4.1.9 Harmonic noise weighting

06.203GPPHalf rate speech transcodingTS

If MODE = 1, 2 or 3, then C(z), the harmonic noise weighting transfer function, is activated. The excitation codebook vector and gains codebook vector are selected to minimize the spectrally and harmonically weighted error. The harmonic weighting filter, C(z), can be expressed as:

(104)

where

(105)

Cpitch, Gpitch and Lpitch were determined during the open loop lag search where the subscript m denoting the subframe has been dropped from Lpitch,m and hnw,m for notational convenience. Lpitch can take on fractional values so the interpolating filter employed for the open loop lag search is utilized to generate the fractionally delayed samples. Let x(n) represent the input of the harmonic noise weighting filter and y(n) represent the output. The filter can then be described by equation (106) and equation (107).

y(n) = x(n) – hnw x(n – Lpitch) (106)

(107)

where:

Figure 4 incorporates harmonic noise weighting (for MODE = 1, 2 or 3) and shows the VSELP excitation source. All error minimizations done after lag selection utilize the combination of spectral and harmonic noise weighting.

For MODE ¹ 0, P(n) is the input speech signal weighted by W(z)C(z) minus the input response of H(z)C(z). For MODE = 0, P(n) is the input speech signal weighted by W(z) minus the zero input response of H(z).

Figure 4: Long term predictor lag and code search

The zero state response of H(z)C(z) to the pitch predictor (adaptive codebook) vector, bL(n), needs to be computed prior to the search of the excitation codebook. This spectrally and harmonically weighted adaptive codebook vector is represented by b"L(n).