Studying the Effect of Simultaneous Modeling of Spectral and Temporal Envelopes on Zero-Shot Children's ASR

The work presented in this paper aims at enhancing the recognition performance of zero-shot children's speech recognition task through simultaneous modeling of spectral and temporal envelopes. We have used the Gamma-tone frequency cepstral coefficients (GFCC) for capturing the spectral envelope. The temporal envelope, on the other hand, is captured using the frequency domain linear prediction (FDLP) features. In the case of FDLP, application of linear prediction on the frequency domain representation of speech signal helps in effectively capturing the temporal envelope due the dual nature of time and frequency. In this study, we have not only studied the relative effectiveness of GFCC and FDLP features on zero-shot children's speech recognition task but also performed frame-level concatenation of GFCC and FDLP features in order to simultaneously model the spectral as well as temporal envelopes. Since the two features are highly uncorrelated, frame-level concatenation leads to capturing a wider range of acoustic attributes. This, in turn, reduces the error rates by a large margin. In our experimental explorations, frame-level concatenation led to a relative reduction of 26.5% in word error rate over the baseline system trained using GFCC features.