Page 1

Improvement of Packet Loss Concealment for MP3 Audio Based on Switching of

Concealment Method and Estimation of MDCT Signs

Akinori Ito∗, Kiyoshi Konno∗, Masashi Ito†and Shozo Makino‡

∗Graduate School of Engineering, Tohoku University

6-6-05 Aramaki aza Aoba, Aoba-ku, Sendai, 980-8579 Japan

E-mail: aito@fw.ipsj.or.jp

†Faculty of Engineering, Tohoku Institute of Technology

35-1, Yagiyama-Kasumicho, Taihaku-ku, Sendai, 982-8577 Japan

‡Faculty of Science and Technology, Tohoku Bunka Gakuen University

6-45-1 Kunimi, Aoba-ku, Sendai, 981-8551 Japan

Abstract—This paper describes packet loss concealment

methods for MP3 audio. The proposed methods are based

on estimation of modified discrete cosine transform (MDCT)

coefficients of the lost packets. The estimation of MDCT

coefficients of lower dimensions is performed by switching

two concealment methods: the sign correction method and

the correlation-based method. The concealment methods are

switched based on redundant side information calculated

subband-by-subband for reducing MDCT prediction errors.

Next, a method for improving estimation of MDCT coefficients

of higher dimensions was proposed. The method estimates the

absolute value and sign of an MDCT coefficient independently.

The subjective evaluation experiment proved that both of

the improvement methods for lower and higher dimensions

effectively improved the subjective audio quality.

Keywords-packet loss concealment; MP3; MDCT;

I. INTRODUCTION

Recently many Web services have provided audio and

video streaming services over the Internet. While most of

them employ the transport control protocol (TCP) that is

a unicast protocol based on connection, a connectionless

protocol such as the real-time transport protocol (RTP) [1] is

beneficial because of its suitability to real-time application

and broadcasting. On using the RTP as a transport protocol,

we have to solve a problem of packet losses. Many tech-

niques for recovery from packet losses have been proposed

so far [2], [3]. Considering audio streaming, only simple

recovery methods based on packet repetition is employed

for commercial audio streaming service for speech signal

such as VoIP [4]. However, this is insufficient for applying

to high quality audio streaming such as MP3 [5].

As a previous work, we proposed packet loss conceal-

ment (PLC) techniques for MP3-coded audio signal [6].

In this work, we proposed PLC methods that used signs

of modified discrete cosine transform (MDCT) coefficients

as side information. In addition to the side information,

correlation of MDCT coefficient was also exploited. The

sign information and correlation provided good estimation

of the lost signal [7]. As a result, we could improve signal-

to-noise ratio (SNR) of the decoded signal using the PLC

methods.

In this paper, we first conduct a subjective evaluation of

the previously proposed PLC methods, and point out that

results of the objective and subjective tests are inconsistent.

Next, we investigate the reason of the inconsistency, and

propose a method for further improvement of subjective

audio quality. Finally, we propose a method to improve

quality of high frequency part of the signal where no side

information is given. Combining the proposed methods, we

prove the effectiveness of the methods through subjective

evaluation experiment.

II. PACKET LOSS CONCEALMENT OF MP3-CODED AUDIO

SIGNAL USING SIDE INFORMATION

A. Overview of the method

In this section, we briefly describe the PLC method

proposed in [6]. In MP3 framework, the input audio signal

is split into 32 subbands using a polyphase filter bank, and

the signal of each subband is converted into 18 frequency

components using MDCT. As a result, 576 coefficients are

obtained for one frame of one channel. In this section we

only consider the single-channel case for simplicity, but the

multi-channel signal can be treated in the same way.

Let i-th MDCT coefficients in t-th frame be Mi(t). The

t-th frame M(t) is defined as

M(t) = (M1(t),...,M576(t)).

(1)

Next, we define sign information S(t) as

S(t) = (S1(t),...,SK(t))

(2)

where 1 ≤ K ≤ 576 and

Si(t) = sign(Mi(t))

(3)

sign(x) =

?

1

−1

if

otherwise

x ≥ 0

(4)

2010 Sixth International Conference on Intelligent Information Hiding and Multimedia Signal Processing

978-0-7695-4222-5/10 $26.00 © 2010 IEEE

DOI 10.1109/IIHMSP.2010.132

518

Page 2

In the previous work [6], K was set to 50 that gave suf-

ficient audio quality. Then we transmit MDCT coefficients

of a certain frame combined with sign information of the

different frame. As one MP3 packet of a channel contains

two frames, the k-th packet P(k) is composed as

P(k) = (M(2k −1),S(2k −2),M(2k),S(2k +1)). (5)

When k-th packet is lost, the sign information of frames in

the k-th packet can be obtained from (k−1)-th and (k+1)-

th packets. Then we estimate the MDCT coefficients of t-th

frame as follows.

˜

Mi(t) =

?

f(Si(t),Mi(t + ∆t))

Mi(t + ∆t)

if

otherwise

i ≤ K

(6)

where

∆t=

?

1

−1

t = 2k

t = 2k − 1.

(7)

In addition, the interleaving technique can be used for

increasing robustness against burst packet losses.

Estimation of the lost MDCT coefficients is performed

using the function f(s,m). In the sign correction method,

the following f0is used:

f(s,m) = f0(s,m) = s|m|

(8)

The other method is the correlation-based estimation

method, where the following f1is used:

f(s,m) = f1(s,m) = s

?

wi|m| + (1 − wi)

?

2σ2

π

i

?

(9)

wi=

|ρi1|(|ρi2| − 1)

2|ρi1ρi2| − |ρi1| − |ρi2|

(10)

where σ2is variance of Mi(t), ρi1is correlation coefficient

between |Mi(t)| and |Mi(t − 1)|, and ρi2 is that between

Mi(t) and Si(t).

As predicted in [7], the correlation-based method gave

higher SNR than the sign correction method.

B. Objective and subjective evaluation of the conventional

methods

Although the correlation-based method showed better

SNR, it does not necessarily mean that subjective quality

of the signal restored by the correlation-based method is

better than that by the sign correction method. To investigate

the relationship between objective and subjective quality,

we carried out an evaluation experiment. The experimental

conditions are shown in Table I.

SNR was used as an index of objective evaluation, and

5-grade DMOS (degradation mean opinion score) was used

as that of subjective evaluation. The result is shown in Table

II. As this result depicts, the correlation-based method im-

proved the objective evaluation, but the subjective evaluation

was not improved.

Table I

EXPERIMENTAL CONDITIONS

Testing materialsThree pieces from the RWC music DB

[8] drawn from three genres (Classic, Pop,

Jazz), 15s each

44.1kHz 16bit sampling, 2 channels

2 frames (2304 samples/packet/ch)

10%, 20% (random loss)

MP3, 160kbps

4 males

Signal format

Packet length

Loss rate

Encoding

Subjects

Table II

RESULT OF OBJECTIVE AND SUBJECTIVE EVALUATIONS

Evaluation

SNR [dB]

DMOS

Sign correctionCorrelation-based

11.7

3.99

12.7

3.50

We then investigated the cause of this inconsistency. To

this end, we focused on degradation that occur locally in

time-frequency domain. To analyze the degradation, we

calculated SNR subband by subband. One subband in this

analysis consists of five dimensions of the MDCT coeffi-

cients. Examples of subband SNR are shown in Figure 1.

As these examples show, the correlation-based method gives

better SNR for most of subbands when the degradation is

not severe, especially lower subbands that have larger power.

However, when severe degradation occurs, subbands with

very low SNR are observed in the correlation-based method,

which seems to be a cause of degradation of subjective audio

quality. These local degradations of SNR happen at which

the subband power change rapidly, which causes large error

of estimated values of correlation coefficients and variance

of MDCT coefficients.

III. IMPROVEMENT OF MDCT ESTIMATION USING

CONCEALMENT METHOD SWITCHING

As revealed in the previous experiment, the correlation-

based method gives better SNR for most of subbands and

frames, but the SNR occasionally takes very low value at

several subbands and frames. Therefore, the basic idea of

improving the correlation-based method is to avoid using

the correlation-based method when estimation error of an

MDCT coefficient is large. This idea can be achieved by

examining estimation errors by the two estimation meth-

ods (sign correction and correlation-based method) at the

encoder side and sending information of better method as

another side information.

This improvement can be realized as follows. First, we

consider sending the side information for switching es-

timation method for every 5 dimensions of the MDCT

coefficients when sending packets. Assuming K be a mul-

tiple of 5, we calculate errors of the t-th subband ε2

x(t,j)

519

Page 3

Figure 1.

degradation

Examples of subband SNR for frames with little and severe

(j = 1,...,K/5) as

ε2

x(t,j) =

5j

?

i=5j−4

(Mi(t) − fx(Si(t),Mi(t + ∆t)))2

(11)

where x is either 0 or 1 where f0 or f1 is used as an

estimation method, respectively. Then the information for

method switching is calculated as follows:

Qj(t) =

?

1

0

if

otherwise

ε2

1(t,j) < ε2

0(t,j)

(12)

Using Qj(t), a packet is composed as

P(k) = (M(2k − 1),S(2k − 2),Q(2k − 2),

M(2k),S(2k + 1),Q(2k + 1))

(13)

where

Q(t) = (Q1(t),...,QK/5(t)).

(14)

When t-th packet is lost, the MDCT coefficient is esti-

mated as follows:

˜

Mi(t) =

?

fQj(t)(Si(t),Mi(t + ∆t))

Mi(t + ∆t)

if

otherwise

i ≤ K

(15)

where

j =

?i − 1

5

?

+ 1.

(16)

Note that this method needs another side information

Q(t), which means we need to use smaller K for using same

amount of side information as the conventional methods.

IV. ESTIMATING HIGHER MDCT COEFFICIENTS

WITHOUT SIDE INFORMATION

The method described in the previous section is to im-

prove estimation of MDCT coefficients using side informa-

tion. However, as shown in Eq. (6) and (15), the MDCT

coefficients higher than the K-th dimension are estimated

by just using the coefficients of the neighboring frames.

Therefore, if we can improve estimation of these coefficients,

we achieve further improvement of recovered audio quality.

Two ideas are introduced to achieve this: one is to improve

estimation of absolute value of MDCT coefficients, and

the other one is to improve sign estimation of MDCT

coefficients.

In the previous method, the absolute values of Mi(2k−1)

and Mi(2k) for i > K are estimated as

|˜

Mi(2k − 1)|

|˜

Mi(2k)|

=

=

|Mi(2k − 2)|

|Mi(2k + 1)|.

(17)

(18)

Here, we propose a method to estimate the absolute value

using a linear interpolation:

Di(k) =|Mi(2k + 1)| − |Mi(2k − 2)|

3

(19)

|˜

Mi(2k − 1)|

|˜

Mi(2k)|

=

=

|Mi(2k − 2)| + Di(k)

|Mi(2k − 2)| + 2Di(k).

(20)

(21)

Next, we propose an estimation of sign of an MDCT coef-

ficient. In the previous method, signs of MDCT coefficients

(i > K) are estimated as

˜Si(2k − 1)

˜Si(2k)

=

=

Si(2k − 2)

Si(2k + 1).

(22)

(23)

Instead of just using signs of the previous or next frame,

we propose a method of estimating signs based on bigram

probability, as follows.

˜Si(t) = arg max

s

P(s|Si(t − 1),t,i)

(24)

where Pt(s|s′,t,i) is a conditional probability of s where

the sign of the previous frame is s′(s,s′∈ {−1,1}). The

probability is estimated from the previous N frames Si(t−

N),...,Si(t−1). In the experiment of the next section, the

previous 50 frames were used (N = 50) according to the

result of a preliminary experiment. The actural calculation

of˜Si(t) is performed in simpler way, as follows.

Ti(t) =

t−2

?

τ=t−N

Si(τ)Si(τ + 1)

(25)

˜Si(t) = sign(Ti(t))Si(t − 1)

(26)

520

Page 4

V. EXPERIMENT

A. Experimental conditions

We conducted subjective evaluation of the conventional

and proposed methods. The sign correction method and the

switching method were examined as the estimation method

of MDCT coefficients of lower dimensions. For the higher

dimensions, we compared conventional method (denoted as

Copy) and the proposed method that estimates absolute

values and signs independently (denoted as Interp and sign

est.). 50 bit/frame were used as the side information of the

conventional method (K = 50), and 48 bit/frame were used

for the proposed method (K = 40, 40 bits for signs and

additional 8 bits for method switching).

We employed 9 male subjects for evaluation, and the

packet loss condition was limited to 20% random loss. The

other conditions were same as shown in Table I.

The evaluation was conducted based on the XAB test [9].

The subjective difference grades (SDG) were calculated as

scores of the test signals, which varies from -4 to 0, 0 to be

the best (degradation is imperceptible).

B. Experimental results

Figure 2 shows the experimental result. This result showed

that both improvement of lower dimensions (the conceal-

ment method switching) and that of higher dimensions

(linear interpolation of absolute values and sign estimation)

were effective for improving the subjective audio quality.

From result of one-way layout ANOVA, statistically signifi-

cant differences were found between concealment methods.

Then we conducted multiple comparison test with Bon-

ferroni correction, and significant differences were found

between every two concealment methods, as shown in Figure

2.

As a result, we obtained SDG score of -1.78, which was

1.3 point higher than the conventional method. Moreover,

absolute quality of the proposed method (-1.78) was reason-

ably high (better than “slightly annoying”) considering that

the packet loss rate was severe (20%).

VI. CONCLUSION

We proposed methods for improving subjective audio

quality. First, we conducted subjective evaluation for con-

ventional PLC methods and investigated the cause of quality

degradation using the correlation-based method. Next, we

proposed a method that switches the sign correction method

and the correlation-based method based on subband-by-

subband errors. Furthermore, an improvement method of

MDCT coefficients of higher dimensions was proposed. The

subjective evaluation experiment proved that both of the

improvement methods for lower and higher dimensions were

effective.

Figure 2.Result of the subjective evaluation experiment

REFERENCES

[1] C. Perkins, ”RTP: Audio and Video for the Internet,”

Addison-Wesley, 2003.

[2] C. Perkins, O. Hodson, and V. Hardman, “A survey of packet

loss recovery techniques for streaming audio,” IEEE Network,

vol. 12, no. 5, pp. 40–48, 1998.

[3] B. W. Wah, X. Su, and D. Lin, “A survey of error-concealment

schemes for real-time audio and video transmissions over the

Internet,” in Proc. IEEE Int. Symp. on Multimedia Software

Engineering, 2000.

[4] ETSI, “Substitution and muting of lost frames for full rate

speech channels.” ETSI Recommendations GSM 6.11, 1992.

[5] ISO, “Information technology — coding of moving pictures

and associated audio for digital storage media at up to about

1,5 Mbit/s — part 3: audio.” ISO/IEC 11172-3, 1993.

[6] A. Ito, T. Sakai, K. Konno, S. Makino and M. Suzuki,

“Packet loss concealment for MDCT-based audio codec us-

ing correlation-based side information,” Int. J. of Innovative

Computing, Information and Control, vol. 6, pp. 1347–1362,

2010.

[7] A. Ito and S. Makino, “Designing side information for

multiple description coding,” J. of Information Hiding and

Multimedia Signal Processing, vol. 1, pp. 10–19, 2010.

[8] M. Goto, “Development of the RWC music database,” in

Proc. of 18th Int. Cong. on Acoustics, pp. 553–556, 2004.

[9] International Telecommunication Union, “Methods for the

subjective assessment of small impairments in audio systems,

including multichannel sound systems.” Recommendation

ITU-R BS. 1116-1, 1997.

521