Robustness Weighting Techniques For Noisy Speech Signal Processing

After more than a century of developments in speech signal processing,noise remains the most ubiquitous deterrent. Suffering from the noise problem,practical application of speech recognition techniques stagnated in the pastyears. To apply speech techniques in practical noisy environments brings usnew challenges.Basically, we can deal with the influence of noise in two ways: the first,noise-focused cancellation or compensation, the second, processing-focusedenhancement. When noise characteristics can be estimated or trained, thenoise-focused techniques, such as spectral subtraction and parallel modelcombination, are expected to be more effective. While not aiming to a specificnoise, the processing-focused techniques, such as Missing Data techniques inrobust speech recognition, are more compatible and thus can be more widelyemployed.In this paper, we study the robustness of summing systems, and propose arobustness weighting technique, which can be widely applied in systemscontaining summing expressions. The idea for robustness weighting is: whenthe elements in the summing system are corrupted by noise to differentextents, we can use a weighting approach to reduce the system's totalnoise-sensitivity, where the weighting coefficients can be determined by theelements' local robustness. To detail the theory, we study the total distortionin the worst case as well as in the statistic average case, and present theabstract propositions with mathematical proofs.The proposed technique is applied to three practical applications onspeech signal processing. First, we derive a robustness weighting algorithmfor harmonic signal reconstruction, which can obtain a relative 10% reductionin reconstruction distance in white noise. The weighted harmonicreconstruction algorithm is further employed as an enhancement front-end forspeech recognition system. Compared with the baseline, the proposed schemecan on average reduce the word error rate by 45.9%, or even 73.6% if wecalculate pitch with clean waves. In the second application, we derive a modelcompensation algorithm for robust speech recognition. Experimentation withAurora 2 tasks and comparison in word error rate, we gain 39.6% relativeimprovement with the proposed algorithm. In the multi-technique combinationexperiments, the algorithm brings a further improvement of 35.7% to singlespectral subtraction scheme and 10.9% to single Missing Data scheme. In thefinal application, we build a mandarin short phrase recognition system forportable devices, and derive a low-cost algorithm from robustness weightingtheory. The experiments in a practical environment, which is severemismatched with the training conditions, show that the proposed algorithmcan relatively reduce the word error rate by 15.2% to the 1-best recognitionsystem and 26.2% to the 3-best recognition system.