A Two-stage Prosodic Structure Generation Strategy

for Mandarin Text-to-speech Systems

DONG Yuan1

ZHOU Tao1

DONG Cheng-Yu2

WANG Hai-La2

Abstract Prosodic structure generation is the key component in improving the intelligibility and naturalness of synthetic speech for
a text-to-speech (TTS) system. This paper investigates the problem of automatic segmentation of prosodic word and prosodic phrase,
which are two fundamental layers in the hierarchical prosodic structure of Mandarin, and presents a two-stage prosodic structure
generation strategy. Conditional random elds (CRF) models are built for both prosodic word and prosodic phrase prediction at
the front end with dierent feature selections. Besides, a transformation-based error-driven learning (TBL) modication module is
introduced in the back end to amend the initial prediction. Experiment results show that the approach combining CRF and TBL
achieves an F-score of 94.66 %.
Key words Text-to-speech (TTS), prosodic structure generation, conditional random elds (CRF), transformation-based errordriven learning (TBL)

Text-to-speech is an important technique to generate the

articial speech from text dependent application[1] . This
technique has been widely applied in many elds, such as
telecommunication services, embedded mobile application
and entertainment.
When people talk, they rarely speak out a whole sentence without a break. Instead, an utterance is divided into
smaller units with perceivable boundaries between them.
These phonetic spurts or chunks of speech, which are commonly known as prosodic units, signal the internal structure of the message and serve important functions in helping people to clearly express their ideas as well as their
feelings. These units are dierent in their categories, levels, and boundary strength. Smaller- and lower-level units
are contained in larger- and higher-level units to form a
prosodic hierarchy. In Mandarin, this hierarchical structure is often simplied to three layers (from bottom to up):
prosodic word, prosodic phrase, and intonation phrase[2] .
In a current text-to-speech (TTS) system, the input
text is rstly processed by a text analysis model, whose
output is a string of syntactic words, each with a partof-speech (POS) tagging. Then, its prosodic structure is
expected to be constructed without linguistic information
in order to enhance the naturalness and understandability
of the synthetic speech[3] . However, as the grammatical
structure does not necessarily correspond to its prosodic
counterpart, misjudgments often occur in assigning proper
prosodic boundary, which has become the major impediment for TTS systems to achieve human-like performance.
That is why the prediction of prosodic structure is an important step for a TTS system. In fact, uent spoken language is never produced in a smooth and unvarying stream.
Furthermore, variation in phrasing can change the meaning
by the utterances of a given sentence.
Researches have shown that the relative size and location of prosodic boundaries provide an important cue for
resolving syntactic ambiguity. More and more attention has
been paid to addressing the problem of automatic prosodic
boundary prediction.
Researches have discovered a hierarchical prosodic strucManuscript received March 25, 2009; accepted July 1, 2010
Supported by National Natural Science Foundation of China
(90920001), the Key Project of the Ministry of Education of China
(108012), and Joint-research Project between France Telecom R&D
Beijing and Beijing University of Posts and Telecommunications
(SEV01100474)
1. Beijing University of Posts and Telecommunications, Beijing
100876, P. R. China 2. France Telecom R&D (Beijing), Beijing
100190, P. R. China
DOI: 10.1016/S1874-1029(09)60063-X

ture for Chinese prosody, which constitutes the rhythm

of Chinese speech. Researchers also found that the main
prosodic elements in Chinese speech are prosodic word,
prosodic phrase, and intonation phrase.
In the earlier time, rule-based methods were usually
adopted[4] . It mainly started from Gee and Grosjean s work
on performance structures, and has had various extensions
over the years[5] . For Chinese, similar investigation has also
been carried out into the formation of prosodic constitutes,
such as those reported by Cao and Wang[67] . The central
idea of all these works is to nd out some explicit rules
that are able to recreate the prosodic structure of a sentence from syntax through a large number of experiments
and empirical observations. This method is easily explicable and understandable, but also poses strict demand for
the system developer to summarize these rules. Moreover,
it is hard to update and improve, and the set of rules is usually constrained to one branch of language, which hinders
its general application.
With the availability of increasing prosodically annotated corpora and the rapid development of machine learning, stochastic-based approach has been more and more
popular in prosodic boundary prediction[8] . As in most
cases, it is assumed that syntactic word is the smallest unit
(i.e., leaf node) in a prosodic hierarchy tree, the task of
building prosodic structure could be reduced to decide the
type for each syntactic word boundary, which is actually
a classication problem. Thus, many dierent statistical
methods used for classication have been tried, such as
classication and regression tree (CART) used by Wang et
al.[9] , and hidden Markov model proposed by Paul et al.[10] .
Researchers have also begun to adopt this approach in Chinese during recent years. Besides those mentioned above,
attempt was made to predict prosody phrase break based
on the maximum entropy model[1112] . Generally, these
methods relate each boundary site with some features (e.g.,
length and POS of adjacent words). By extracting and absorbing these features from a large collection of annotated
sentences, a statistical model is trained and then applied
to unlabeled texts. For each potential boundary site in the
text, a probability is estimated for each possible outcome,
and the one with the largest likelihood is determined as the
correct type.
This paper proposes to predict prosodic boundaries
by using a two-stage strategy: conditional random elds
(CRF) models and a transformation-based error-driven le-

1570

ACTA AUTOMATICA SINICA

arning (TBL) modication module[1314] . The strategy is

applied to both prosodic words and prosodic phrase boundary labeling.
The rest of the paper is organized as follows. Section 1
introduces the prosodic structure briey. The CRF algorithm and corresponding feature selection are presented
in Section 2. Section 3 describes the TBL modication
module[15] . Experiments and performance results are given
in Section 4. Section 5 draws the conclusion.

complished with a conditional random elds model.

Prosodic structure

Generally, the prosodic hierarchy includes three tiers,

which are prosodic word (PW), prosodic phrase (PP), and
intonation phrase. Prosodic word is the most elementary
rhythmic unit among these three tiers. In real speech,
prosodic word should be uttered continuously and closely
without breaks. Prosodic phrase and intonation phrase are
the combination of several PWs. Thus, the segmentation
of prosodic word will aect the segmentation of prosodic
phrase and intonation phrase, and will also play an important role in increasing the naturalness of synthesized
speech. The segmentation module is able to make boundaries of lexicon words (LW). However, the boundaries of
lexicon words are quite dierent from prosodic structure.
Prosodic word is the basic element of prosodic boundary,
which has three types of relationship with lexicon word:
1) A prosodic word is a lexicon word; 2) A prosodic word
is combination of several short lexicon words; 3) A prosodic
word is part of a long lexicon word. Most of the cases belong to the rst two types. The rst step of our work is to
make a proper prosodic word boundary from the output of
segmentation and POS module.
Prosodic phrase can be treated as the combination of
several prosodic words. Prosodic phrase usually contains
2 3 prosodic words, even one single prosodic word, with a
length of 3 9 characters. Prosodic phrase plays the most
important role in prosodic boundary prediction, as it contains most information of human pronunciation attribute.
So, the F-score of prosodic phrase takes the majority work
in determining the naturalness of voice synthesis in TTS
system. In this paper, prosodic phrase boundaries are generated based on the output of the prosodic word prediction.
Intonation phrase can mostly be indicated by the punctuations, which can be easily detected. Punctuations can
be regularly marked as a sign of prosodic boundaries. This
paper does not build a single CRF model for intonation
phrase, but denes it as an intonation phrase by the punctuations.
The framework of the two-stage strategy is shown in
Fig. 1.

Vol. 36

Conditional random elds model

By using a probability map model, CRF expresses the

ability of a long-range dependence and overlapping. CRF
has the advantage to solve the problem of bias and normalize the overall situation in order to nd the optimal solution. Two distinct CRF models are built by incorporating
dierent features for prosodic word and prosodic phrase
prediction, respectively.
Given a string of consecutive syntactic words, for each
boundary between two of them (say wi and wi+1 ), there
are three types: LW (wi and wi+1 are within the same
prosodic word), PW (within the same prosodic phase but
dierent prosodic words) and PP (within dierent prosodic
phrases). Then, our task comes down to decide the right
type for each syntactic word boundary, which could be ac-

Fig. 1

2.1

Flow of prosodic structure prediction

CRF algorithm modeling

Consider probability distributions over sets of random

variables V = X Y , where X is a set of input variables
that are assumed to be observed, and Y is a set of output
variables that we wish to predict. Every variable v V
takes outcomes from a set , which can be either continuous or discrete although we discuss only the discrete case in
this section. We denote an assignment to X by x, and we
denote an assignment to a set A X by A , and similarly
for Y . We use the notation 1{x=x } to denote an indicator
function of x which takes the value 1 when x = x and 0
otherwise.
A graphical model is a family of probability distributions
that factorize according to an underlying graph. The main
idea is to represent a distribution over a large number of
random variables by a product of local functions each of
which depends on only a small number of variables.

A (xA , yA )
Z=
x,y

The constant Z is a normalization factor dened as

above, which ensures that the distributions sum to 1. The
quantity Z, considered as a function of the set F of factors,
is called the partition function in the statistical physics and
graphical models communities. Computing Z is intractable
in general, but much work exists on how to approximate it.
Graphically, we represent the rst factorization by a factor
graph. A factor graph is a bipartite graph G = (V, F, E),
in which a variable node s V is connected to a factor
node A F if s is an argument to A .
A directed graphical model, also known as a Bayesian
network, is based on a directed graph G = (V, E). A directed model is a family of distributions that factorize as

p(|())
p(y, x) =
V

where () are the parents of in G.

We use the term generative model to refer to a directed
graphical model in which the outputs topologically precede

No. 11

DONG Yuan et al.: A Two-stage Prosodic Structure Generation Strategy for

the inputs, that is, no x X can be a parent of an out put

y Y . Essentially, a generative model is one that directly
describes how the outputs probabilistically generate the
inputs.
2.2

Feature selection

Feature selection is an important part in the research of

machine learning. Good data can make a proper supplement to the corresponding algorithm. In this paper, the
feature template is manually designed.
For specic application, most commonly used features
include POS (part of speech tagging), WLen (length in syllables), and Word (the word itself) of the words surrounding the boundary, which have also proved to be the most
important determinants of prosodic boundary types. On
account of this, we add them into both templates, with a
window length of 2 for POS, i.e., we consider the POS of 2
words immediately before and after the boundary in question, and a window length of 1 for WLen and Word. A point
to note, though, is that word has dierent meanings under the two scenarios. For prosodic word, it indicates syntactic word and is readily available from the input text. For
prosodic phrase, which is built upon prosodic words rather
than syntactic words, the meaning accordingly changes to
prosodic word. Here, POS property of a prosodic word
is acquired by simply concatenating POS s of the syntactic
words it contains.
Another feature LastType is also introduced into the
templates, which denotes the last prosodic boundary type.
The motive for adding this information came from the observation that current boundary type is inuenced by that
of the last one, which applies to prosodic word as well as
prosodic phrase boundary. For example, a LastType of
PP could well reduce the possibility that current boundary
is still PP.
So, the feature selection in this paper is mainly focusing on three facts: POS, WLen, and Word. Based on the
analysis of the result to the segmentation and POS tagging,
six types are generated for both prosodic words model and
prosodic phrases model. These six types are presented in
Table 1.
Table 1

Types of features

Feature

Meaning

LastType

The type of last prosodic boundary

Word

The current word

WLen

The length of the current word

POS

The POS information of the current word

Word&&POS

The POS information and current word

WLen&&POS

The word length and POS information

Besides, two more types, DistPre and DistNxt, are specially designed for the prosodic phrase model. To some
extent, insertion of prosodic phrase boundaries in natural
spoken language is to balance the length of the constituents
in the output. Hence, it is not surprising that most PP
breaks occur in the middle part of a long sentence, and a
prosodic phrase is usually 5 7 syllables long, but rarely
shorter than 3 or longer than 9 syllables. For this reason,
we took into consideration length measures by including
DistPre and DistNxt in our templates for prosodic phrase
prediction, which means the distance from current boundary to the last and next nearest PP location.
For some type, we make an extension in the range from
previous two words to the following two words. Take feature POS for example, ve features are designed, as shown

1571

in Table 2.
What is more, the certain combination of the two types
can hold some special information for prosodic boundaries. Some features are generated in this way as shown
in Table 3.
Table 2

The extension of features

Feature

Meaning

POS2

The POS of the word before last word

POS1

The POS of last word

POS

The POS of current word

POS+1

The POS of next word

POS+2

The POS of the word after next word

Table 3

The combination of features

Feature

Meaning

Word2POS2

The word before last word and its POS

Word1POS1

The last word and its POS

Word0POS0

The current word and its POS

Word+1POS+1

The next word and its POS

The above four features are separately standing for: the

word before the last one and the POS of the word before
the last one; the last word and the POS of the last word;
the current word and the POS of the current word; the next
word and its POS.
Altogether, 34 features are selected for prosodic word
model and 38 features are designed for the prosodic phrase
model.

Transformation-based error-driven
learning modication module

In our preliminary experiment using only CRF model

for prediction, we nd that there are always some obvious
mistakes that humans would never commit as they obviously contradict to some accustomed xed patterns. It
then happened to us that these mistakes might be corrected
by rules. The transformation-based error-driven learning
modication module is added at the back-end of the system to amend the mistakes left by CRF model.
3.1 The design of TBL modication module
Transformation-based error-driven learning, commonly
referred to as transformation-based learning or TBL, is
an automatic machine learning technique. The output of
TBL is an ordered list of rules, whose application to data
can result in a reduction of error.
TBL algorithm starts from an initial state, and by use
of a series of transformation rules, it modies the result bit
by bit to achieve the best score according to the objective
function used.
Fig. 2 illustrates how the transformation-based errordriven learning works. First, unannotated text is passed
through an initial-state annotator. The initial-state annotator can range in complexity from assigning random structure to assigning the output of a sophisticated manually
created annotator. For syntactic parsing, we have explored
initial state annotations ranging from the output of a sophisticated parser to random tree structure with random
nonterminal labels.
Once the text passes through the initial-state annotator,
it is compared to the truth. A manually annotated corpus
is used as our reference for truth. An ordered list of transformations is learned that can be applied to the output of

1572

ACTA AUTOMATICA SINICA

the initial-state annotator to make it better resemble the

truth. There are two components to a transformation: a
rewrite rule and a triggering environment.

corpus is rst processed by the corpus preprocessor component, which extracts the features used for prosodic
structure prediction. The second component is the TBL
trainer component, which learns a set of rules by the
transformation-based error-driven learning algorithm. This
component also needs the prepared rule template as the input. The raw rule template is stored in a le and parsed
by the template parser component.
Testing process applies the rules generated by TBL
trainer for prosodic structure prediction. The features
used for prosodic structure prediction are word, part-ofspeech, character, prosodic word boundary, and prosodic
phrase boundary. In real application, the information can
be gotten from word segmenter, part-of-speech tagger, and
pinyin-lookup module. In our package, the testing corpus
is extracted from the annotated corpus and rst processed
by the corpus preprocessor component, which extracts the
features used for prosodic structure prediction from the annotated corpus. Then, we apply all the rules learned from
the TBL training system to prosodic structure prediction.
3.3

Fig. 2

Framework of TBL

In all of the applications, we have examined to date the

following greedy search is applied for deriving a list of transformations: At each iteration of learning, the transformation is found whose application results in the best score
according to the objective function being used; that transformation is then added to the ordered transformation list
and the training corpus is updated by applying the learned
transformation. Learning continues until no transformation can be found whose application results in an improvement to the annotated corpus. Other more sophisticated
search techniques could be used, such as simulated annealing or learning with a look-ahead window, but we have not
yet explored these alternatives.
3.2

Applying TBL to prosodic structure prediction

This paper applies TBL to the prosodic structure prediction of Mandarin TTS system.
According to the counting result of the performance with
CRF model, this paper presents 20 Chinese characters to
TBL learning for further disambiguation. These 20 Chinese characters occur with most mistakes frequently in the
test corpus, such as , , and .
For each character, we prepare two sets of corpus, one
for training and another for testing. The two corpora are
extracted from the annotated People s Daily Corpus 2000,
which was annotated manually by the Institute of Computational Linguistics.
The whole TBL modication module contains two processes: the training process and the testing process.
The training process accepts the annotated corpus. The
Table 5

Vol. 36

Feature selection and template design

The previous research shows that the following features

are useful for prosodic structure prediction: character and
word feature: LC (lexicon character), LW (lexicon word);
syntax feature: POS (part-of-speech); other features: POSITION (the position in the syntax word), LEN (the length
of each word), BEGIN (if in the begin of the sentence). Besides, we design other features such as PW (prosodic word
boundary) and PP (prosodic phrase boundary) for prosodic
structure prediction.
The current system makes use of the features of character (LC), word (LW), prosodic word boundary (PW) and
prosodic phrase boundary (PP), and part-of-speech (POS).
In order to investigate the disambiguation ability for a
dierent feature set, 18 dierent kinds of templates are designed with performance in Table 4.
Table 4

Templates for TBL learning

Rule template (P)

Type

Rule template (P)

Type

R0 (0.921 775)

T0

R9 (0.910 721)

T4+T0

R1 (0.869 216)

T1

R10 (0.910 432)

T5+T0

R2 (0.865 277)

T2

R11 (0.930 861)

T1,T0

R3 (0.868 538)

T3

R12 (0.932 088)

T2,T0

R4 (0.862 322)

T4

R13 (0.927 600)

T3,T0

R5 (0.859 158)

T5

R14 (0.931 081)

T4,T0

R6 (0.908 081)

T1+T0

R15 (0.919 302)

T5,T0

R7 (0.910 238)

T2+T0

R16 (0.929 745)

T1+T0,T0

R8 (0.907 211)

T3+T0

R17 (0.931 252)

T2+T0,T0

Each rule template consists of the atom templates: T0,

T1, T2, T3, T4, and T5. The atom templates are described
as Table 5.

Atom templates

Atom

Type

T0

POS itself

Contents

T1

Char

Char n (n = 3, 2, 1, 1, 2, 3)

Char n&Char n 1

Char n&Char n 1&Char n 2

T2

Word

Word n

Word n&Word n 1

Word n&Word n 1&Word n 2

PW n&PW n 1&PW n 2

POS 0

T3

Prosodic word

PW n

PW n&PW n 1

T4

Prosodic phrase

PP n

PP n&PP n 1

PP n&PW n 1&PW n 2

T5

Context POS

POS n

POS n&POS n 1

POS n&POS n 1&POS n 2

No. 11
3.4

DONG Yuan et al.: A Two-stage Prosodic Structure Generation Strategy for

Rule design and generation

Every item of the template is made up of the combination

of several features referred. And every rule is formatted,
for example:

4.3

Experiments

Experiments are designed to test the performance of

the two-stage prosodic structure generation strategy. Experiments are designed and implemented in BaiLing TTS
system, a real TTS system authorized by France Telecom
R&D, Beijing. Experiments here can not only reect the
performance of the two-stage strategy but also be able to
show the great practicality in actual speech synthesis.
4.1 Corpus preparation
Our raw corpus comprises 20 000 sentences randomly selected from the corpus of People s Daily 2000. Each sentence had been manually segmented into syntactic words
with POS tags according to Specication for Corpus Processing at Peking University: Word Segmentation, POS
Tagging and Phonetic Notation.
On average, one sentence contains 45.24 syllables and
25.15 syntactic words. Then, this corpus was prosodically
annotated by two trained people; more than 90 % of their
annotating results were consistent with each other.
Based on the annotation of syllables and syntactic words,
both prosodic word and prosodic phrase boundaries were
marked out. The whole process was guided and supervised
by Cao Jian-Fen of Chinese Academy of Social Sciences.
19 000 sentences were randomly selected for CRF model
training and rest 1 000 sentences were used for testing of
all the following experiments. The two sets do not overlap
with each other.
In this paper, CRF++ 0.51 tool was used for
the model training, which can be downloaded from:
http://crfpp.sourceforge.net/.
For TBL module training, 30 words were selected which
are mostly happening with errors in prosodic boundaries
generation in the above test experiments with CRF model.
4.2 Evaluation criteria
Evaluation criteria are precision (P ), recall (R), and Fscore (F ). P shows the ratio between the number of correctly tagged prosodic boundaries and the number of all
tagged prosodic boundaries, while R indicates the proportion of correctly tagged prosodic boundaries number and
all manually tagged prosodic boundaries number. Besides,
F-score obeys the following equation:
F =

2P R
P +R

Experiments results

The baseline uses the maximum entropy (ME)-based

method for Mandarin prosodic structure prediction. In Table 6, the ME-based method achieves the F-score of 87.05%.
Table 6

POS(Y, 0)&LW(Y, 1) : A B
where Y indicates the feature value. 1 indicates the
relative position to the character (we constrain the oset
of each feature within the range of (3, 3)), A and B
indicate the initial prosodic boundary and the transformed
prosodic boundary, respectively, : is used to separate the
condition and prosodic boundary tag, is used to separate the current tag (left) and correct tag (right), & is
used to separate each sub-condition. Each sub-condition
comprises two parts: function name and its parameter
list. All the items are composed of ve features: POS,
LC, LW, PW, and PP. Through TBL learning,
we could generate the corresponding rule from the template whose items just match the context of the prosodic
structure boundary.

1573

Experiment performances
Precision (%)

Recall (%)

ME (PW)

87.95

86.16

F-score (%)
87.05

CRF (PW)

90.21

92.94

90.67

CRF&TBL (PW)

93.77

95.56

94.66

ME (PP)

71.77

77.24

74.40

CRF (PP)

78.61

81.55

80.05

CRF&TBL (PP)

83.91

85.33

84.61

The ME-based method and CRF-based method are using the same training and testing corpus.
As shown in Table 6, compared with the ME-based
method, the two-stage prosodic structure prediction strategy of CRF & TBL, respectively, improves the F-score by
7.61%.
It shows that the two-stage strategy is much more effective than the ME-based method for prosodic boundaries prediction. Moreover, compared with the simple CRF
based method, the method that combines CRF and TBL
works well.
Still, some inexible errors occur in our test of prosody
boundary prediction. Such as the following sentence
, the result of the two-stage strategy
is: /vq/PW /n/PP /vi/LW /vi/LW,
and the standard answer is: /vq/PW /n/PW
/vi/LW /vi/LW.
The strategy gives the PP prediction after the word
, as well as the real answer should be PW
boundary. Two reasons can be brought to explain such
case. For one reason, the length of the word is
three and is a verb. The strategy may probably
treat it as a V-O construction, in order to make the PP
boundary prediction. For another reason, the word
is quite new to the corpus of People s Daily 2000. This
word rarely or even not appear in the corpus, so that the
strategy is not able to get enough information neither in
CRF model training nor in TBL rule learning. Finally, the
strategy get the wrong prosody boundary prediction.
According to the two reasons above, the next work will
be making certain rule to x the error boundary and updating the corpus timely.
Due to dierence in the corpora and evaluation metric,
these results may not be comparable in all respects. Yet
from the statistics above, the approach presented in this
paper is a successful attempt towards prosodic boundary
prediction.

Conclusion

This paper makes an extensive investigation on Mandarin prosodic structure prediction of TTS system. Based
on the analyzing of a large-scale corpus, a two-stage
prosodic structure prediction strategy is proposed: conditional random elds model for the rst prediction and
transformation-based error-driven learning modication
module for the back-end of the system to amend the errors from CRF prediction.
Experiments results indicate that this approach is able to
achieve a good performance with high precision, recall, and
F-score. Moreover, the approach of the two-stage predic-

1574

ACTA AUTOMATICA SINICA

tion strategy for Mandarin prosodic boundaries works well

in real TTS system. It also shows great general application
in further processing of TTS system.
References
1 Chou F C, Tseng C Y, Lee L S. Automatic generation of
prosodic structure for high quality Mandarin speech synthesis. In: Proceedings of the 4th International Conference on Spoken Language. Philadelphia, USA: IEEE, 1996.
16241627
2 Cao Jian-Fen. Prediction of prosodic organization based on
grammatical information. Journal of Chinese Information
Processing, 2003, 17(3): 4146 (in Chinese)
3 Chu M, Lu S N. A text-to-speech system with high intelligibility and high naturalness for Chinese. Chinese Journal of
Acoustics, 1996, 15(1): 8190
4 Niu Z Y, Chai P Q. Segmentation of prosodic phrases for
improving the naturalness of synthesized Mandarin Chinese
speech. In: Proceedings of the International Conference on
Spoken Language Processing. Beijing, China: ISCA, 2000.
350353
5 Gee J P, Grosjean F. Performance structures: a psycholinguistic and linguistic appraisal. Cognitive Psychology, 1983,
15(4): 411458
6 Cao J F, Zhu W B. Syntactic and lexical constraint in
prosodic segmentation and grouping. In: Proceedings of the
Conference on Speech Prosody. Aix-en-Provence, France:
ISCA, 2002. 203206
7 Wang Hong-Jun. Prosodic words and prosodic phrases in
Chinese. Zhongguo Yuwen, 2000, 23(6): 525536 (in Chinese)
8 Mao X N, Dong Y, Han J Y, Huang D Z, Wang H L. Inequality maximum entropy classier with character features
for polyphone disambiguation in Mandarin TTS systems. In:
Proceedings of the IEEE International Conference on Acoustics, Speech, and Signal Processing. Honolulu, USA: IEEE,
2007. 705708
9 Wang M Q, Hirschberg J. Predicting intonational boundaries
automatically from text. In: Proceedings of the Workshop
on Speech and Natural Language. Pacic Grove, USA: Association for Computational Linguistics, 1991. 378383
10 Taylor P, Black A W. Assigning phrase breaks from partof speech sequences. Computer Speech and Language, 1998,
12(4): 99117
11 Zhou T, Dong Y, Huang D Z, Liu W, Wang H L. A
three-stage text normalization strategy for Mandarin textto-speech systems. In: Proceedings of the 6th International
Symposium on Chinese Spoken Language Processing. Kunming, China: IEEE, 2008. 14
12 Mao X N, Dong Y, Han J Y, Wang H L. A comparative
study of diverse knowledge sources and smoothing techniques via maximum entropy for polyphone disambiguation
in Mandarin TTS systems. In: Proceedings of the IEEE
International Conference on Natural Language Processing
and Knowledge Engineering. Beijing, China: IEEE, 2007.
162169
13 Laerty J D, McCallum A, Pereira F C N. Conditional random elds: probabilistic models for segmenting and labeling sequence data. In: Proceedings of the 18th International
Conference on Machine Learning. San Francisco, USA: Morgan Kaufmann, 2001. 282289

Vol. 36

14 Mao X N, Dong Y, Fang W B, He S K, Wang H L. A two-step

approach for Chinese named entity recognition based on conditional random elds and maximum entropy. In: Proceedings of the 7th International Conference on Chinese Computing. Wuhan, China: Publishing House of Electronics Industry, 2007. 434442
15 Brill E. Transformation-based error-driven learning and natural language processing: a case study in part of speech tagging. Computational Linguistics, 1995, 21(4): 543565

DONG Yuan
Received his Ph. D. degree in telecommunication from Shanghai
Jiao Tong University in 1999. From 1999
to 2001, he was with Nokia Research Center as a research and development scientist,
working on voice recognition on Nokia mobile phone. From 2001 to 2003, he worked
as a post doctoral research sta at the
Engineering Department, Cambridge University, UK, working on European speech
recognition project CORETEX. Since
2003, he has worked as an associate professor at Beijing University of Posts and Telecommunications. He is also a senior
research consultant at the France Telecom R&D Beijing. His
research interest covers speaker recognition, speech synthesis,
speech recognition, and multimedia content indexing.
E-mail: [email protected]
ZHOU Tao Received his master degree
in signals and information processing from
Beijing University of Posts and Telecommunications in 2010. His research interest covers text-to-speech synthesis and natural language processing. Corresponding
author of this paper.
E-mail: [email protected]

DONG Cheng-Yu Received his master

degree from Beijing University of Posts and
Telecommunications in 2007. Now, he is a
researcher at France Telecom R&D Beijing.
His research interest covers speech recognition, speech synthesis, and speaker recognition.
E-mail: [email protected]

WANG Hai-La Received his Ph. D. degree from University of Paris, France in
1986. Then, he worked at France Telecom
R&D for six years. From 1998 to 2004,
he was with several France telecom departments. Since 2004, he has been with France
Telecom R&D Beijing as a chief technology
ocer. His research interest covers valueadded multimedia service.
E-mail: [email protected]