Speech recognition using quantized LSP parameters and their transformations in digital communication

A comparison of front-ends for bitstream-based ASR over IP

“…Recent papers ( [3][4][5]7,10], among others) have established that more robust parameterizations can be obtained by transforming some of the parameters sent by the coder, instead of decoding the speech signal and using a conventional ASR front-end. This means that selecting just the necessary information from the bitstream is better than extracting it from the decoded waveform.…”

“…There are a number of reasons that motivate this choice: first, they are highly predictable (they give smooth frame to frame transitions); second, their interpretation as frequencies eases the integration of auditory-related concepts; finally, they offer the possibility of performing a straight-forward stability check. However, the use of LSP as feature vectors has proved to be unsuitable for current ASR systems [4]. Therefore, they must be transformed into MFCC-type (Mel Frequency Cepstral Coefficients) parameters, which nowadays are still the most successful parameters for ASR.…”

“…On the one hand, a proposal by Kim et al [9,4] called pseudocepstrum provides a straight and computationally efficient approximation to the LPCC parameterization. On the other, the LPCC parameterization can be computed in an exact and computationally more demanding way.…”

“…In particular, we have tested both parameterizations in several simulated VoIP scenarios using two codecs (G.723.1 and G.729) and a wide range of Packet Loss Rates (PLRs) and Mean Burst Lengths (MBLs). This realistic testing environment adds to the analysis of the LSP quantization effects of [4] an evaluation of the influences of the whole codec process (for example on the energy parameter extraction or the frame rate provided) and the network distortions. This allows us to discuss when the approximation given by pseudocepstrum is advantageous and when, on the contrary, the exact LSP to MFCC conversion is preferable.…”

“…In this framework, bitstream-based approaches that obtain the ASR feature vectors directly from the coded bitstream, avoiding the speech decoding process, have been proposed ( [4,7,10,15], among others) to improve the robustness of ASR systems. LSP (Line Spectral Pairs) are the preferred set of parameters for the description of the speech spectral envelope in most of the modern speech coders.…”