Harmonic Speech Coding - Digital Speech: Coding for Low Bit Rate Communication Systems - page 267

(a)

(a)

(b) (b)

(c)

(c)

100 150 200

0 0.5 0 0.5 100 150 200

samples samples

(b)

(a)

Figure 8.4 Harmonic speech synthesis: (a) original speech, (b) original harmonic
phases, and (c) IMBE

add method with the unvoiced part of the preceding frame. The synthesized
speech

ˆ

s(n) is then given by,

ˆ s (n) = ˆ s v (n) + ˆ s uv (n) (8.24)

An interesting feature of the IMBE coder is its simple phase model. The
fundamental phase is computed as the integral of the linearly-interpolated
pitch frequency, and the multiples of the fundamental phase are used as the
harmonic phases. The effect of this phase model is illustrated in Figure 8.4.
The coherent phase model used in IMBE concentrates the speech energy at
the phase locations corresponding to the multiples of 2 π of the fundamental
phase. For reference, the speech waveforms synthesized using the original
harmonic phases are also shown and they are very similar to the original
speech waveforms.

8.4.3 Split-BandLinearPredictiveCoding
The split-band linear predictive coding (SB-LPC) coder operating at 4 kb/s
employs time-domain LPC filtering and uses a multi-band type of excitation
signal. However the excitation signal of SB-LPC consists of only two bands,
separated by a frequency marker, below which the spectrum is declared
voiced and above which it is declared unvoiced. The estimation of the
frequency marker of SB-LPC is different from the technique used in STC.
The SB-LPC estimates a voicing decision for each harmonic band using
a similar multi-band approach described in Section 8.3.1. The estimated
voicing decisions are used to determine the voicing frequency marker, which
has eight possible equally-spaced locations in the spectrum, the first being