A Log-energy Scaling Normalization Scheme for Robust Speech Recognition

The log-energy parameter, as an auxiliary but influential feature, has been commonly used to augment Mel-frequency cepstral coefficients (MFCCs) to improve the recognition accuracy in automatic speech recognition (ASR). In this paper, a new and effective scaling approach named log-energy scaling normalization (LESN), which utilizes special nonlinear scaling functions on noisy speech data for log-energy normalization, is investigated. The scaling function is contrived conceptually from the relationship between clean and noisy data. In the experiments carried out with the Aurora-2 database, LESN is shown average word error rate (WER) improvements of 37.53%, 43.93% and 9.66% for Test Sets A, B, and C, respectively, when compared with the baseline processing. The results also show that LESN outperforms other similar approaches, such as log-energy dynamic range normalization (ERN). Furthermore, as well as being integrated with cepstral mean and variance normalization (CMVN), LESN is further applied to the LVCSR task and a considerable gain is achieved.