ANSI SCTE 24-21-2017 pdf free download – BV16 Speech Codec Specification for Voice over IP Applications in Cable Telephony.
Due to the small frame size, the parameters of the short-term predictor (also called the “LPC predictor”) and the long-term predictor (also called the “pitch predictor”) are both transmitted and updated once a frame. The gain of the excitation signal is transmitted once every frame. The excitation VQ uses a vector dimension of 4 samples. Hence, there are 10 excitation vectors in a frame.
The BV16 encoder first passes the input signal through a fixed pole-zero high-pass pre-filter to remove possible DC bias or low frequency rumble. The resulting signal is then used to derive the LPC predictor coefficients. To keep the complexity low, BV16 uses a relatively low LPC predictor order of 8, and the LPC analysis window is 20 ms (160 samples) long. The LPC analysis window is asymmetric, with the peak of the window located at the center of the current frame, and the end of the window coinciding with the last sample of the current frame. Autocorrelation LPC analysis based on Levinson-Durbin recursion is used to derive the coefficients of the 8 th -order LPC predictor. The derived LPC predictor coefficients are converted to Line-Spectrum Pair (LSP) parameters, which are then quantized by an inter-frame predictive coding scheme. The inter-frame prediction of LSP parameters uses an 8 th -order moving-average (MA) predictor. The MA predictor coefficients are fixed.
The time span that this MA predictor covers is 8 × 5 ms = 40 ms. The inter-frame LSP prediction residual is quantized by a two-stage vector quantizer. The first stage employs an 8-dimensional vector quantizer with a 7-bit codebook. The second stage uses an 8-dimensional sign-shape VQ with 1 bit for sign and 6 bits for shape. For long-term prediction, a three-tap pitch predictor with an integer pitch period is used. To keep the complexity low, the pitch period and the pitch taps are both determined in an open-loop fashion. The three pitch predictor taps are jointly quantized using a 5-bit vector quantizer. The distortion measure used in the codebook search is the energy of the open-loop pitch prediction residual. The 32 codevectors in the pitch tap codebook have been “stabilized” to make sure that they will not give rise to an unstable pitch synthesis filter. The excitation gain is also determined in an open-loop fashion to keep the complexity low. The average power of the open-loop pitch prediction residual within the current frame is calculated and converted to the logarithmic domain.
The resulting log-gain is then quantized using inter- subframe MA predictive coding. The MA predictor order for the log-gain is 8, corresponding to a time span of 8 × 5 = 40 ms. Again, the log-gain MA predictor coefficients are fixed. The log-gain prediction residual is quantized by a 4-bit scalar quantizer. The 4-dimensional excitation VQ codebook has a simple sign-shape structure, with 1 bit for sign, and 4 bits for shape. In other words, only 16 four-dimensional codevectors are stored, but the mirror image of each codevector with respect to the origin is also a codevector. In the BV16 decoder, the decoded excitation vectors are scaled by the excitation gain. The scaled excitation signal passes through a long-term synthesis filter and a short-term synthesis filter. Figure 4 shows the block diagram of the BV16 decoder.ANSI SCTE 24-21 pdf download.