Response of Delta Modulation to Gaussian Signals

On Using the Itakura-Saito Measures for Speech Coder Performance Evaluation

“…This can be seen from the application of the likelihood ratio measure in vector quantization for voice coding.?" In fact, the code words designed for vector quantization using (14) are substantially consistent with the vowel triangle of Peterson and Barney from an acoustic-phonetic point of view," It also has been shown that the log likelihood ratio measure usually leads to a better recognition rate in speech recognition schemes. lO ,21> (Note that the log likelihood ratio and the likelihood ratio measure make no significant difference in most speech recognition applications.…”

“…1. Since~n(Z) is the optimal Mth-order inverse filter of Xn(z), En(z) is then the X~(Z) under this condition is thus (21) which is determined by the two all-pole filters, and has the same expression as the likelihood ratio measure of (14). This result gives us a convenient means of modifying a signal in order to achieve a prescribed distortion level from the original signal.…”

An interpretation of the log likelihood ratio as a measure of waveform coder performance

“…S shows a block diagram of the waveform coder distortion model which we have investigated. This model, also used by Aaron et al [10] for a delta modulator, is composed of a time-varying linear filter and an additive noise source. The rationale for this model is that the time-varying filter h (n) models the "linearly correlated" distortions in the coder (i.e., attenuation, delay, band limiting, reverberation) and the noise source e(n) accounts for the nonlinear distortions in 7a the coder (i.e., additive noise, tonal noise, clicks, etc.).…”

“…In terms of predicting subjective quality, it is known that an equal weighting of these two components is not the most desirable [9]. However, the functional form of the log likelihood ratio seems to be a good candidate for predicting subjec- (10) tive quality when only one of the components of distortion is significant. For example, in waveform coders in which there is no loss of bandwidth or attenuation of certain frequency bands, this measure can be useful in predicting subjective quality [91.…”