Harmonic Speech Coding - Digital Speech: Coding for Low Bit Rate Communication Systems - page 264

However the cross-correlation-based method has a higher complexity and
requires a high-precision pitch estimate.
The unvoiced amplitudes are calculated as the rms spectral energy over the
unvoiced spectral bandwidth, given by,

b k

S w (m)

1
κ

m = a k

A k uv = (8.13)

b k − a k

The harmonic amplitude estimation techniques described may be applied to
either the speech spectrum or the LPC residual spectrum.

8.4 Common Harmonic Coders

This section describes three examples of low bit-rate harmonic coders: sinu-
soidal transform coding (STC) [2], improvedmulti-band excitation (IMBE) [8],
and split-band linear predictive coding (SB-LPC) [4]. The STC and IMBE apply
sinusoidal analysis and synthesis techniques to the original speech signal and
SB-LPC uses the LPC residual signal. All three examples restrict the synthesis
of sinusoidal components to be harmonics of the fundamental frequency.

8.4.1 Sinusoidal TransformCoding
The sinusoidal transform coding (STC) operating at 4.8 kb/s divides the
speech spectrum into two voicing bands using the sinusoidal model approach
described in Section 8.3.1. The lower part of the spectrum, which is declared
as voiced, is synthesized as follows:

L v

A(lω 0 ) exp jlφ 0 (n) + jφ s lω 0 ˆ s v (n) = for − N / 2 ≤ n ≤ N / 2 (8.14)

l = 1

where

nω 0 + φ 0 φ 0 (n) = (8.15)

and ω k − 1

0

+ ω 0 N 2 φ 0 = φ k − 1 + (8.16)

0

1 is the frame length, ω 0 is the normalized fundamental frequency
of the k th frame, N is the duration between the analysis points, A(ω) is
the spectral envelope obtained by interpolating the selected peaks of the
magnitude spectrum, φ s (ω) is the phase spectrum derived from the spectral

where N +