Understanding the Internet Low Bit Rate Coder

Jan Linden Vice President of Engineering

Global IP Sound

Presented by

Jan Skoglund

Sr. Research Scientist

Global IP Sound

iLBC – Background info

• Development started in Summer 2000• Contributed to IETF as an internet draft in Feb 2002• Accepted as work item in IETF AVT group Mar 2002• Contributed to CableLabs RFP in June 2002• Improved version to IETF, Fall 2002• ECR submitted in May 2003• Support for 20 ms frames spring 2003• Successful interoperability events• Past Working Group Last call in IETF Jan 2004• April 2004 added as a mandatory codec in PacketCable 1.1 • December 2004 IETF process finalized (became

Experimental RFC 3951 and 3952)

Design Principles

• Free of 3rd party IPRo extensive experience in speech coding patents by design teamo patent and research situation monitored since 2000o has been public in IETF since March 2002 and reviewed by independent

speech coding researchers

• Packet independencyo no coding interdependency between frameso increased packet loss robustnesso suitable for IP networks

• Linear Predictive Codingo well know highly successful coding modelo novel coding techniques of residual signal

iLBC Features

• Sampling Rate: 8 kHz • Supports 30 ms and 20 ms speech frame modes• Bitrate

o 13.3 kbps (399 bits, packetized in 50 bytes) for 30 ms frameso 15.2 kbps (303 bits, packetized in 38 bytes) for 20 ms frames

• Computational complexity (TI C54x)o 30 ms frames: appr. 18 MIPS/channelo 20 ms frames: appr. 15 MIPS/channel

• Memoryo 400 Words/channel state memory (RAM)o less than 4 kWords table memory (ROM)o Stack and program memory requirements similar to other low bit rate

codecs (e.g. G.729A)

The Core iLBC method

• Start state encoding• Gain-shape waveform matching forward in time• Gain-shape waveform matching backward in time• Pitch enhancement• Packet loss concealment

iLBC Encoding

Incoming speech

Packets to network

iLBC Decoding

Packets from network

Decoded speech

20 ms vs 30 ms sub-blocks

• 20 ms frame size mode - 4 sub-blocks with the total length of 160 samples

• 30 ms frame size mode - 6 sub-blocks with the total length of 240 samples

0 39 79 119 159 +---------------------------------------+ | 1 | 2 | 3 | 4 | +---------------------------------------+ 20 ms frame 0 39 79 119 159 199 239 +-----------------------------------------------------------+ | 1 | 2 | 3 | 4 | 5 | 6 | +-----------------------------------------------------------+ 30 ms frame

20 ms vs 30 ms mode – bit allocation

240 samples encoded to 399 bits = 13.3 kbit/s (50 oct)

Parameter Bits

LPC

Start state position

Start state scale

Start state samples

Shapes

Gains

40

4

6

174

115

60

Total 399

160 samples encoded to 303 bits = 15.2 kbit/s (38 oct)

Parameter Bits

LPC

Start state position

Start state scale

Start state samples

Shapes

Gains

20

3

6

171

67

36

Total 303

Advantage over CELP

original

iLBC

g729

g723

PLC Staterecovery

iLBC Performance vs G.729A & G.723.1old version from Winter 2002

Source: Dynastat

iLBC Performance

Equivalent or slightly lower performance than G.729E in clean.

Improved robustness to packet loss compared to G.729E.

iLBC showed better than G.728 in other testing.

Implementation

Floating Point Source

Fixed Point Source

DSP Source

• Significant signal processing skills necessary

• Quality / efficiency trade-off

• ~ 6 Months

• Optimization skills

• ~ 4 Months

iLBC Specifications

• Available in floating point , fixed point ANSI C, TIc54x, TIc55x, TIc64x,…

• Supports 20 and 30 ms speech frames• Algorithmic delay: Same as frame size• Sampling Rate: 8 kHz• Bit rate: 13.333 kpbs for 30ms and 15.2 kpbs for 20ms

ProductFrame size

Complexity (max)

Program Memory

Data Memory Static Data

Memory DynamicEncoder Decoder Fix

Per channel

GIPS iLBC TIc54x

20 ms11.5 MIPS

4.1 MIPS

17.6 2.4 1.4 2.3GIPS iLBC TIc54x

30 ms13.5 MIPS

4.4 MIPS

GIPS iLBC TIc55x

20 ms 7.5 MIPS 3.0 MIPS

15.6 2.4 1.4 2.3GIPS iLBC TIc55x

30 ms 8.8 MIPS 3.1 MIPS

Memory in kWord16