A comparison of different methods for noise adaptation in a HMM-based speech recognition system
Résumé
Hidden Markov models (HMMs) have been successfully applied in speech recognition, but their performances dramatically drop in noisy conditions. This paper presents a comparison of different methods to increase the robustness of an HMM automatic speech recognition system. We have evaluated two types of approaches: the first one estimates a transformation from a few noisy sentences to adapt the initial models trained in clean speech. The second one tries to remove the noise from the signal without modifying the HMM models. We have compared the following methods: Parallel Model Combination(PMC) Maximum A Posteriori(MAP) Maximum Likelihood Linear Regression(MLLR) Multivariate-Gaussian-based Cepstral Normalization(RATZ) Vector Taylor Series(VTS) Spectral subtraction. Tests have been conducted on the noisy database of a voice command task: a multi-speaker navigation system using a limited vocabulary. On a subset of the database with a SNR of 10 dB, we have obtained the following results: baseline system:85% PMC:93% RATZ:91% MLLR:93% In the full paper, we will give the results for all the methods and all the noisy conditions and we will discuss the advantages and drawbacks of each method regarding the real time capabilities and the size of the adaptation set.