This article was downloaded by: [National Chiao Tung University 國立交通大學] On: 25 April 2014, At: 06:40
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK
Journal of the Chinese Institute of Engineers
Publication details, including instructions for authors and subscription information: http://www.tandfonline.com/loi/tcie20
Channel
‐optimized error mitigation for distributed
speech recognition over wireless networks
Cheng‐Lung Lee a & Wen‐Whei Chang b
a
Department of Communications Engineering , National Chiao‐Tung University , 1001 University Road, Hsinchu 300, Taiwan, R.O.C.
b Department of Communications Engineering , National Chiao‐Tung University , 1001
University Road, Hsinchu 300, Taiwan, R.O.C. Phone: 886–3–5731826 E-mail: Published online: 04 Mar 2011.
To cite this article: Cheng‐Lung Lee & Wen‐Whei Chang (2009) Channel‐optimized error mitigation for distributed speech recognition over wireless networks, Journal of the Chinese Institute of Engineers, 32:1, 45-51, DOI: 10.1080/02533839.2009.9671481
To link to this article: http://dx.doi.org/10.1080/02533839.2009.9671481
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and should be independently verified with primary sources of information. Taylor and Francis shall not be liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http:// www.tandfonline.com/page/terms-and-conditions
CHANNEL-OPTIMIZED ERROR MITIGATION FOR DISTRIBUTED
SPEECH RECOGNITION OVER WIRELESS NETWORKS
Cheng-Lung Lee and Wen-Whei Chang*
ABSTRACT
This paper investigates the error mitigation algorithms for distributed speech recognition over wireless channels. A MAP symbol decoding algorithm which ex-ploits the combined a priori information of source and channel is proposed. This is used in conjunction with a modified BCJR algorithm for decoding convolutional codes based on sectionalized code trellises. Performance is further enhanced by the use of the Gilbert channel model that more closely characterizes the statistical dependencies between channel bit errors. Experiments on Mandarin digit string recognition task indicate that our proposed mitigation scheme achieves high robustness against chan-nel errors.
Key Words: channel error mitigation, distributed speech recognition.
*Corresponding author. (Tel: 886-3-5731826; Email: wwchang @cc.nctu.edu.tw)
The authors are with the Department of Communications Engineering, National Chiao-Tung University, 1001 University Road, Hsinchu 300, Taiwan, R.O.C.
I. INTRODUCTION
The increasing use of mobile and IP networks for speech communication has lead to distributed speech recognition (DSR) systems being developed (ETSI ES 202 212 v1.1.1., 2003). The basic idea of DSR consists of using a local front-end from which speech features are extracted and transmitted through a data channel to a remote back-end recognizer. F or transmission, speech features are grouped into pairs and compressed via vector quantizers (VQs) in order to meet bandwidth requirements. The VQ encoder operates by mapping a large set of input vectors into a finite set of representative codevectors. The trans-mitter sends the index of the nearest codevector to the receiver, while the receiver decodes the codevector associated with the received index and uses it as an approximation of the input vector. Transmitting VQ data over noisy channels changes the encoded infor-mation and consequently leads to degraded recogni-tion performance. In the case of packet-erasure channels, several packet loss compensation techniques such as interpolation (Bernard and Alwan, 2002) and error control coding (Boulis et al., 2002) have been introduced for DSR. For wireless channels, joint
source-channel decoding (JSCD) techniques (Peinado et al., 2003; Reinhold and Valentin, 2004; Fingscheidt and Vary, 2001) have been shown effective for error mitigation using source residual redundancy assisted by bit reliability information provided by the soft-output channel decoder. However, the usefulness of these techniques may be restricted because they only exploit the bit-level source correlation on the basis of a memoryless AWGN channel assumption.
In this paper, we attempt to capitalize more fully on the a priori knowledge of source and channel and then develop a DSR system with increased robustness against channel errors. The first step toward realiza-tion is to use quantizer indexes rather than single index-bits as the bases for the JSCD, since the depen-dencies of quantizer indexes are stronger than the cor-relations of the index-bits. The next knowledge source to be exploited is the channel error characteristics. Transmission errors encountered in most real commu-nication channels exhibit various degrees of statistical dependency that are contingent on the transmission medium and on the particular modulation technique used. A typical example occurs in digital mobile radio channels, where speech parameters suffer severe deg-radation from error bursts due to the combined effects of fading and multipath propagation. A standard tech-nique for robust VQ over a channel with memory is to use interleaving to render the channel memoryless and then design a decoding algorithm for the memoryless channel. This approach, however, often introduces large
46 Journal of the Chinese Institute of Engineers, Vol. 32, No. 1 (2009)
decoding delays and does not utilize the channel memory information. Further improvement can be realized through a more precise characterization of the channel on which the decoder design is based (Kanal and Sastry, 1978). For this investigation, we focused on the two-state Markov chain model proposed by Gilbert (Gilbert, 1960). This model has several practical advantages over the Gaussian channel (Peinado et al., 2003) and binary Markov channels (Wang and Moayeri, 1993). First, the Gilbert model is relatively simple and can characterize a wide range of digital channels, as evi-denced by its applicability to performance analysis of various error control schemes (Drukarev and Yiu, 1986). Second, as we shall see later, the channel transition probabilities of the Gilbert model have a recursive formula that can be represented in terms of model parameters.
II. DSR TRANSMISSION SYSTEM The standard, ETSI ES 202 212, describes the speech processing, transmission, and quality aspects of a DSR system. The local front-end consists of a feature extraction algorithm and an encoding scheme for speech input to be transmitted to a remote recognizer. Each speech frame is represented by a 14-dimension feature vector containing log-energy logE and 13 Mel-frequency cepstral coefficients (MFCCs) ranging from C0 to C12. For the cepstral analysis speech signals
are sampled at 8 kHz and analyzed using a 25 ms Ham-ming window with 10 ms frame shift. These features are further compressed based on a split vector codebook where the set of 14 features is split into 7 subsets with two features in each. Each feature pair is quan-tized using its own codebook. MFCCs C1 to C10 are
quantized with 6 bits each pair, (C11, C12) is
quan-tized with 5 bits, and (C0, logE) is quantized with 8
bits. Two quantized frames are grouped together and protected by a 4-bit cyclic redundancy check creat-ing a 92-bit pair packet. Twelve of these frame-pairs are combined and appended with overhead bits resulting in an 1152-bit multiframe packet represent-ing 240 ms of speech. Multiframe packets are con-catenated into a bit-stream for transmission via a data channel with an overall data rate of 4800 bits/s.
This work is devoted to channel error mitiga-tion for DSR over burst error channels. Fig. 1 gives the block diagram of the transmission scheme for each
DSR feature pair. Suppose at time t, the input vector vt is quantized to obtain a codevector ct∈ {c(i), i = 0,
1, ..., 2k – 1} that, after bit mapping, is represented
by a k-bit combination ut = (ut(1), ut(2),..., ut(k)).
Each bit combination ut is assigned to a quantizer
index i ∈ {0,1, ..., 2k –1} and we write for simplicity
ut = uit to denote that ut represents the i-th quantizer
index. Due to constraints on coding complexity and delay, the VQ encoder exhibits considerable redun-dancy within the encoded index sequence, either in terms of a non-uniform distribution or in terms of correlation. If only the non-uniform distribution is considered and the indexes are assumed to be inde-pendent of each other, the redundancy is defined as the difference between the index length k and the entropy given by
H(ut) = –
Σ
P(ut)⋅log2P(ut)ut
. (1)
If inter-frame correlation of indexes is considered by using a first-order Markov model with transition prob-abilities P(ut|ut – 1), the redundancy is then defined as
the index length k and the conditional entropy given by H( ut ut – 1) = –
Σ
u t P(ut, ut – 1)⋅log2P( utut – 1)Σ
ut – 1 . (2) Table 1 shows the index lengths and entropies for the seven feature pairs of the ETSI DSR frond-end. For each column in Table 1, the probabilities P(ut) andP(ut|ut – 1) have to be estimated in advance from a
training speech database. From it we see that the DSR index sequence is better characterized by a first-or-der Markov process. For error protection individual index-bits are fed into a binary convolutional encoder consisting of M shift registers. The register shifts
Feature extraction Input signal Parameter estimation Index APP computation VQ encoder and Bit mapping Channel encoder
Source a priori knowledge Channel information Channel yt vt xt ut vt P[ut | y1 ] ^ T
Table 1 Entropies for DSR feature pairs
Parameter (ut) C1, C2 C3, C4 C5, C6 C7, C8 C9, C10 C11, C12 C0, logE
Bits/Codeword 6 6 6 6 6 5 8
H(ut) 5.75 5.71 5.68 5.80 5.82 4.85 7.33
H(ut|ut – 1) 3.17 3.42 3.85 4.14 4.25 3.64 3.46
Fig. 1 Transmission scheme for each DSR feature pair
one bit at a time and its state is determined by the M most recent inputs. After channel encoding, the code-bit combination corresponding to the quantizer index ut is denoted by xt = (xt(1), xt(2), ..., xt(n)) with the
code rate R = k/n.
One of the principal concerns in transmitting VQ data over noisy channels is that channel errors cor-rupt the bits that convey information about quantizer indexes. Assume that a channel’s input xt and output
yt differ by an error pattern et, so that the received bit
combination is yt = (yt(1), yt(2), ..., yt(n)) in which
yt(l) = xt(l) ⊕ et(l), l = 1, 2, ..., n, and ⊕ denotes the
bitwise modulo-2 addition. At the receiver side, the JSCD decoder will find the most probable transmit-ted quantizer index given the received sequence. The decoding process starts with the formation of an a posteriori probability (APP) for each of the possibly transmitted indices ut = i, which is followed by choosing
the index value ^
i that corresponds to the maximum a posteriori (MAP) probability for that quantizer index. Once the MAP estimate of the quantizer index is determined, its corresponding codevector becomes the decoded output ^
vt = c( ^
i ). The APP is the probability
that a decoded index ut = i can be derived from the
joint probability P(ui
t, st, yT1), where st is the channel
encoder state at time t and yT
1 =(y1, y2, ..., yT) is the
received sequence from time t = 1 through some time T. We have chosen the length T = 24 in compliance with the ETSI bit-streaming format, where each multiframe message packages speech features from 24 frames. Proceeding in this way, the symbol APP can be obtained by summing the joint probability over all encoder states, as follows:
P(ut= i y1T) = P(uti, s t, y1T) P(y1T)
Σ
st , i = 0, 1, ..., 2 k –1. (3) III. MODIFIED BCJR ALGORITHM Depending upon the choice of the symbol APP calculator, a number of different MAP decoder imple-mentations can be realized. For the transmissionscheme with channel coding, a soft-output channel decoder can be used to provide both decoded bits and their reliability information for further processing to improve the system error performance. The most well-known soft-output decoding algorithm is the BCJR algorithm (Bahl et al., 1974) that was devised to minimize the bit error probability. This algorithm is a trellis-based decoding algorithm for both linear block and convolutional codes. The derivation psented in Bahl et al. led to a forward-backward re-cursive computation on the basis of a bit-level code trellis. In a bit-level trellis diagram, there are two branches leaving each state and every branch repre-sents a single index-bit. Proper sectionalization of a bit-level trellis may result in useful trellis structural properties (Lin and Costello, 2004) and allow us to devise MAP decoding algorithms which exploit bit-level as well as symbol-bit-level source correlations. To advance with this, we propose a modified BCJR al-gorithm which parses the received code-bit sequence into blocks of length n and computes the APP for each quantizer index on a symbol-by-symbol basis. Un-like a conventional BCJR algorithm that decodes one bit at a time, our scheme proceeds with decoding the quantizer indexes in a frame as nonbinary symbols according to their index length k. By parsing the code-bit sequence into n-code-bit blocks, we are in essence merg-ing k stages of the original bit-level code trellis into one. As an example, we illustrate in Fig. 2 two stages of the bit-level trellis diagram of a rate 1/2 convolu-tional encoder with generator polynomial (5, 7)8. The
solid lines and dashed lines correspond to the input bits of 0 and 1, respectively. Fig. 2 also shows the decoding trellis diagram when two stages of the origi-nal bit-level trellis are merged together. In general, there are 2k branches leaving and entering each state
in a k-stage merged trellis diagram. Having defined the decoding trellis diagram as such, there will be one symbol APP corresponding to each branch which represents a particular quantizer index ut = i. For
convenience, we say that the sectionalized trellis dia-gram forms a finite-state machine defined by its state transition function S(ui
t, st) and output function X(uit,
Stage 1
(a) Bit-level trellis diagram Previous state Stage 2
00 01 10 11 Next state 00 01 10 11
(b) Merged trellis diagram Previous state 00 01 10 11 Next state 00 01 10 11
Fig. 2 Trellis diagrams used for (a) the encoder and (b) the MAP decoder
48 Journal of the Chinese Institute of Engineers, Vol. 32, No. 1 (2009)
st). Viewed from this perspective, the code-bit
com-bination xt = X(uit, st) is associated with the branch
from state st to state st + 1 = S(uit, st) if the
correspond-ing quantizer index at time t is ut = i.
We next modified the BCJR algorithm based on sectionalized trellis to exploit the combined a priori information of source and channel. We begin our de-velopment of the modified BCJR algorithm by rewrit-ing the joint probability in Eq. (3) as follows:
P(ui
t, st, yT1) = αit(st)βit(st), (4)
where αi
t(st) = P(uit, st, y1t) and βit(st) = P(yTt + 1|uit, st,
yt
1). For the MAP symbol decoding algorithm, the
forward and backward recursions are to compute the following metrics: αti(st) =
Σ
P(uti, st, ut – 1j , st – 1, yt, y1t – 1) jΣ
st – 1 =Σ
αt – 1j (st – 1)γi, j(yt, st, st – 1) jΣ
st – 1 , (5) βt i(s t) = P(ut + 1 j , s t + 1, yt + 1, yt + 2T uti, st, y1t)Σ
jΣ
st + 1 =Σ
βt + 1j (st + 1)γj, i(yt + 1, st + 1, st) jΣ
st + 1 , (6) in which γi, j(yt, st, st – 1) = P(ui t, st, yt|ujt – 1, st – 1, y1t – 1) = P(st|u j t – 1, st – 1, y1t – 1)P(uit|st, u j t – 1, st – 1, y1t – 1) . P(yt|uit, st, ujt – 1, st – 1, y1t – 1). (7)Having a proper representation of the branch metric
γi, j(yt, st, st – 1) is the critical step in applying MAP
symbol decoding to error mitigation and one that con-ditions all subsequent steps of the implementation. As a practical manner, several additional factors must be considered to take advantage of source correla-tion and channel memory. First, making use of the sectionalized structure of a decoding trellis, we write the first term in Eq. (7) as
P( st ut – 1j , s t – 1, y1t – 1) = P( stut – 1j , st – 1) = 1 , st= S(ut – 1 j , s t – 1) 0 , othertwise . (8)
The next knowledge source to be exploited is the re-sidual redundancy remaining in the DSR features. Assuming that the quantizer index is modelled as a first-order Markov process with transition probabilities P(ut|ut – 1), the second term in Eq. (7) is reduced to
P(ui t|st, u
j
t – 1, st – 1, y1t – 1) = P(ut = i|ut – 1 = j). (9)
In addition to source a priori knowledge, specific knowledge about the channel memory must be taken into consideration. There are many models describ-ing the correlation of bit error sequences. If no chan-nel memory information is considered, which means that the channel bit errors are assumed to be random, the third term in Eq. (7) is reduced to
P(yt|uit, st, ujt – 1, st – 1, y1t – 1)
= P(yt|xt = X(uti, st))
= P(et) = εl(1 – ε)n – 1, (10)
where ε is the channel bit error rate (BER) and l is the number of ones occurring in the error pattern et.
When intraframe and interframe memory of the chan-nel are considered, the third term in Eq. (7) becomes
P(yt|uit, st, ujt – 1, st – 1, y1t – 1)
= P(yt|xt = X(uti, st), yt – 1, xt – 1 = X(ujt – 1, st – 1))
= P(et|et – 1). (11)
IV. PROBABILITY RECURSIONS FOR GILBERT CHANNEL
Designing a robust DSR system requires that pa-rameterized probabilistic models be used to summa-rize some of the most relevant aspects of error statistics. It is apparent from previous work on chan-nel modelling (Kanal and Sastry, 1978) that we are confronted with contrasting requirements in select-ing a good model. A model should be representative enough to describe real channel behavior and yet it should not be analytically complicated. To permit theoretical analysis, we assumed that the encoded bits of DSR features were subjected to the sample error sequences typical of the Gilbert channel. The Gil-bert channel model consists of a Markov chain hav-ing an error-free state G and a bad state B, in which errors occur with the probability (1 – h). The state transition probabilities are b and g for the G to B and B to G transitions, respectively. The model state-tran-sition diagram is shown in Fig. 3. The effective BER produced by the Gilbert channel is ε = (1 – h)b/(g + b). Notice that in the particular case of a Gilbert model with parameter values b = 1, g = 0, h = 1 –ε,
1-b G B 1-g
b
g
Fig. 3 Gilbert channel model
the channel model reduces to a memoryless binary symmetric channel with the BER ε.
The effectiveness of the MAP symbol decoding depends crucially on how well the error characteris-tics are incorporated into the calculation of channel transition probabilities P(et|et – 1). Although using
channel memory information was previously proposed for MAP symbol decoding (Turin, 2001), the empha-sis was placed upon channels with no interframe memory. When only access to the intraframe memory is available, it was shown that the channel transition probabilities of the Gilbert channel have closed-form expressions that can be represented in terms of model parameters {h, b, g}. Under such conditions, we can proceed with the MAP symbol decoding in a manner similar to the work of (Turin, 2001). Extensions of these results to channels with both intraframe and interframe memory have been found difficult. Rec-ognizing this, we next develop a general treatment of probability recursions for the Gilbert channel. The main result is a recursive implementation of MAP symbol decoder being closer to the optimal for chan-nels with memory. For notational convenience, channel bit error et(l) will be denoted as rm, in which the bit
time m is related to the frame time t as m = n(t – 1) + l, l = 1, 2, ..., n. Let qm∈ {G, B} denote the Gilbert
channel state at bit time m. The memory of the Gil-bert channel is due to the Markov structure of the state transitions, which lead to a dependence of the cur-rent channel state qm on the previous state qm – 1.
To develop a recursive algorithm, it is more con-venient to rewrite the channel transition probabili-ties as P( et et – 1) = P(rm= 1 rm 0 m – 1)rmP(r m= 0 rm0 m – 1)1 – rm
Π
m = n(t – 1) + 1 nt , (12) where rm 0 m – 1 = (rm0, rm0 + 1, ..., rm – 1) represents the bit error sequence starting from bit m0 = n(t – 2) + 1.
The following is devoted to a way of recursively com-puting P(rm = 1|rm0
m – 1) from P(r
m – 1 = 1|rm0
m – 2). The
Gilbert channel has two properties, P(qm|qm – 1, rm0
m – 1) = P(qm|qm – 1) and P(rm|qm, rm0
m – 1) = P(r
m|qm), which
facilitate the probability recursions. By successively applying the Bayes rule and the Markovian property of the channel, we have
P( rm= 1 rm 0 m – 1) = P( rm= 1 qm= B , rm 0 m – 1)P( q m= B rm0 m – 1) = (1 – h)P( qm= B rm 0 m – 1) , (13) in which P( qm= B rm 0 m – 1) = P( qm= B qm – 1= G, rm0 m – 1 )P( qm – 1= G rm0 m – 1 ) + P( qm= B qm – 1= B, rm0 m – 1 )P( qm – 1= B rm0 m – 1 ) = bP( qm – 1= G, rm – 1 rm0 m – 2) P( rm – 1 rm 0 m – 2) + (1 – g)P( qm – 1= B, rm – 1 rm0 m – 2) P( rm – 1 rm 0 m – 2) = b + (1 – g – b)P( qm – 1= B, rm – 1 rm0 m – 2) P( rm – 1 rm0 m – 2 ) = b + (1 – g – b)P( rm – 1 qm – 1= B) P( rm – 1 rm 0 m – 2) ⋅P( rm – 1= 1 rm0 m – 2) 1 – h . (14) V. EXPERIMENTAL RESULTS
Computer simulations were conducted to evaluate three MAP-based error mitigation schemes for DSR over burst error channels. First a bit-level trellis MAP decoding scheme BMAP is considered that uses the standard BCJR algorithm to decode the index-bits. The decoders SMAP1 and SMAP2 exploit the sym-bol-level source redundancy by using a modified BCJR algorithm based on a sectionalized trellis structure. The SMAP1 is designed for a memoryless binary sym-metric channel, whereas the SMAP2 exploits the channel memory though the Gilbert channel characterization. The channel transition probabilities to be used for the SMAP1 is P(et) in Eq. (10), and P(et|et – 1) in Eq. (11)
for the SMAP2. For purposes of comparison, we also investigated an error mitigation scheme (Peinado et al., 2003) which applied the concept of softbit speech decoding (SBSD) and achieved good recognition per-formance for AWGN and burst channels. A prelimi-nary experiment was first performed to evaluate various decoders for reconstruction of the feature pair (C0,
logE) encoded with the DSR front-end. A rate R = 1/ 2 convolutional code with memory order M = 6 and the octal generator (46,72)8 is chosen as the channel
code. Table 2 presents the signal-to-noise ratio (SNR) obtained from transmission of the index-bits over Gilbert channel with BER ranging from 10–3 to 10–1. The
results of these experiments clearly demonstrate the improved performance achievable using the SMAP1 and SMAP2 in comparison to those of BMAP and SBSD. Furthermore, the improvement has a tendency to increase for noisy channels with higher BER. This indicates that the residual redundancy of quantizer indexes is better exploited at the symbol level to achieve more performance improvement. A comparison of
50 Journal of the Chinese Institute of Engineers, Vol. 32, No. 1 (2009)
the SMAP1 and SMAP2 also revealed the importance of matching the real error characteristics to the chan-nel model on which the MAP symbol decoder design is based. The better performance of SMAP2 can be attributed to its ability to compute the symbol APP taking interframe and intraframe memory of the channel into consideration, as opposed to the memoryless chan-nel assumption made in the SMAP1.
We further validate the proposed decoding algorithms for the case where error sequences were generated using a complete GSM simulation. The simu-lator is based on the CoCentric GSM library (CoCentric System Studio-Referenec Design Kits, 2003) with TCH/ F4.8 data and channel coding, interleaving, modulation, a channel model, and equalization. The channel model represents a typical case of a rural area with 6 propa-gation paths and a user speed of 50 km/h. Further, cochannel interference was simulated at various car-rier-to-interference ratios (CIR). In using the SMAP1 and SMAP2 schemes, the channel transition probabilities have to be combined with a priori knowledge of Gil-bert model parameters which can be estimated once in advance using the gradient iterative method (Chouinard et al., 1988). For each simulated error sequence, we first measured the error-gap distribution P(0l|1) by computing the probability that at least l
suc-cessive error-free bits will be encountered next on the condition that an error bit has just occurred. The op-timal identification of Gilbert model parameters was then formulated as the least square approximation of the measured error-gap distribution by exponential curve fitting. Table 3 gives estimated Gilbert model parameters for the GSM TCH/F4.8 data channels operating at CIR = 1, 4, 7, 10 dB. The next step in the present investi-gation concerned the performance degradation that may result from using the SMAP2 scheme under channel mismatch conditions. In Table 4, CIRd refers to the
CIR value assumed in the design process, and CIRa
refers to the true CIR used for the evaluation. The best results are in the main diagonal of the table, where channel-matched Gilbert model parameters are used for the channel transition probability computation of Eq. (12). The performance decreases in each column below the main diagonal when the CIRd is increased.
The investigation further showed that the SMAP2 is not very sensitive to a channel mismatch between the design and evaluation assumptions.
We next considered the speaker-independent rec-ognition of Mandarin digit strings as the task without restricting the string length. A Mandarin digit string database recorded by 50 male and 50 female speakers was used in our experiments. Each speaker pronounced 10 utterances and 1-9 digits in each utterance. The speech of 90 speakers (45 male and 45 female) was used as the training data, and the other 10 as test data. The total numbers of digits included in the training and test data were 6796 and 642, respectively. The reference recog-nizer is based on the HTK software package. A 38-dimension feature vector used in the recognizer con-sisted of 12 MFCCs, delta-MFCC, delta-delta-MFCC, delta-energy and delta-delta-energy. The digits were modelled as whole word Hidden Markov Models (HMMs) with 8 states per word and 64 mixtures for each state. In addition, a 3-state HMM was used to model pauses before and after the utterance and a one-state HMM was used to model pauses between digits. The DSR results obtained by various error mitigation algo-rithms for the Gilbert channel are shown in Fig. 4. It Table 2 SNR(dB) performance for various
decod-ers on a Gilbert channel
BER BMAP SBSD SMAP1 SMAP2 0.001 26.84 26.88 26.93 27.51 0.0025 26.37 26.51 26.56 27.10 0.0063 25.21 25.83 25.91 26.41 0.0158 22.30 22.71 23.31 25.13 0.0398 17.51 20.67 21.13 24.67 0.1 14.12 16.88 18.52 23.94
Table 3 Estimated Gilbert model parameters for GSM TCH/F4.8 data channels
CIR(dB) 1 4 7 10
g 0.001 0.01 0.02 0.05
b 0.0197 0.0034 0.0022 0.0034
h 0.7528 0.6086 0.7511 0.9403
Table 4 SNR performance of the SMAP2 over the GSM data channel under channel mis-match conditions
CIRa = 1 CIRa = 4 CIRa = 7 CIRa = 10
CIRd = 1 11.86 16.68 27.02 30.25
CIRd = 4 11.62 16.78 27.19 30.40
CIRd = 7 11.51 16.72 27.24 30.41
CIRd = 10 11.31 16.32 27.01 30.64
Fig. 4 Recognition performances for DSR transmission over a Gilbert channel 100 95 90 85 80 75 70 65 60 55 -3 -2.8 -2.6 -2.4 -2.2 -2 BER (dB) -1.8 -1.6 -1.4 -1.2 -1 SMAP2 SMAP1 SBSD BMAP Digit Accuracy (%)
can be seen that employing the source a priori infor-mation, sectionalized trellis MAP decoding, and chan-nel memory constantly improves the recognition accuracy. The SMAP2 scheme performs the best in all cases, showing the importance of combining the a priori knowledge of source and channel by means of a section-alized code trellis and Gilbert channel characterization.
VI. CONCLUSIONS
A JSCD scheme which exploits the combined source and channel statistics as an a priori informa-tion is proposed and applied to channel error mitiga-tion in DSR applicamitiga-tions. We first investigate the residual redundancies existing in the DSR features and find ways to exploit these redundancies in the MAP symbol decoding process. Also proposed is a modified BCJR algorithm based on sectionalized code trellises which uses Gilbert channel characterization for better decoding in addition to source a priori knowledge. Experiments on Mandarin digit string recognition indicate that the proposed decoder achieved significant improvements in recognition accuracy for DSR over burst error channels.
ACKNOWLEDGEMENT
This study was jointly supported by MediaTek Inc. and the National Science Council, Republic of China, under contract NSC 95-2221-E-009-078.
NOMENCLATURE
b the probability for the state transition G to B
B the bad state of the Gilbert channel
ct the codeword of the VQ at time t
et the bit-error pattern at time t
g the probability for the state transition B to G
G the good state of the Gilbert channel
k the length of ut
st the state of the channel encoder
ut the binary index representing ct
vt the source vector
xt the code-bit combination after channel encoding
yt the received bit combination at the receiver
ε bit error rate of the noisy channel
REFERENCES
Bahl, L. R., Cocke, J., Jelinek, F., and Raviv, J., 1974, “Optimal Decoding of Linear Codes for Minimizing Symbol Error Rate,” IEEE Transactions on In-formation Theory, Vol. IT-20, No. 2, pp. 284-287. Bernard, A., and Alwan, A., 2002, “Low-bitrate Distrib-uted Speech Recognition for Packet-based and Wire-less Communication,” IEEE Transactions on Speech
and Audio Processing, Vol. 10, No. 8, pp. 570-579. Boulis, C., Ostendorf, M., Riskin, E., and Otterson, S., 2002, “Graceful Degradation of Speech Rec-o g n i t i Rec-o n P e r f Rec-o r m a n c e Rec-o v e r P a c k e t - e r a s u r e Networks,” IEEE Transactions on Speech and Au-dio Processing, Vol. 10, No. 8, pp. 580-590. Chouinard, J. Y., Lecours M., and Delisle, G. Y., 1988,
“Estimation of Gilbert’s and Fritchman’s Models Parameters Using the Gradient Method for Digital Mobile Radio Channels,” IEEE Transactions on Vehicular Technology, Vol. 37, No. 3, pp.158-166. CoCentric System Studio-Referenec Design Kits, 2003, Synopsys, Inc., Mountain View, CA, USA. Drukarev, A. I., and Yiu, K. P., 1986, “Performance of Error-correcting Codes on Channels with Memory,” IEEE Transactions on Communications, Vol. COM-34, No. 6, pp. 513-521.
ETSI ES 202 212 v1.1.1. Digital Speech Recognition; Extended Advanced Front-end Feature Extraction Algorithm; Compression Algorithms; Back-end Speech Reconstruction Algorithm. November 2003. Fingscheidt, T., and Vary, P., 2001, “Softbit Speech Decoding: a New Approach to Error Conceal-ment,” IEEE Transaction on Speech and Audio Processing, Vol. 9, No. 3, pp. 240-251.
Gilbert, E. N., 1960, “Capacity of a Burst-noise Channel,” The Bell System Technical Journal, Vol. 39, No. 1, pp. 1253-1265.
Kanal, L. N., and Sastry, A. R. K., 1978, “Models for Channels with Memory and their Applications to Error Control,” Proceedings of IEEE, Vol. 66, No. 7, pp. 724-744.
Lin, S., and Costello, D. J., 2004, Error Control Coding, Prentice Hall, NJ, USA.
Peinado, A. M., Sanchez, V., Perez-Cordoba, J. L., and Torre, A., 2003, “HMM-based Channel Er-ror Mitigation and its Application to Distributed Speech Recognition,” Speech Communication, Vol. 41, No. 2, pp. 549-561.
Reinhold, H. U., and Valentin, I., 2004, “Soft Fea-tures for Improved Distributed Speech Recogni-tion over Wireless Networks,” Proceedings of 8th
International Conference on Spoken Language Processing, Jeju Island, Korea, pp. 2125-2128. Turin, W., 2001, “MAP Symbol Decoding in
Chan-nels with Error Bursts,” IEEE Transactions on Information Theory, Vol. 47, No. 5, pp. 1832-1838. Wang, H. S., and Moayeri, N., 1993, “Modeling, Capacity, and Joint Source/Channel Coding for Rayleigh Fading Channels,” Proceedings of IEEE Vehicular Technology Conference, Secaucus, NJ, USA, pp. 473-479.
Manuscript Received: Oct. 23, 2007 Revision Received: June 30, 2008 and Accepted: July 30, 2008
