Conversational Speech Quality - The Dominating Parameters in VoIP Systems

Transcription

1 Conversational Speech Quality - The Dominating Parameters in VoIP Systems H.W. Gierlich, F. Kettler HEAD acoustics GmbH Typical IP-Scenarios: components and their influence on speech quality testing techniques for speech quality parameters, selected examples Summary and future work 1

2 Sound Quality & Naturalness Intelligibility Speech Characteristics Listening & Talking Effort Speech Quality Individual Listening Conversational Effort Environmental Conditions Double Talk Performance Transmission of Background Noise Expectation Network Conditions Parameter Influencing Quality TGE_001 E1/T1 E1/T1 PABX IP- Gateway IP IP- Gateway PABX PSTN - PSTN over IP PSTN to IP IP to IP The different IP-Scenarios 2

3 Background noise reduction H(f) - Coder Buffer IP VAD H(f) Decoder Buffer IP VAD Echo canceller Voice controlled attenuation/switching - Comfort noise insertion The IP-Scenario from the speech quality point of view (example) delay, time variant coding ( also with packet loss) (speech controlled) attenuation of signals (e.g. based on VAD) echo cancellation noise reduction and comfort-noise - injection packet loss interactivity perception of echo impairments interactivity clipping effects missing words (sound)-quality clipping effects missing words noise modulation masking e. g. of echo impairments Signal processing and speech quality 3

4 talking situation speech quality conversational situation listening situation Speech quality from the users perspective listening situation talking situation conversational situation - standard"- parameters, e.g. Loudness Ratings - psychoacoustic motivated proced., e.g. PESQ, TOSQA - ext. measurements, e.g. switching characteristics - background noise transmission, e.g. Relative Approach - delay - echo attenuation - echo-/sidetonedistortion - proposals for psychoacoustically motivated tests, e.g. PESQM (perceptual echo and sidetone quality measure) - background noise transmission, e.g. Relative Approach - delay - echo attenuation - echo-/sidetonedistortion - switching characteristics - time variant echo characteristics Relevant quality parameters in the three situations 4

5 SLR CLR CLR RLR 7dB 0dB 0dB 3dB TELR a(echo) a(echo) TELR RLR CLR CLR SLR 3dB 0dB 0dB 7dB OLR = 10dB Fig. 1 G.131 Delay and echo attenuation during single talk TGE_014 echo during double talk: the effect of echo level increase during double talk on subjects judgement Echo level offset single talk/double talk db ,5 2 2,5 3 3,5 4 4,5 -> MOS MOS > loudness variation single/double talk level variation between single- and double talk: the effect on subjects judgement Double talk, important performance limits 5

6 Measurement-technique - selected examples: Listening speech quality Echo & switching during double talk background noise transmission (Relative Approach) comfort-noise injection Measurement for handsets and headsets 6

7 Instrumental Measures based on Hearing Models: Modeling the Results of Auditory Tests by Comparison of Reference Speech Signal with Processed Speech Signals Typical Processing Steps (Schematic): Results of Listening Tests! Adaptation Hearing Model Comparison, Reference, Reduction, Signal Value S index Q Realizations: spectral distance measures PESQ (P.862) TOSQA... General construction of perceptual models TG_e ITU-T 8kbit/sec char. test, exp1 ρ = Combinations of ADPCM Systems ρ = TMOS (TOSQA) single codecs mobile TMOS (TOSQA) real connections simulations 2*tandem single codecs 3*tandem 4+tandem 1.0 MNRU 1.0 MNRU Auditory MOS Auditory MOS 5.0 ETSI - TIPHON 5.0 ETSI - TIPHON div. low bitrate codecs under packet loss div. low bitrate codecs under packet loss ρ = 0.95 ρ = TMOS (TOSQA) G.729 G.728 TMOS (TOSQA) G.723 GSM-FR&G.723 GSM-EFR GSM-FR&G.729 single codecs single codecs 1.0 MNRU 1.0 MNRU Auditory MOS Auditory MOS TOSQA validation, some Results [Berger] 7

8 sending receiving (double talk) transmitted signal in sending, extracted during double talk original test signal in sending test signal in sending test signal in receiving cut off from measured sending signal test signal transfer function switching characteristics H S () f = F s F s () t S { () t } ES Test signal combination for double talk situations test signal in receiving switching during double talk measured signal in receiving echo during double talk Principle of double talk tests 8

9 strong echo components clipping Example of a double talk test strong clipping comfort noise Example of a double talk test 9

10 echo components desyncronisation (delay variation) Example of a double talk test Q = f ( N, S) + f ( 24 F ( i 1) F F (, i n+ ) w (, i F ())) i G 1 2 ( i) w ( if ( i)) + G G 1 G i= 1 n= 1 G ] T F ( i, n) G Basic principle of the Relative Approach: Comparison between short term and long term averaging of signal energies in critical bands (app. 2 s vs. 2 ms) based on a hearing model [Sottek] Background noise transmission: Relative Approach [Genuit] 10

11 onset peak, level variation (level = -35,1 db m0, SLR 16,7 db) clearly visible structure in time and frequency clear structures between 1,6 and 3,5 khz Background noise transmission analysis (example) analysis of the background noise when call was put on hold (example from the beginning) Background noise transmission analysis (example) 11

12 Background noise in sending simultaneously with receiving signal (from loudspeaker) transmitted signal in sending power density spectrum spectral analysis Relative Approach analysis Original -34,6 db m0 comfort noise -36,2 db m0 Comfort noise insertion analysis (example) Summary and outlook: - Speech quality influenced by * condition and load of the network * interaction of network components * interaction of terminal and network * environmental conditions at the users location - Test methods available for various parameters - Further investigations needed on noise transmission overall quality - number Summary 12