Use of speech presence uncertainty with MMSE spectral energy estimation for robust automatic speech recognition

In this paper, we investigate the use of the minimum mean square error (MMSE) spectral energy estimator for use in environment-robust automatic speech recognition (ASR). In the past, it has been common to use the MMSE log-spectral amplitude estimator for this task. However, this estimator was originally derived under subjective human listening criteria. Therefore its complex suppression rule may not be optimal for use in ASR. On the other hand, it can be shown that the MMSE spectral energy estimator is closely related to the MMSE Mel-frequency cepstral coefficient (MFCC) estimator. Despite this, the spectral energy estimator has tended to suffer from the problem of excessive residual noise. We examine the cause of this residual noise and show that the introduction of a heuristic based speech presence uncertainty (SPU) can significantly improve its performance as a front-end ASR enhancement regime. The proposed spectral energy SPU estimator is evaluated on the Aurora2, RM and OLLO2 speech recognition tasks and can be shown to significantly improve additive noise robustness over the more common spectral amplitude and log-spectral amplitude estimators.

Figure optionsDownload as PowerPoint slideResearch highlights
► The spectral energy estimator is investigated for the purpose robust automatic speech recognition.
► A speech presence uncertainty modification is proposed for the spectral energy estimator.
► When combined with speech presence uncertainty, the spectral energy estimator is shown to out-perform the log-spectral amplitude estimator for recognition robustness.