Speech Compression

Abstract

Speech compression is a key technology underlying digital cellular communications, VoIP, voicemail, and voice response systems. We trace the evolution of speech coding based on the linear prediction model, highlight the key milestones in speech coding, and outline the structures of the most important speech coding standards. Current challenges, future research directions, fundamental limits on performance, and the critical open problem of speech coding for emergency first responders are all discussed.

Log Likelihood Spectral Distance, Entropy Rate Power, and Mutual Information with Applications to Speech Coding

“…AMR is also the codec (more specifically, its wideband version) that is planned for use in U.S. next generation emergency first responder communication systems [22]. For more information on these codecs, the speech coding techniques, and the cellular applications, see Gibson [23].…”

“…The quality of the spectral match in Figure 3 would appear to be unsatisfactory and so the statistically significant threshold is also somewhat validated; however, it would be useful if something more interpretive or carrying more of a physical implication could be concluded for this D value. Motivated by the variation in the values of the log likelihood ratio across frames, we calculate the percentage of frames that fall below the statistically significant threshold, in between the statistically significant and perceptually significant thresholds, and above the perceptually significant threshold for each G.726 bit rate for the sentence "We were away a year ago" and for the sentence "A lathe is a big tool", and list these values in Tables 1 and 2, respectively, along with the corresponding signal-to-noise ratios in dB and the PESQ-MOS values [23]. Motivated by the variation in the values of the log likelihood ratio across frames, we calculate the percentage of frames that fall below the statistically significant threshold, in between the statistically significant and perceptually significant thresholds, and above the perceptually significant threshold for Entropy 2017, 19, 496 9 of 14 each G.726 bit rate for the sentence "We were away a year ago" and for the sentence "A lathe is a big tool", and list these values in Tables 1 and 2, respectively, along with the corresponding signal-to-noise ratios in dB and the PESQ-MOS values [23].…”

“…Even though G.729 is not widely used at this point, we examine its performance because of its historical importance as the forerunner of today's best codecs, and also so that we can compare AMR-NB at 7.95 Kbps to G.729 at 8 Kbps, both of which fall into the category of code-excited linear prediction (CELP) approaches [23]. Tables 3 and 4 show the percentage of frames in the various D ranges along with the PESQ-MOS value for the two sentences "We were away a year ago" and "A lathe is a big tool".…”

“…Table 7 shows the bits/20 ms frame for each AMR-NB bit rate for each of the major codec parameter categories, predictor coefficients, pitch delay, fixed codebook, and codebook gains. Without discussing the AMR-NB codec, we point out in summary that, first, the AMR-NB codec does not quantize and code the predictor coefficients directly, but quantizes and codes a one-to-one transformation of these parameters called line spectrum pairs [15,23], but we use the label "predictor coefficients", since these are the model parameters discussed in this paper. Further, as shown in Figure 5, the synthesizers or decoders in these speech codecs (as well as G.729) are linear prediction filters with two excitations added together, the adaptive codebook, depending on the pitch delay, and the fixed codebook, which is intended to capture the elements of the residual error that are not predictable using the predictor coefficients and the long-term pitch predictor.…”

“…This same trend is not observed in Table 5. The reason for this appears to be that the sentence "We were away a year ago" is almost all what is called Voiced speech, which is well-modeled by linear prediction [23], and the bits allocated for the predictor coefficients remain almost constant over those rates. However, the sentence "A lathe is a big tool" has considerable Unvoiced speech content, which has more of a noise-like spectrum not captured well by a linear prediction model.…”