Quality of Coding Schemes in C#

Scanner PDF417 in C# Quality of Coding Schemes

G7231

G728

G711

Standard G711 G726 G728 G729 G729a G7231 G7231 Codex PCM ADPCM LD-CELP CS-ACELP CS-ACELP MPMLQ ACELP Bit Rate 64 32 16 8 8 63 53 MIPS 34 13 33 20 105 16 16 Comp Delay (ms) 075 1 3 5 10 10 30 30

2

summarizes some of the key parameters that have been discussed in this section Linear Prediction Analysis-by-Synthesis (LPAS) Coding

Basic Mechanisms The ITU-T Recommendations G7231, G728, and G729 belong to a class of linear prediction analysis-by-synthesis (LPAS) vocoders CELP vocoders are the most common realization of the LPAS technique Figure 2-17 shows a block-diagram view of LPAS vocoders Some general principles discussed in Section 212 are applied herewith to specific ITU standards The decoded speech is produced by filtering the signal produced by the excitation generator through both a long-term (LT) predictor synthesis filter and a short-term (ST) predictor synthesis filter The excitation signal is found by minimizing the mean-squared error signal (the difference

between the original and decoded signal) over a block of samples100 It is weighted by filtering the signal through an appropriate filter Both ST and LT predictors are adapted over time Because the encoder analysis procedure includes the decoder synthesis procedure, the description of the encoder also defines the decoder The ST synthesis filter models the short-term correlations in the speech signal This is an all-pole filter with an order between 8 and 16 The predictor coefficients of the ST predictor are adapted in time, with rates varying from 30 to as high as 400 times per second The LT predictor filter models the long-term correlations in the speech signal Its parameters are a delay and a gain coefficient For periodic signals, the delay corresponds to the pitch period (or possibly an integral number of pitch periods); for nonperiodic signals, the delay is random Typically, the LT predictor coefficients are adapted at rates varying from 100 to 200 times per second99 A frequently used alternative for the pitch filter is the adaptive code book (this was briefly described in Section 212) Here, the LT synthesis filter is replaced by a code book that contains the previous excitation at different delays These vectors are searched, and the one that provides the best match is selected To simplify the determination of the excitation for delays smaller than the length of the excitation frames, an optimal scaling factor can be determined for the selected vector To achieve a low bit rate, the average number of bits per sample for each frame of excitation samples must be kept small The multipulse excitation vocoder represents the excitation as a sequence of pulses located at nonuniformly spaced intervals The excitation analysis procedure determines both amplitudes and positions of the pulses Finding these parameters all at once is a difficult problem and simpler procedures, such as determining locations and amplitudes one pulse at a time, are typically used The number of pulses required for acceptable speech quality varies from four to six pulses every 5 ms For each pulse, both amplitude and location have to be transmitted, requiring about 7 or 8 bits per pulse99 CELP vocoders approach the issue of reducing the number of bits per sample as follows Both the encoder and decoder store the same collection of C possible sequences of length L in a code book and the excitation for each frame is described by the index to an appropriate vector in the code book This index is typically found by searching the code book vectors and identifying the one that produces the smallest error between the original and decoded signals To simplify the search procedure, many implementations use a gain-shape code book where the gain is searched and quantized separately The index requires (log2C)/L bits per sample, typically 02 to 2 bits per sample, and the gain requires 2 to 5 bits for each code book vector ACELP introduces further simplifications by populating the code book vectors with a multipulse structure By using only a few nonzero unit pulses in each code book vector, the search procedure can be sped up The