a statistical model-based double-talk detection incorporating soft ...

A STATISTICAL MODEL-BASED DOUBLE-TALK DETECTION
INCORPORATING SOFT DECISION
Yun-Sik Park, Ji-Hyun Song, Sang-Ick Kang, Woojung Lee and Joon-Hyuk Chang
School of Electronic Engineering
Inha University, Incheon, Korea
E-mail: [yspark, jhsong, sikang, wjlee]@dsp.inha.ac.kr, changjh@inha.ac.kr
ABSTRACT
In this paper, we propose a novel double-talk detection
(DTD) technique based on a soft decision in the frequency
domain. The proposed method provides an efficient procedure
to detect the double-talk situation by the use of the
global near-end speech presence probability (GNSPP) and
voice activity detection (VAD) of the near-end and far-end
signal. Specifically, the GNSPP is derived based on a statistical
method of speech and is employed to determine the
double-talk presence in a given frame. The performance of
our approach is evaluated by objective tests under different
environments, and it is found that the suggested method yields
better results compared with the conventional scheme.
Index Terms— Double-Talk Detection, Speech Presence
Probability, Voice Activity Detection
1. INTRODUCTION
In most hands-free mobile communication systems, since
the loudspeaker and microphone are acoustically coupled,
acoustic echoes occur. In efforts to address this problem,
numerous acoustic echo cancellation (AEC) techniques incorporating
an adaptive filter such as the least mean square
(LMS) and normalized LMS (NLMS) have been reported [1]-
[3]. One of the major problems of AEC techniques, however,
is that the performance significantly degrades during
the double-talk periods, in which signals from both the nearend
and far-end coexist because the double-talk acts as very
large interference to the adaptive filter. The problem can be
alleviated by freezing the adaptive filter coefficients through
the use of a double talk detection (DTD) algorithm [4]. In
this regard, many studies have been dedicated to the problem
of DTD. In practice, cross-correlation and coherence-based
algorithms are relevant, as they present straightforward approaches.
Adopting hard decisions [4]-[6], these schemes
classify each frame into one of two (i.e., double-talk or not)
This work was supported by National Research Foundation of Korea
(NRF) grant funded by the Korean Government (MEST) (NRF-2009-
0072319) and This work was supported by the IT R&D program of
MKE/KEIT. [2008-F-045-02].
cases by comparing decision statistics and given threshold
values. However, they are sensitive to optimized parameters
and do not always provide reliable performance under various
conditions.
In this paper, we propose a novel DTD algorithm based
on a global soft decision [7], where the term ‘global’ means
that DTD is performed globally in a given frame and ‘soft decision’
[8], [9] denotes that the probability of double-talk is
introduced as a decision and is applied to update the adaptive
filter in the acoustic echo suppressor (AES) algorithm [10].
Specifically, the global near-end speech presence probability
(GNSPP) based on a statistical model is computed in each
frame to apply the proposed DTD algorithm in conjunction
with results of voice activity detection (VAD) of the near-end
and far-end signal. It is worth noting that our approach provides
for the first time an effective framework of DTD based
on a soft decision by taking advantage of a statistical model,
in contrast with the conventional hard decision-based method.
The performance of the proposed algorithm is evaluated by
echo return loss enhancement (ERLE) and speech attenuation
(SA) tests during double-talk and is demonstrated to be better
than that of the conventional method.
2. GLOBAL NEAR-END SPEECH PRESENCE
PROBABILITY
In this section, we consider how to derive the global nearend
speech presence probability (GNSPP) in the frequency
domain. To this end, we first assume that two hypotheses,
H 0 and H 1 , indicate near-end speech absence and presence
as follows:
H 0 : near-end speech absent : Y(i) =D(i) (1)
: near-end speech present : Y(i) =D(i)+S(i)
H 1
where D(i) = [D(i, 1),D(i, 2),...,D(i, M)], S(i) =
[S(i, 1),S(i, 2),...,S(i, M)] and Y(i) =[Y (i, 1),Y(i, 2),
...,Y(i, M)], respectively, represent the Fourier domain
spectra of the echo signal, the near-end speech and the microphone
input signal with a frame index i. Also, X(i) =
[X(i, 1),X(i, 2),...,X(i, M)] denote the Fourier spectrum
978-1-4244-4296-6/10/$25.00 ©2010 IEEE 5082
ICASSP 2010

of the far-end signal as shown in Fig. 1. The background
noise is not taken into account since we assume that near-end
speech absence is not correlated with the background noise.
Under the assumption that D(i, k) and S(i, k) are characterized
by separate zero-mean complex Gaussian distributions,
the following is obtained [7].
p(Y (i, k)|H 0 ) =
p(Y (i, k)|H 1 ) =
1
[
πλ d (i, k) exp −
|Y (i,
]
k)|2
λ d (i, k)
1
π(λ s (i, k)+λ d (i, k))
[
|Y (i, k)| 2 ]
exp −
λ s (i, k)+λ d (i, k)
where λ s (i, k) and λ d (i, k) are the variance of the near-end
speech and estimated echo, respectively. Accordingly, the
GNSPP p(H 1 |Y(i)) is derived from Bayes’ rule, such that
[7]:
p(Y(i)|H 1 )p(H 1 )
p(H 1 |Y(i)) =
p(Y(i)|H 0 )p(H 0 )+p(Y(i)|H 1 )p(H 1 )
(4)
where p(H 0 )(= 1−p(H 1 )) represents the a priori probability
of near-end speech absence. Since the spectral component in
each frequency bin is assumed to be statistically independent,
(4) can be rewritten as [7]
p(H 1 |Y(i)) (5)
M∏
p(H 1 ) p(Y (i, k)|H 1 )
k=1
=
M∏
M∏
p(H 0 ) p(Y (i, k)|H 0 )+p(H 1 ) p(Y (i, k)|H 1 )
k=1
M∏
q Λ k (Y (i, k))
k=1
=
M∏
1+q Λ k (Y (i, k))
k=1
k=1
in which q = p(H 1 )/p(H 0 )(= 1) which is determined by
the rough estimate of the ratio of absence time duration and
presence time duration for near-end speech and Λ k (Y (i, k))
is the likelihood ratio computed in the kth frequency bin, as
given by [7]:
Λ k (Y (i, k)) = p(Y (i, k)|H 1)
(6)
p(Y (i, k)|H 0 )
1
[ γ(i, k)ξ(i, k)
]
=
1+ξ(i, k) exp 1+ξ(i, k)
where the a posteriori signal-to-echo ratio (SER) γ(i, k) and
the a priori SER ξ(i, k) are defined by
(2)
(3)
Microphone
Near-end
Loudspeaker
DFT
IDFT
VAD
Decision
GNSPP
DTD
SER Esimation
Echo Path Response
Estimation
x
G
Send path
IDFT
DFT
Far-end
Receive path
Fig. 1. Block diagram of the proposed DTD algorithm.
ξ(i, k) ≡ λ s(i, k)
(8)
λ d (i, k)
where λ d (i, k) is estimated by ˆλ d (i, k). The power spectrum
of the echo signal is obtained in the case of the absence of the
near-end speech signal, as given by
ˆλ d (i, k) =ζ Dˆλd (i − 1,k)+(1− ζ D )|Ŷ (i, k)|2 (9)
in which ζ D (= 0.93) is the smoothing parameter. Also, in (8),
ξ(i, k) is estimated with the help of the well-known decisiondirected
approach with α DD =0.6 [11]. Then,
|Ŝ(i − 1,k)|2
ˆξ(i, k) =α DD
λ d (i − 1,k) +(1− α DD)P {γ(i, k) − 1}
(10)
where P {z} = z if z ≥ 0, and P {z} =0otherwise. As
specified in (9), the robust estimation of the echo magnitude
spectrum Ŷ (i, k) plays an essential role in the performance.
In our approach, we follow the parameter estimation procedure
proposed in [10] as follows:
|Ŷ (i, k)| = Ĥ(i, k)|X(i, k)| (11)
where Ĥ(i, k) is the estimate for the echo path response mimicking
the actual echo path. Specifically, Ĥ opt (i, k) is obtained
based on the magnitude of the least squares estimator
as follows [10]:
Ĥ opt (i, k) =
E[X ∗ (i, k)Y (i, k)]
∣E[X ∗ (i, k)X(i, k)] ∣ (12)
where ∗ denotes the complex conjugate and E[·] represents
the expected value. Note that there exist some delay between
the far-end speech X(i, k) and the microphone input signal
Y (i, k) (due to a digital amplifier, e.g.). In our approach, it is
assumed that the echo time-delay is separately estimated and
compensated (i.e., no delay) at the near-end. Since the echo
path is time varying, the estimated echo path response Ĥ(i, k)
is obtained using the iterative procedure such that[10]
γ(i, k) ≡
|Y (i, k)|2
λ d (i, k)
(7)
Ĥ(i, k) =
C(i, k)
R(i, k)
(13)
5083

Far−end Echo
Signal
Near−end Speech
Signal
Microphone Input
Signal
p(H 1
|Y(i))
1
0
−1
1
0
−1
1
0
−1
1.0
0.5
0.0
Noise Far−end echo Double−Talk Near−end Speech Noise
Table 1. ERLE and SA test results obtained from the proposed
DTD algorithm based on a soft decision with those
yielded by the conventional hard decision method during
double-talk.
Environments ERLE (dB) SA (dB)
Noise SNR (dB) Park-based proposed Park-based proposed
10 3.23 2.96 1.75 1.64
White 20 3.46 3.30 1.99 1.86
30 3.51 3.36 2.03 1.90
10 3.36 3.23 1.81 1.71
Babble 20 3.48 3.33 1.20 1.87
30 3.52 3.36 2.03 1.90
10 2.95 2.84 1.56 1.47
Vehicle 20 3.40 3.25 1.96 1.83
30 3.50 3.35 2.02 1.90
Clean speech ∞ 3.52 3.37 2.03 1.91
0 1 2 3 4
Time (sec)
Fig. 2. DTD results for the acoustic echo signal under the
vehicular noise condition (SNR=20 dB).
where
C(i, k) =ζ C C(i − 1,k)+(1− ζ C )|X ∗ (i, k)Y (i, k)| (14)
R(i, k) =ζ R R(i − 1,k)+(1− ζ R )|X ∗ (i, k)X(i, k)|, (15)
and ζ C (= 0.998) and ζ R (= 0.998) are smoothing parameters.
Note that this update iteration achieves the room change
tracking.
3. PROPOSED DOUBLE TALK DETECTION
BASED AES
As noted earlier, the update of the echo path response must
be frozen in the case of the double-talk. For this, we propose
the DTD technique to incorporate the newly derived GNSPP,
p(H 1 |Y(i)), with the help of the VAD results of the near-end
and far-end signal, as shown in Fig. 1. We inherently consider
the near-end speech presence in the case of far-end signal
presence, where the GNSPP substantially determine the
double-talk situation and is used to update the echo path response
based on (12). Note that the VAD has an impact on
the near-end speech presence and the far-end speech presence
only. Specifically, we derive a novel update routine of the
echo path response by utilizing the soft decision as follows:
⎧
⎪⎨
ˇĤ(i, k) =
p(H 1 |Y(i)) ˇĤ(i − 1,k)
+(1 − p(H 1 |Y(i)))Ĥopt(i, k)
, if I(Y(i)) = 1 and I(X(i)) = 1
⎪⎩
Ĥ opt (i, k), otherwise
(16)
where I(·) denotes an indicator function of the VAD result
provided by the IS-127 noise suppression algorithm since it
is known that it gives us a robust performance under various
noise environments [12]. Furthermore, we modified the
VAD algorithm to reduce the false decisions. For example,
I(Y(i)) = 1 if the near-end signal Y(i) exists at the ith frame
and I(Y(i)) = 0 otherwise. Therefore, the update of ˇĤ(i, k)
is finally addressed such that
ˇĤ(i, k) replaces ˇĤ(i − 1,k)
(i.e., no update) within the double-talk regions on each frequency
bin and (12) in the case of single-talk. In particular, in
the case of abrupt transient periods between double-talk and
single-talk, as shown in Fig. 2, the GNSPP could be a soft
value between 0 and 1. This accounts for why the soft decision
scheme is more insensitive to detection error compared
to conventional hard decision methods.
Based on this proposed DTD method, we finally apply it to
the AES algorithm proposed by Faller et al. [10] as follows:
Ŝ(i, k) =G(i, k)Y (i, k) (17)
where the Wiener filter gain G(i, k) is given by [10]
[ max(|Y (i, k)|−| Ŷ (i, k)|, 0)
]
G(i, k) =
. (18)
|Y (i, k)|
4. EXPERIMENTAL RESULTS
In order to verify the performance of the proposed DTD algorithm,
we conducted objective comparison experiments under
various noise conditions. Twenty test phrases, spoken by
seven speakers and sampled at 8 kHz, were used as the experimental
data. For assessing the performance of the proposed
method, we artificially created 20 data files, where
each file was produced by mixing the far-end signal with the
near-end signal. Each frame of the windowed signal was
transformed into its corresponding spectrum through a 128-
point DFT after zero padding. We then constructed 16 frequency
bands through combination of subbands to cover all
frequency ranges (∼4 kHz) of the narrow band speech signal,
5084

which is analogous to that of the IS-127 noise suppression algorithm
[12]. The far-end speech signal was passed through
a filter simulating the acoustic echo path modeled by a timeinvariant
FIR filter based on the analysis of room acoustics
before being mixed electrically [13], [14]. The simulation environment
was designed to fit a small office room having a
size of 5×4×3 m 3 . The echo level measured at the input microphone
was 3.5 dB lower than that of the input near-end
speech on average. In order to create noisy conditions, white,
babble and vehicular noises from the NOISEX-92 database
were added to clean near-end speech signals at signal-to-noise
ratios (SNRs) of 10, 20, and 30 dB. For the purpose of an objective
comparison, we evaluated the performance of the proposed
scheme and that of the conventional DTD algorithm
proposed by Park et al. [6], wherein the cross-correlation
coefficients-based double-talk detection method is used. The
performances of the approaches were measured in terms of
echo return loss enhancement (ERLE) and speech attenuation
(SA) during double-talk [14].
Given the three types of noise environments, the ERLE and
SAs scores were averaged to give final mean score results,
as presented in Table 1. From Table 1, it is evident that in
most noisy conditions, the proposed DTD algorithm based on
a soft decision yielded a lower ERLE compared to the hard
decision-based conventional technique. Also, from Table 2,
we can observe that the SAs (measured during double-talk periods)
of the proposed scheme based on a soft decision were
better than those of the previous method [6] in all the tested
conditions. It is noted that the performance gain of the proposed
method becomes smaller as the SNR becomes lower.
This is attributed to imperfection of the GNSPP under adverse
noise conditions. Summarizing the overall results, the
proposed approach is found to be effective in the AES technique.
5. CONCLUSIONS
In this paper, we have proposed a novel DTD algorithm
based on a soft decision scheme in the frequency domain.
The GNSPP based on a statistical model of the near-end and
far-end signal is applied to the DTD algorithm in conjunction
with VAD decisions for effective echo suppression. The
performance of the proposed algorithm has been found to be
superior to that of the conventional technique through objective
evaluation tests.
6. REFERENCES
[1] P. S. R. Diniz, Adaptive Filtering: Algorithm and Practical
Implementation. Norwell, MA: Kluwer, 1997.
[3] H. Ye and B. X. Wu, “A new double-talk detection algorithm
based on the orthogonality theorem,” IEEE Trans.
Commun., vol. 39, pp. 1542-1545, Nov. 1991.
[4] T. Gänsler, M. Hansson and C. J. Ivarsson, “A doubletalk
detector based on coherence,” IEEE Trans. Commun.,
vol. 44, pp. 1421-1427, Nov. 1996.
[5] J. Benesty, D. R. Morgan and J. H. Cho, “A new class of
doubletalk detectors based on cross-correlation,” IEEE
Trans. Speech Audio Processing, vol. 8, no. 10, pp. 168-
172, Mar. 2000.
[6] S. J. Park, C. G. Cho, C. Lee and D. H. Youn, “Integrated
echo and noise canceler for hands-free applications,”
IEEE Trans. on Circuits and Systems II, vol. 49,
issue 3, pp. 186-195, Mar. 2002.
[7] N. S. Kim and J.-H. Chang, “Spectral enhancement
based on global soft decision,” IEEE Signal Processing
Letters, vol. 7, no. 5, pp. 108-110, May 2000.
[8] Y.-S. Park and J.-H. Chang, “A novel approach to a robust
a priori SNR estimator in speech enhancement,”
IEICE Trans. on Communications, vol. E90-B, no. 8, pp.
2182-2185, Aug 2007.
[9] Y.-S. Park and J.-H. Chang, “A probabilistic combination
method of minimum statistics and soft decision for
robust noise power estimation in speech enhancement,”
IEEE Signal Processing Letters, vol. 15, no. 5, pp. 95-
98, Jan 2008.
[10] C. Faller and C. Tournery, “Robust echo control using
a simple echo path model,” in Proc. IEEE Int. Conf.
Acoust., Speech Signal Processing, vol. 5, pp. V281-
V284, 2006.
[11] Y. Ephraim and D. Malah, “Speech enhancement using
a minimum mean-square error short-time spectral amplitude
estimator,” IEEE Trans. Acoust., Speech, Signal
Process., vol. ASSP-32, no. 6, pp. 1109-1121, Dec.
1984.
[12] TIA/EIA/IS-127, “Enhanced variable rate codec, speech
service option 3 for wideband spread spectrum digital
systems,” 1996.
[13] S. McGovern, A Model for Room Acoustics, 2003 [Online].
Available: http://2pi.us/rir.html
[14] S. Y. Lee and N. S. Kim, “A statistical model based
residual echo suppression,” IEEE Signal Processing Letters,
vol. 14, no. 10, pp. 758-761, Oct. 2007.
[2] C. Avendano, “Acoustic echo suppression in the STFT
domain,” in Proc. IEEE Workshop on Appl. of Sig. Proc.
to Audio and Acoust., Oct. 2001.
5085