Unsupervised noise model estimation for model-based robust speech recognition
Martin Graciarena, Horacio Franco · 2004
Within the framework of a generalization of Rose's integrated parametric model (IPM) to the Gaussian mixture hidden Markov model (HMM) formulation to model speech in noisy environments, we further extend an algorithm for the maximum likelihood (ML) estimation of the noise HMM component. We propose: (1) a gain normalization algorithm in the log filterbank domain, which uses a gain estimate that is less degraded in broadband noise conditions; (2) the use of an augmented feature to incorporate dynamic information, which enables the use of the same probability density computation as the instantaneous feature; and (3) a technique for unsupervised noise model estimation using a phone loop grammar, which does not require an initial recognition pass. This algorithm does not require speech/nonspeech detection. In noisy digit recognition experiments, using HTK and NOISEX-92 databases, the noise estimation algorithm achieves, in supervised and unsupervised cases, performance similar to using noise models trained with the noise source data only.