A unified compensation approach for speech recognition in severely adverse environment

The performance of speech recognition systems is greatly degraded in adverse environments, such as in heavy noise, and in articulation variation adaptively made by the speaker. This variation gives rise to the Lombard effect characterized with changes in loudness, speed, emphasis, and clarity of speech. This paper presents a unified approach for speech recognition in adverse environments based on hidden Markov model (HMM) to compensate the additive noise and the Lombard effect. Unlike the existing methods which alleviate additive noise in recognition phase, we design a spectral addition algorithm in the training phase to compensate for the additive noise. Taking both the variance of Mel-scaled frequency cepstrum coefficients (MFCC) and the duration of different HMM states in different acoustic units into consideration, we compensate the Lombard effect by HMM state labeling. Experiments show great robustness in severe adverse environment without sacrificing the performance under normal environment. Since the compensation for noise and the Lombard effects is made in the training phase, it also greatly reduces the computational complexity in the recognition phased