Combined simulated data adaptation and piecewise linear transformation for robust speech recognition

This paper proposes a combination of simulated data adaptation and piecewise linear transformation (PLT) for robust continuous speech recognition. The original PLT selects an appropriate acoustic model using tree-structured HMMs and the acoustic model is adapted by the input speech in an unsupervised scheme. This adaptation can improve the acoustic model if the input speech is long enough and is correctly transcribed in the adaptation process. Indeed, an incorrect transcription can drastically degrade the acoustic model. Our proposed method increases the size of adaptation data by adding noise portions from the input speech to a set of pre-recorded clean speech, of which correct transcriptions are known. We investigate various configurations of the proposed method. Evaluations are performed with additive noisy continuous speech. The experimental results show that the proposed system reaches higher recognition rates than MLLR and PLT.