Experiments on children's speech recognition under acoustically mismatched conditions

In this paper, we have explored the effectiveness of some of the existing acoustic features for the task of recognizing children's speech using acoustic models trained on adults' speech. Among the explored features, the Mel-frequency cepstral coefficients (MFCC) and the perceptual linear prediction cepstral coefficients (PLPCC) are the most commonly used ones in speech recognition. The third feature explored in this work is the one based on normalized first-order spectral moments (SMAC-features). The SMAC-features have not been explored for such mismatched ASR task as reported in the presented work. Due to large acoustic mismatch that exists between the training and test data, the recognition performance is observed to be highly degraded in all the explored cases. At the same time, the SMAC-features are noted to be superior to the other two. The same has been verified experimentally in this paper under clean as well as noisy test conditions. To address the acoustic mismatch, a low-rank projection is learned on the adults' training data using heteroscedastic linear discriminant analysis. The derived transform happens to emphasize the principal dimensions of acoustic variations in the adults' speech. The low-rank projection is then applied to both the training and test data. During testing, the transform maps children's test data to the space of the training data and thus alleviates the acoustic mismatch. The low-rank projection is found to result in significant improvements in the recognition performance when applied to the SMAC-features.