Efficient MMSE-based channel error mitigation techniques. Application to distributed speech recognition over wireless channels

Combining Media-Specific FEC and Error Concealment for Robust Distributed Speech Recognition Over Loss-Prone Packet Channels

“…In this work we have extended the formulation previously proposed for wireless channels [24] in order to include any type of bursty distortion. While in wireless channels distortion is due to bit errors, when VQ replicas are used, distortion is caused by the strong quantization process.…”

“…In previous papers [23], [24], we showed that minimum mean square error (MMSE) estimation is a powerful technique to mitigate the errors introduced by a wireless channel. In [23], [24] it is assumed that all vectors are received at the back end, although affected by channel distortion.…”

“…In [23], [24] it is assumed that all vectors are received at the back end, although affected by channel distortion. In wireless channels, transmission errors can be observed at the receiver as bit errors.…”

“…In our previous work [23] we named this technique as forward-backward MMSE (FBMMSE) estimation in order to remark the use of past and future frames by introducing the forward and backward probabilities and a decoding delay. Further details of this technique, along with an analysis of the complexity of the HMM-based MMSE estimation, can be found in [24].…”

“…Example of the reliability evolution of MFCC (1) with time when VQ replicas (R) are used within a burst (starting at S and ending at E ). (24) therefore, (25) where represents vector quantization. Thus, finally, the estimate of the reliability function can be defined as when when (26) Since the cross-covariance decays relatively quickly, only the normalized cross-covariances for a few values must be precalculated and stored, while the remaining ones can be assumed to be zero.…”