Hilman F. Pardede

On noise robust feature for speech recognition based on power function family

In this paper, a new feature robust against environmental noise is proposed for automatic speech recognition (ASR). This feature has similar extraction process with Power-Normalized Cepstral Coeffients (PNCC) except on two aspects. First, a generalization of the log function called the q-logarithmic function is applied to replace the power function and secondly, the mean normalization process is implemented before discrete cosine transform (DCT) instead of after it as in many traditional feature extraction algorithms. The proposed feature, called Q-Log Normalized Cepstral Coeffients (QLNCC), is shown more robust compared to two traditional features: MFCC and PLP. It is also better than PNCC without adding much complexity.

On the impact of normalizing power-based features on robustness against noise for speech recognition

Many power based features have been proposed in previous studies as alternative to the conventional feature, i.e. MFCC, for speech recognition. These features are of interest because they are empirically shown to be more robust than MFCC in noisy environments. Some studies argue that the compressions of power functions which are less sensitive than log for low energy spectra is one of the reasons while others agree that their performances are heavily affected by their normalization techniques. In this paper, the author focuses on investigating the effect of two normalization methods on power based features for the robustness of speech recognition in noisy environments. The paper specifically addresses features based on the q-log function, an example of power functions widely used in Tsallis statistics. The analysis suggests subtracting the mean may not be the best normalization technique due to its non-additive properties which is confirmed by the experimental results on Aurora-2. The results also confirm that the normalization may contribute more to the robustness of the features than the effect of the q-log compression.

Speech recognition features: Comparison studies on robustness against environmental distortions

The robustness against environmental distortions of various features used in speech recognition: MFCC, PLP, LPCC, FBANK, MELSPEC, ETSI - AFE, and PNCC are compared in this paper. These features are evaluated on Aurora-2, English spoken digit recognition task, a popular corpus often used to evaluate the robustness of speech recognition approaches. The results show that the use of different types of filter bank such as mel-scale filter bank in MFCC and Bark scale filter bank in PLP, achieves similar performance. The robustness of speech recognition features against environmental distortions are improved by using DCT even though the performances of features with and without DCT are comparable in clean conditions. PNCC, the current state-of-the-art feature generally shows a better performance compared to traditional features, except ETSI - AFE. Need to be noted that ETSI - AFE is found to be bias on Aurora-2 task.