Robust automatic speech recognition with missing and unreliable acoustic data
Speech Communication
Hi-index | 0.00 |
An adaptation of independent component analysis (ICA) and missing feature theory (MFT)-based ASR for two simultaneous continuous speech recognition is described. We have reported on the utility of a system with isolated word recognition, but the performance of the MFT-based ASR is affected by the configuration, such as an acoustic model. The system needs to be evaluated under a more general condition. It first separates the sound sources using ICA. Then, spectral distortion in the separated sounds is estimated to generate missing feature masks (MFMs). Finally, the separated sounds are recognized by MFT-based ASR. We estimate spectral distortion in the temporal-frequency domain in terms of feature vectors, and we generate MFMs. We tested an isolated word and the continuous speech recognition with a cepstral and spectral feature. The resulting system outperformed the baseline robot audition system by 13 and 6 points respectively on the spectral features.