The performance of Mel-frequency cepstrum based automatic speech recognition system significantly degrade in noisy environments. In this article, the feasibility of utilizing the bio-inspired auditory features to improve noise robustness is investigated. The features are based on auditory characteristics, which include gammatone filtering and modulation spectral processing to emulate the mechanisms performed in the cochlea and middle ear aimed to improve robustness in human ear. The robust noise resistant features that emulate cochlea frequency resolution are extracted by gammatone filtering. And then a long-term modulation spectral processing, which preserves speech intelligibility in the signal is performed. Compared and discussed are the features based on the performance on Aurora5 database, comprising the meeting recorder digit task recorded with four different microphones in a hands-free mode at a real meeting room and living room and office room simulated data corrupted with different levels of additive noises. The performance of these features is also investigated for CHiME challenge, aiming at speech separation and recognition in noise background that has been collected from a real family room using binaural microphones. The experimental results show that the proposed features provide considerable improvement with respect to the standard feature extraction techniques for both the versions of the database.
AUDITORY PROCESSING INSPIRED ROBUST FEATURE ENHANCEMENT FOR SPEECH RECOGNITION
Maganti, Hari Krishna;Matassoni, Marco
2012-01-01
Abstract
The performance of Mel-frequency cepstrum based automatic speech recognition system significantly degrade in noisy environments. In this article, the feasibility of utilizing the bio-inspired auditory features to improve noise robustness is investigated. The features are based on auditory characteristics, which include gammatone filtering and modulation spectral processing to emulate the mechanisms performed in the cochlea and middle ear aimed to improve robustness in human ear. The robust noise resistant features that emulate cochlea frequency resolution are extracted by gammatone filtering. And then a long-term modulation spectral processing, which preserves speech intelligibility in the signal is performed. Compared and discussed are the features based on the performance on Aurora5 database, comprising the meeting recorder digit task recorded with four different microphones in a hands-free mode at a real meeting room and living room and office room simulated data corrupted with different levels of additive noises. The performance of these features is also investigated for CHiME challenge, aiming at speech separation and recognition in noise background that has been collected from a real family room using binaural microphones. The experimental results show that the proposed features provide considerable improvement with respect to the standard feature extraction techniques for both the versions of the database.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
