A noisy Lombard and Loud speech compensation approach for speech recognition in extremely adverse environment

This paper proposes a unified approach for the noisy Lombard and Loud speech recognition based on training data compensation. A spectral addition to the training data is applied to the additive noise which is derived from the reversed point of spectral subtraction, while the compensation in Mel frequency cepstrum (MFC) domain for the Lombard and loud speech is based on HMM state labeling of the training data which take jointly the Mel frequency cepstrum coefficient (MFCC) variance and duration of different states in different acoustic units into account. The new approach is of great robustness in extremely noise and does not worsen the performance under normal environment and normal style. Meanwhile, since the compensation is made in the training phase, it does not increase the complexity of recognition.