Root adaptive homomorphic deconvolution schemes for speech recognition in noise

In this contribution we address the problem of speech recognition in noise and mismatch (there is mismatch when the conditions of training are different from those of testing). We extend our previously reported work in two directions. First, a comparison of normalisation functions demonstrates that the root-based cepstral representation can still be significantly enhanced by an appropriate reinforcement of the most energetic speech portions, leading to a 10-15% increase in performance on raw data Second, we extend this concept and propose two root-adaptive schemes. Recognition tests demonstrate the noise robustness of the proposed analysis that further improves the results by a significant amount