changes and a variety of variations under different acoustic unit length and relative pitch length changes. This data-based multi-mode compensation automatically adapts the model to a variety of noise and speech variations and is robust against strong noisy environments without affecting the recognition performance of the recognition system in normal or normal speech. Since the compensation is obtained during the training, the computational complexity of the recognition is not increased.