Enhancing Conversational Speech Quality of VoIP in a Wired/Wireless Environment

Batu Sat and Benjamin W. Wah

Illinois Center forWireless Systems

Background

Proposed Solutions

Results

VoIP• Providing interactive speech among multiple users • Utilizing public and private wired/wireless IP networks • Independent of locations of users and devices usedIP Networks• Long-haul, WAN, LAN wired/wireless networks• Non-stationary real-time packet arrivals and losses• Large disparity in delay and loss behavior among clients• Complex QoS with multiple IP providers and without cost model• Quality measured and maintained at end-points• Better scalability with end-to-end strategies

Conversational Dynamics• Different network delays among clients• Multiple realities in VoIP in contrast to face-to-face conversation• Perception of delays and efficiency affected by conversational switching (turn-taking) frequency

Conversational Dynamics & Quality

Goals

Challenges

Design of VoIP End Clients• Achieving high and consistent perceptual conversational quality• Enabling natural and efficient conversation among users• Real-time adaptation to changing network delay & loss conditions• Suitable for any communication device using any IP network

Quality Metrics• No objective metrics for quantifying conversational speech quality• Costly non-repeatable subjective tests with full implementation

Design of Play-out Scheduling and Loss Concealment• Under dynamic packet delays and losses

• Trade-offs among mouth-to-ear delay (MED), redundancy, and

amount of packets not received in time for play-out (UCFLR)

• Difficulty under dynamic delay spikes and bursty losses • With longer MED

• Improved one-way speech quality• Degraded symmetry and efficiency of interactive conversation

• Trade-off between minimizing pair-wise MED and maintaining a

balance among MEDs perceived by users in a conversation

Conversational Speech Quality • Multiple dimensions in user perception of quality • Quality of one-way speech segments• Naturalness and rhythm of conversation, mutual-silence durations

Trade-Offs

Collection of Traces on Delays and Losses• Using Planet-Lab nodes for collecting end-to-end traces• With packet periods and payloads typical of VoIP applications

Modeling of Two-Party and Multi-Party Conversations• Utilizing human psychological models when possible• Subjective tests to obtain parameters for simulating dynamics

Evaluation of Conversational Speech Quality (CSQ)• Identification of human-observable and system-measurable metrics• Modeling CSQ as function of these metrics• Designing human subjective tests

Designing Play-out Scheduling/Loss concealment schemes• Trade-offs on system measurable and human-observable metrics• Schemes for real-time collection and relay of network statistics• Schemes for real-time adaptive POS and LCS

Face-to-face setting:(A & B’s common perspective)

A B A B time

VoIP setting:

(A’s perspective)

(B’s perspective)

time

time

A

A’ B

B’ A

A’ B

B’

A speaks

B speaks

A thinks

B thinks

MED(AB)

MED(BA)

Legend:

PublicInternetPublic

Internet

Wireless

Wireless

Private IP Network

Private IP Network

Trace

#Jitter Loss

2 Low 1.7%

5 - 17%

9 High 0.1%

10 - 33%

None of the previous algorithms provides consistent balance between one-way speech quality and conversational interactivity

Our scheme• Hugging delay curve closely• Minimizing delay degradations• Providing good one-way quality• Maximizing human quality perception