VoIP Seminar

1
Speech Coding for VoIP

• Konsta Koppinen
• Signal Processing Laboratory
• konsta.koppinen_at_tut.fi

2
Why Speech Coding?

• The purpose of coding is to reduce the required
  bit rate
• For speech, compression ratio of 81 possible
  with little degradation in quality

3
Important Coder Attributes in VoIP

• Delay caused by coding (typically lt 50 ms)
• Bit rate (typically 4-64 kbps)
• Coder complexity (usually lt 30 DSP MIPS)
• Compatibility with existing codecs
• international standards of ITU-T, e.g. G.711,
  G.723, G.729,
• Nokia recommends using mobile standards

4
Choice of Coder Structure

• Different coder structures exist
• waveform coding
• vocoding
• hybrid coding
• Currently linear prediction-based analysis by
  synthesis (LPAS) coders offer best performance

5
Basic Idea of LPAS Coders

• We have a coder with various parameters
• Try synthesizing signals with different parameter
  values
• Pick the best one
• Hence the name analysis-by-synthesis
• Also known as closed-loop optimization

6
DPCM Example

• DPCM? (differential pulse code modulation)

Q
-
P(z)

P(z)
7
Analysis-by-Synthesis, Ver. 0.0
Input
Signal generator
Output
()2
?
8
Signal generator for speech?

• Excitation

• Vocal cords

• Vocal tract

noise
1/A(z)
Z-T
9
Analysis-by-Synthesis, Ver. 1.0
Input

• decoder

noise
1/A(z)
-
Z-T
Output
10
Comments on ver. 1.0

• All parameters (synthesis filter, pitch delay,
  gains) must be time-dependent
• e.g. every 5 ms
• All parameters must be quantized before sending
• Squared error in the output is minimized
• Some preprocessing useful
• normalize bandwidth
• speech enhancement

11
Perceptual Weighting

• Noise is not perceived uniformly
• Rule of thumb frequency bands with more energy
  tolerate more error (masking)
• This can be represented as a time-dependent
  frequency weighting (filtering)

original
SNR 15 dB
15 dB, weighted
12
Analysis-by-Synthesis, Ver. 2.0
Input frame
preprocessing
Z0
noise
1/A(z)
Z0
W(z)
Z-T
Output
13
Comments

• Synthesis and weighting filters (both IIR) should
  have memory, i.e. operate continuously
• improves performance
• Frames divided into subframes
• pitch filter has to be updated more frequently
• reduces complexity

14
Synthesis Filter

• Obtained from LP analysis of the input
• smooth windowing to reduce artifacts
• bandwidth expansion
• Update rate 10 ms
• Can be interpolated e.g. every 5 ms
• Quantized and interpolated using LSFs
• Has memory

15
Long-Term Predictor (LTP)

• Can be implemented using an adaptive codebook
  (AC)
• delay and gain into past excitation
• Fractional delay also possible by interpolating
  past excitation

16
Weighting Filter

• Also needs memory
• A good choice is

17
Fixed Codebook (FC)

• Represents residual error or system input
• Should be reasonably white (no correlation)
• Difficult to determine without a structured
  codebook
• stochastic codebook
• multi-pulse, regular pulse
• algebraic codebook

18
Choice of Parameters

• Ideally, all parameters should be chosen jointly
  to minimize the weighted error
• Practically, this is impossible (e.g. 2120
  possibilities)
• The parameters are chosen sequentially
• LPC ? AC lag and gain ? FC index and gain

19
Analysis-by-Synthesis, Ver. 3.0
preprocessing
LPC
-
output
-
1/A(z)
W(z)