Performance Study of Objective Speech Quality Measurement for Modern Wireless-VoIP Communications

Wireless-VoIP communications introduce perceptual degradations that are not present with traditional VoIP communications. This paper investigates the e ﬀ ects of such degradations on the performance of three state-of-the-art standard objective quality measurement algorithms—PESQ, P.563, and an “extended” E-model. The comparative study suggests that measurement performance is signiﬁcantly a ﬀ ected by acoustic background noise type and level as well as speech codec and packet loss concealment strategy. On our data, PESQ attains superior overall performance and P.563 and E-model attain comparable performance ﬁgures.


Introduction
Due to the "best-effort" nature of current Internet Protocol (IP) connections, real-time speech quality monitoring is needed in order to maintain acceptable quality of service for voice over IP (VoIP) communications [1].Traditionally, subjective quality assessment tests, such as the mean opinion score (MOS) test [2,3], are used to quantify perceived speech quality.Subjective tests, however, are expensive and time-consuming and, for the purpose of real-time quality monitoring, have been replaced by objective speech quality measurement methods.
For VoIP communications, objective methods can be classified as either signal or parameter based.Signal-based methods use perceptual features extracted from the speech signal to estimate quality.Parameter-based methods, on the other hand, use VoIP connection parameters, such as codec, packet loss pattern, loss rate, jitter, and delay, to compute impairment factors which are then used to estimate speech quality.Such parameters are commonly obtained from the real-time transport protocol (RTP) header [4], real-time transport control protocol (RTCP) [5], and RTCP extended reports (RTCP-XRs) [6].
Current state-of-the-art signal based quality estimation algorithms perform well for traditional telephony applications but recent studies have found large "per-call" quality estimation errors and error variance [7][8][9].Large estimation errors limit the use of signal-based methods for online quality monitoring and control purposes [10].Parameterbased methods, on the other hand, can provide lower percall quality estimation errors [11,12] and have been widely deployed in VoIP communication services.The major disadvantage of parameter-based measurement is that distortions that are not captured by the connection parameters are not accounted for.Examples of such distortions include acoustic noise type, temporal clippings, noise suppression artifacts, as well as distortions in tandem connections caused by unidentified equipment and signal conditions.
Today, with the emergence of advanced technologies such as wireless local and wide area networks, the number of wireless-VoIP connections has grown substantially [13,14].Recent research by consulting firm ON World has suggested that by 2011 the number of wireless-VoIP users around the world will rise to 100 million from 7 million in 2007 [15].Wireless-VoIP inter-networking results in tandeming of heterogeneous links which can produce new impairment combinations that are not addressed by current standard quality measurement algorithms.Representative combinations of distortions can include (i) acoustic background noise (with varying levels and types) combined with packet loss concealment artifacts, (ii) acoustic background noise combined with temporal clipping artifacts (resultant from voice activity detection errors), or (iii) noise suppression artifacts (and residual noise) combined with speech codec distortions.In this paper, the effects of such "modern" degradation combinations on the performance of three International Telecommunications Union ITU-T standard algorithms are investigated.Focus is placed on listening quality, hence, factors such as jitter and delay, which affect conversational quality [16], are not considered.
The remainder of this paper is organized as follows.First, a brief overview of subjective and objective quality measurement is given in Section 2. Simulated wireless-VoIP impairments and details about the subjective listening test are presented in Section 3. Analysis of variance tests and quantitative algorithm performance comparisons are described in Section 4 and the quantification of overall performance loss is presented in Section 5. Lastly, conclusions are drawn in Section 6.

Speech Quality Measurement
In this section, a brief overview of subjective and objective speech quality measurement methods is given.

Subjective Measurement.
Speech quality is the result of a subjective perception-and-judgment process, during which a listener compares the perceptual event (speech signal heard) to an internal reference of what is judged to be good quality.Subjective assessment plays a key role in characterizing the quality of emerging telecommunications products and services, as it attempts to quantify the end user's experience with the system under test.Commonly, the mean opinion score (MOS) test is used wherein listeners are asked to rate the quality of a speech signal on a 5point scale, with 1 corresponding to unsatisfactory speech quality and 5 corresponding to excellent speech quality [2,3].The average of the listener scores is termed the subjective listening MOS, or as suggested by ITU-T Recommendation P.800.1 [17], MOS-LQS (listening quality subjective).Formal subjective tests, however, are expensive and time consuming, thus unsuitable for "on-the-fly" applications.

Objective Measurement.
Objective speech quality measurement replaces the listener panel with a computational algorithm, thus facilitating automated real-time quality measurement.Indeed, for the purpose of real-time quality monitoring and control on a network-wide scale, objective speech quality measurement is the only viable option.Objective measurement methods aim to deliver quality estimates that are highly correlated with those obtained from subjective listening experiments.As mentioned previously, for VoIP communications objective quality measurement can be classified as either signal based or parameter based.Such measurement methods are described in the subsections to follow.

Signal-Based Measures.
Signal based methods can be further classified as double-input (Figure 1(a)) or singleinput (Figure 1(b)) depending on whether a clean reference signal is required or not, respectively.Such schemes are commonly referred to as double-ended or single-ended, respectively.Research into double-ended signal based quality measurement dates back to the early 1980s [18].ITU-T Recommendation P.862 [19] (better known as perceptual evaluation of speech quality, PESQ) is the current state-ofthe-art double-ended standard measurement algorithm.An in-depth description of the PESQ algorithm is available in [20,21].
Single-ended measurement, on the other hand, is a more recent research field, and only recently (late 2004) has an algorithm been standardized.The ITU-T Recommendation P.563 [22] represents the current state-of-the-art singleended standard algorithm for traditional telephony applications.A detailed description of the P.563 signal processing steps is available in [23].Throughout the remainder of this paper, listening quality MOS obtained from an objective model will be referred to as MOS-LQO [17].

Parameter-Based Measures.
Parameter based measurement, as depicted in Figure 1(c), was first proposed in the early 1990s by the European Telecommunications Standards Institute (ETSI).The ETSI computation model (so-called E-model) was developed as a network planning tool and describes several parametric models of specific network impairments and their interaction with subjective quality [24].In the late 1990s, the E-model was standardized by the ITU-T as Recommendation G.107 [25].The basic assumption of the E-model is that transmission impairments can be transformed into psychological impairment factors, which in turn, are additive in the psychoacoustic domain.
A transmission rating factor R is obtained from the impairment factors by where I s , I d , and I e−e f f represent impairment factors due to transmission (e.g., quantization distortion), delay, and effective equipments (e.g., codec impairments at different packet loss scenarios), respectively.R 0 describes a base factor representative of the signal-to-noise ratio and A an advantage factor; the R rating ranges from 0 (bad) to 100 (excellent).
If the delay impairment factor I d is not considered, the R rating can be mapped to listening quality MOS by means of equations described in ITU-T Recommendation G.107 Annex B [25].Throughout the remainder of this paper, listening quality MOS obtained from E-model planning estimates will be referred to as MOS-LQE [17].
Several improvements to Recommendation G.107 have been proposed or are under investigation [26] in order to incorporate more modern transmission scenarios.Impairment factors, obtained from subjective tests, are described in [25,27,28] for several common network configurations.Impairment factors for alternate configurations can be obtained either from subjective MOS tests (according to ITU-T Recommendation P.833 [29]) or from objective methods (Recommendation P.834 [30]).As mentioned previously, the E-model is a transmission planning tool and is not recommended for online quality measurement.Hence, several extensions have been proposed to improve E-model performance for online monitoring.Representative extensions include nonlinear impairment combination models to compensate for high levels of "orthogonal" (unrelated) impairments [31], or online signal-to-noise ratio (SNR) estimation to account for varying background noise levels [32].In this study, an "extended" E-model implementation is used.With the extended version, nontabulated equipment impairment factors (e.g., codecs described in Section 3 under 4% random and bursty packet losses) are obtained from subjectively scored speech data [12,29].Moreover, since degraded speech files have been artificially generated, the true noise level is used to compute MOS-LQE for noisecorrupted speech.Note that extended E-model performance may be favored with this unrealistic assumption that true noise level information is available online.In order to investigate a more realistic scenario, the noise level is measured in real time and is incorporated into the E-model in a manner similar to that described in [12,32].Here, the "noise analysis" module available in P.563 [22] is used to estimate noise levels online for the noisy and noisesuppressed speech signals.In controlled experiments, the noise level meter attained a correlation of 0.96 with the true noise level, computed both prior to and post-noisesuppression.
Moreover, equipment impairment factor values are currently not available for noise suppression algorithms and are the focus of ongoing research [26].In fact, artifacts introduced by such enhancement schemes are dependent on the noise type and noise levels.In our experiments, the estimated noise level (post enhancement) is used in the computation of MOS-LQE for noise-suppressed speech.It is important to emphasize, however, that while using the estimated (or measured) noise level is convenient for quantifying noise artifacts that remain after enhancement, noise suppression artifacts that arise during speech activity are not accounted for.As emphasized in Section 5, this is a major shortcoming of parameter based measurement methods.

Experiment Setup
In this section, the degradation conditions available in the datasets-simulated wireless-VoIP, reference, and conventional VoIP-as well as the subjective listening tests are described.

Wireless-VoIP Degradation Conditions.
The source speech signals used in our experiments are in English and French (four signals per language) and have been artificially corrupted to simulate distortions present in modern wireless-VoIP connections.Degradation sources that are commonly present in the wireless communications chain can include signal-based distortions such as acoustic background noise or noise suppression artifacts.These impairments are combined with distortions present in the VoIP chain, which may include codec distortions and packet loss concealment (PLC) artifacts.
To simulate the effects of acoustic background noise and codec distortions (including PLC artifacts), clean speech signals are corrupted by three additive noise sources (hoth, babble, car) at two SNR levels (10 dB and 20 dB).Noisy speech is then processed by three speech codecs: G.711, G.729, and Adaptive Multi rate (AMR).Random and bursty packet losses are simulated at 2% and 4% using the ITU-T G.191 software package [33]; the Bellcore model is used for bursty losses.Losses are applied to speech packets, thus simulating a transmission network with voice activity detection (VAD).The G.729 and AMR codecs are equipped with built-in PLC algorithms to compensate for lost packets.For the G.711 codec, two PLC strategies are investigated: the one described in [34] and a simple silence insertion scheme which is included to investigate the effects of acoustic noise combined with temporal clipping artifacts; the latter is referred to as "G.711 * " throughout the remainder of this paper.Packet sizes are 10 milliseconds for the G.729 codec and 20 milliseconds for the remaining codecs.Moreover,  1.
Noise suppression artifacts combined with codec distortions are used to further simulate impairments introduced by wireless-VoIP connections.Here, the noise suppression algorithm available as a preprocessing module in the SMV codec is used [35].The SMV codec per se is not used in our experiments as ITU-T P.563 has not been fully validated for such technologies [22].Clean speech is corrupted by four noise types (hoth, car, street, and babble) at three SNR levels (0 dB, 10 dB, and 20 dB).Noisy speech is processed by the noise suppression algorithm and the noise-suppressed signal is input to the G.711, G.729 or AMR speech codec.As mentioned above, tandem conditions are also considered for the G.729 and AMR codecs.For noise suppression related impairments, a total of 192 speech signals (half English half French) are available, covering 24 degradation conditions as described in Table 2.

Reference Degradation Conditions.
The multilingual datasets also include 128 reference-condition speech files which are commonly used in subjective listening tests to facilitate validation of test measurements and comparison with measurements from other tests.Reference conditions include modulated noise reference unit (MNRU) [36] at seven different signal-to-noise ratios (5-35 dB, 5 dB increments), as well as G.711, G.729, and AMR codecs operating in clean conditions either singly or in tandem.As described in Section 4.1, these reference conditions are used to map the datasets to a common MOS-LQS scale.

Conventional VoIP Degradation Conditions.
Conventional VoIP degradation conditions are also included with the English and French datasets.With conventional VoIP conditions, clean speech, as opposed to noise-corrupted or noise-suppressed speech, is processed by the G.711, G.711 * , G.729, and AMR codec-PLC schemes (singly or in tandem), under 2% and 4% random and bursty packet loss conditions.A total of 192 speech files (half English half French) covering 24 degradation conditions (no tandem: 4 codec-PLC types × 2 loss types × 2 loss rates; tandem: 2 codecs × 2 loss types × 2 loss rates) are available.Conventional VoIP data is used as a benchmark in Section 5 to quantify the decrease in quality measurement accuracy due to wireless-VoIP distortions.

Subjective Listening Tests. Source speech files were recorded in an anechoic chamber by four native Canadian
French talkers and four native Canadian English talkers.Half of the talkers were male and the other half female.Clean speech signals were filtered using the modified intermediate reference system (MIRS) send filter according to ITU-T Recommendation P.830 Annex D [37].Degraded speech signals were further filtered using the MIRS receive filter.In both instances, speech signals were level adjusted to −26 dBov (dB overload) and stored with 8 kHz sampling rate and 16-bit precision.Similar to the ITU-T Supp.23 dataset [38], each speech file comprises two sentences separated by an approximately 650 milliseconds pause.
Two subjective MOS tests (one per language) were conducted in 2006 following the requirements defined in [2,37].Sixty listeners, native in each language, participated in each listening quality test and rated processed speech files described in Sections 3.1-3.3.Listener gender ratio was roughly one-to-one and listeners consisted of naïve adults (aged 18-50) with normal hearing recruited from the general population.Beyerdynamic DT 770 headphones were used and the listening room ambient noise level was kept below 28 dBA.Statistics for the subjective scores collected in the listening tests are listed in Table 3 for the wireless-VoIP, reference, and conventional VoIP degradation conditions.

Performance of ITU-T Standard Algorithms
In this section, two methods are used to assess the performance of PESQ, P.563, and the extended E-model for the wireless-VoIP distortion combinations described in Section 3.1.The first method, based on analysis of variance tests, investigates the performance sensitivity of current state-ofthe-art algorithms to different wireless-VoIP degradation sources, such as noise type, noise level, packet losses, and codec-PLC type.The second method uses correlation and root-mean-square errors, computed between MOS-LQS and MOS-LQO (or MOS-LQE), to quantify the performance of

Analysis of Variance.
In this section, factorial analysis of variance (ANOVA) is used to assess the effects of different wireless-VoIP degradation on objective quality measurement performance.In particular, the effects of codec-cum-PLC type and acoustic background noise (type and level) are investigated using the noise-corrupted and noise-suppressed speech signals described in Section 3.1.For noise-corrupted speech, the effects of packet loss rates and packet loss patterns (random or bursty) are also investigated.
For the purpose of real-time quality monitoring and control, it is known that objective measures are required to provide low per-call estimation errors.Hence, we use persample MOS residual as the performance criterion; MOS residual is given by MOS-LQO minus MOS-LQS (or MOS-LQE minus MOS-LQS).In the analysis, raw MOS-LQO and MOS-LQE results (without mappings) are used.As shown in Section 4.2, mappings such as the one described in [39] can actually decrease algorithm performance.
In order to obtain a sufficiently large number of samples for variance analysis, a combined English-French speech dataset is used.It is important to emphasize that the English and French datasets were produced concurrently by the same organization, under identical conditions, with the only differences between them being the speakers, the spoken text, and the listener panels.Notwithstanding, in order to remove the differences between the English and French subjective scales, as suggested by the statistics in Table 3, a third-order monotonic mapping is trained between the English reference-condition MOS-LQS values and the French reference MOS-LQS values.The scatter plot in Figure 2 illustrates the French versus English reference MOS-LQS values available in the datasets; the dotted curve represents the obtained polynomial.The combined dataset used for analysis comprises the French dataset described in Section 3.1 and the English dataset mapped to the French scale.This English-to-French scale mapping was chosen as it resulted in a larger reduction in mean absolute error between the two datasets of 0.13 MOS.After the mapping is applied, no significant differences are observed between the scales, as suggested by a t significance test (P = .22). 4 shows F-statistics (F) and Pvalues (P) obtained from factorial ANOVA for noisecorrupted speech files.As can be seen, with a 95% confidence level, codec-PLC and noise type have significant main effects (i.e., P < .05) on the performance of all three objective measures.Noise level is shown to have significant main effects on E-model and PESQ performance and packet loss rate only on E-model performance.The box and whisker plots depicted in Figures 3(a)-3(c) assist in illustrating these behaviors, respectively; the plots illustrating the effects of packet losses on E-model performance are omitted for brevity.

Main Effects. Table
The boxes have lines at the lower quartile, median, and upper quartile values; the whiskers extend to 1.5 times the interquartile range.Outliers are represented by the symbol "+".The vertical width of the notches that cut into the boxes at the median line indicates the variability of the median between samples.When the notches of two boxes do not overlap, their medians are significantly different at the 95% confidence level [40].In the plots, abscissa labels are omitted to avoid crowding; the missing labels can be obtained by periodically replicating the shown labels.Moreover, the abbreviation "LQO-LQS" is used for the ordinate labels to represent the MOS residual for all three measurement algorithms.
From Figure 3(a), it can be seen that larger E-model residual errors are attained for the silence insertion PLC scheme (represented by "G711 * ") followed by the AMR codec-cum-PLC.Furthermore, P.563 performance is lower for G.711-processed speech irrespective of the PLC strategy.According to [22], P.563 has only been validated for PLC schemes in CELP (codebook-excited linear prediction) codecs such as G.729; this can explain the poor performance obtained for G.711.Nonetheless, for the G.729 codec, P.563 attains residual errors that can be greater than one MOS point; on a five-point MOS scale, this can be the difference between having acceptable and unacceptable quality [9].
From Figures 3(b) and 3(c), it can be observed that E-Model and PESQ underestimate MOS-LQS for speech corrupted by car noise; E-model underestimates MOS-LQS for all noise types and levels.Figure 3(c) shows that P.563 performance is not significantly affected by noise level.This may be due to the fact that P.563 is equipped with a noise analysis module which not only estimates the SNR but also takes into account other spectrum-related measures such as high frequency spectral flatness.High frequency analysis, however, may be the cause of P.563 sensitivity to noise type since babble and car noise have low-pass characteristics.Ongoing research is seeking a better understanding of the limitations of signal [41,42], and parameter-based [26] measurement of noisy speech.
Table 5 shows F-statistics and P-values obtained from factorial ANOVA for noise-suppressed speech files.With a 95% confidence level, it can be seen that for noise-suppressed speech, codec-PLC type incurs significant main effects on E-model and PESQ performance.Noise type significantly affects PESQ and P.563 performance, and noise level (prior to noise suppression) incurs significant main effects on the performance of all three algorithms.The box and whisker plots depicted in Figures 4(a)-4(c) help illustrate these behaviors, respectively.As seen from the plots, E-model performance is inferior for AMR-processed speech.Both PESQ and P.563 underestimate MOS-LQS for car noise and overestimate MOS-LQS for hoth and street noise.Moreover, similar effects of noise level on estimation accuracy are observed for all three algorithms, with superior performance attained for speech corrupted by noise at higher SNR levels (10 dB and 20 dB).At low SNR (0 dB prior to noise suppression), all three algorithms overestimate MOS-LQS and PESQ attains superior performance.
Table 6 summarizes all significant main effects of wireless-VoIP impairments on PESQ, P.563, and E-model performance for both noisy and noise-suppressed degradation conditions.For noise-suppressed speech, packet loss rate effects were not included in the datasets hence are represented by the term "NI" in the table.As observed, PESQ and E-model performance are most sensitive to wireless-VoIP distortions.

Two-Way Interactions.
Factorial ANOVA with a 95% confidence level has suggested four significant two-way interaction effects on noise-corrupted speech files: Significant interaction effects were not observed for noise-suppressed speech.Figures 5(a) and 5(b) depict box and whisker plots that help illustrate the two-way interaction effects of codec-PLC and noise type as well as codec-PLC and noise level on E-model performance.Plots illustrating the remaining two-way interactions are omitted for brevity.As can be seen from the plots, inferior performance is attained for babble and car noise (and for SNR = 20 dB) with G.711-processed speech with the silence insertion packet loss concealment scheme (G711 * ).In such scenarios, perceptual artifacts are introduced due to the sudden changes in signal energy (i.e., temporal clippings); such artifacts are not accounted for by E-model quality estimates if the speech sample is additionally corrupted by noise.Other algorithms such as G.729 and AMR are equipped with comfort noise

MOS-LQO versus MOS-LQS for the English dataset obtained
with the E-model, P.563, and PESQ, respectively.Plots (a) and (b) are after 3rd-order polynomial mapping and plot (c) depicts PESQ MOS-LQO before ("•") and after ("×") the mapping described in [39].As can be seen from the plots and from Table 7, PESQ performance decreases once the mapping is applied.This suggests that an alternate mapping function needs to be investigated for modern degradation conditions such as those present in wireless-VoIP communications.

Figure 1 :
Figure 1: Block diagram of (a) double-and (b) single-ended single-based measurement, and (c) parameter-based measurement.

Figure 2 :
Figure 2: Scatter plot of English and French MOS-LQS values obtained from anchor conditions available in the datasets.The dotted line depicts the obtained third-order monotonic mapping used for dataset combination.

Figure 3 :Figure 4 :
Figure 3: Significant main effects of (a) codec, (b) noise type, and (c) noise level on the accuracy of objective quality measurement of noise corrupted speech.

Table 1 :
Description of the 54 available noise-related degradation conditions.

Table 2 :
Description of the 24 available noise suppression related degradation conditions.

Table 3 :
Subjective score (MOS-LQS) statistics, separately for the English and French speech files, for the wireless-VoIP, reference, and conventional VoIP degradation conditions.

Table 7 :
Per-condition performance of E-model, PESQ, and P.563 on wireless-VoIP degradation conditions available in the English and French datasets.Post mapping performance is represented by R * and RMSE * .