© 2006 AudioCodes Ltd. All rights reserved.

AudioCodes Confidential Proprietary

Signal Processing Technologies in Voice over IP

Eli Shoval Audiocodes

Scope

• The purpose of this presentation is to provide an overview of speech processing technologies that are used in Audiocodes VoIP products

Outline

• Signal processing technologies in VoIP

– Line Echo Cancellation

– Acoustic Echo Cancellation

– Speech Compression

– Wideband Speech Compression

– Background Noise Reduction

– Voice Quality Monitoring

Main challenges in VoIP system design

• Bandwidth Efficiency – Need a vocoder with the lowest possible bit rate with voice quality

suitable for the application– Different vocoders will suite different networks (LAN, WAN, Wireless)

• IP Network Inherent problems– Jitter – Delay – Packet Loss

• Voice Quality issues and enhancements– Echo– Background Noise– Gain adjustment

Main challenges in VoIP system designCont’d

• Handling of Non Speech Signals -– Fax– Data Modems– Caller ID– DTMF

• Interoperability – VoIP equipment must be able to communicate smoothly with equipment of other vendors

• Implementation efficiency

Basic DSP processing in VoIP

Voice Signal

I/F

In-Band Signaling Detector

Data Relay Rx

Speech Encoder

Speech Decoder

Host Port I/F

Echo Canceler

Host Data

PCM Highway

Fax Relay Rx

Fax Relay Tx

Data Relay Tx

In-Band Signaling Generator

Voice FaxData

Discrim-inator

CAS Signaling Generator

CAS Signaling Detector

Fax/Modem ByPass Tx

Fax/Modem ByPass Rx

Input Gain

Output Gain

CAS Signal

CAS Signal

Line/Electrical Echo phenomena

• Line echo exists in both networks due to leakage in 2/4 wire hybrid transformers

• In PSTN network: the echo exist but is not perceptible (it is masked)

• In IP network: echo is perceptible due to the added IP delay

TDM Networks vs. IP Networks

TDM Network

Telephone A Telephone B

E1 E2 E3 E4

Acoustical echo

Acoustical echo

Tx Tx

Rx Rx

Regular TDM network

IP network – the IP delay is added to the PSTN side delay

IP Network

Telephone A Telephone B

E1 E2 E3 E4GW A GW B

Acoustical echo

Acoustical echo

RxRx

Tx Tx

Basic structure of Echo Canceler

Echo Filter Estimation

h

Echo Filtering

h

+-

Ne

ar - E

nd

Fa

r -

En

d

Hybridh

DC Remover

Control Circuit

NLP

][nx

][nx ][nx

][ny][ˆ ny

][nv][ns][ns

][ne

][ne][no

EC

h

Example of sparse FIR with 3 active windows to handle 3 hybrids

R1 R2 R3

G.168 Test 2B - convergence

Challenges in Echo Cancellation

• Double talk can cause the adaptive filter to diverge – adaptation in AC49x/AC50x EC is robust to double talk

• Non linearity in the echo path can not be modeled by the linear FIR – AC49x/AC50x EC has a proprietary NLP to reduce the residual echo

• NLP attenuation can cause modulation of background noise level – AC49x/AC50x EC support the injection of comfort noise in order to overcome this issue

• Echo path can change during the call – AC49x/AC50x EC adaptation is fast after such changes are detected

Acoustic Echo Cancellation

• A similar problem to line echo cancellation with some additional complications:

– Longer echo paths, less sparse compared to line echo path

– Worse ERL , can also be negative, howling must be handled

– Worse SNR due to the hands free interface

– Rapid echo path changes

– Higher nonlinearity in the echo path due to loudspeaker characteristics

– Need both in 8KHz and 16KHz sampling rates

• The AC494/AC495 has a an acoustic EC that can handle hands free communication in IP phones.

Speech Compression

• Interoperability is a key issue in VoIP communication systems , therefore Vocoders are usually standardized

• Audiocodes products support a wide range of vocoders

• Support transcoding between different vocoders

Narrowband Vocoders

• G.711 – The most basic vocoder , 64 kbps• G.726 – ADPCM 32 kbps• G.729A – The most popular LBR , 8 kbps• G.723.1 – Developped by Audiocodes, 6.3 kbps, Same

quality less bitrate than G729• iLBC – very robust to packet loss, royalties free, 13 kbps• AMR – used in UMTS, 4.75-12.2 kbps• EVRC – used in CDMA, 8.55 kbps

Next generation speech compression - Wideband Vocoders

• Bandwidth: 50H – 7Khz , 16 bit, 16Khz sampling rate (vs. 300H to 3.4Khz, PCM, 8 Khz rate in NB speech)

• Substantially higher MOS quality

• Superior clarity

• Better Intelligibility (esp. in noise)

• Richer sound

• Similar bit rates (& cost) as NB

• Better speaker recognition (important in conferencing)

• Better quality with music signals

Comparison of NB and WB codecs

Source: ITU G.7291 performance tests

Comparison of NB and WB codecs

• Humans perceives a wideband speech signal as a much higher in quality than narrow band, the difference is big -more than 1 in Wide Band MOS score (4.5 compared to 3.5)

• The MOS difference is even more dramatic when comparing current narrowband codec like G.729 @ 8 kbps to a modern wide band codec like g.729.1 @ 32 kbps

Wide Band Vocoders in AC49x/AC50xx

• G.722 Sub Band Coding ADPCM @ 48,56,64 kbps – used in some high end conferencing systems

• G.722.2 AMR-WB – ACELP @ 12-32 kbps used in UMTS networks

• G.729.1 – CELP @ 8-32 kbps used for VoIP• G.711-WB @ 96 kbps used for VoIP (*)• RTA – Microsoft proprietary Vocoder (*) • Speex – Royalties free vocoder for internet applications• SILK – Skype proprietary vocoder (*)

• (*) – roadmap

Background Noise Reduction

• A new feature planned for AC494 6.6 release

• Used for improved Hands free communication in IP phones

• Optimal Filtering is done in each frequency to suppress the background noise with minimal effect on speech

Noise Reduction Block Diagram

FFT Noise Energy Estimation

Speech Energy EstimationEstimate

Optimal GainIFFT

Noise Reduction Demo

• Male with Car Noise , SNR=12dB – Noisy

– NR

• Male with Car Noise , SNR=6dB– Noisy

– NR

• Female with Office Noise , SNR=18dB– Noisy

– NR

Voice Quality Monitoring

• Telchemy VQ Mon Algorithm – estimate MOS from packet arrival statistics

• RTCP-XR – a standard packet format that carry the Quality parameters

• MOS-CQ – Conversational MOS , takes into account also Echo and Delay influence on the quality

Voice Quality Monitoring

Telchemy VQmon accuracy vs. MOS

Summary

• We described some of the challenges in implementing speech processing algorithms in practical VoIP products

• We described the solutions as implemented in Audiocodes AC490x/AC50x VoIP processors products

Thank you for your time