Content uploaded by Sangramsing N. Kayte
Author content
All content in this area was uploaded by Sangramsing N. Kayte on Jun 11, 2016
Content may be subject to copyright.
International Journal of Computer Applications (0975 – 8887)
Volume 130 – No.3, November 2015
35
Hidden Markov Model based Speech Synthesis:
A Review
Sangramsing Kayte1
Research Scholar
Deprtment of Computer
Science & IT
Dr. Babasaheb Ambedkar
Marathwada University,
Aurangabad
Monica Mundada1
Research Scholar
Deprtment of Computer
Science & IT
Dr. Babasaheb Ambedkar
Marathwada University,
Aurangabad
Jayesh Gujrathi, PhD
Department of chemistry,
Pratap College, Amalner
ABSTRACT
A Text-to-speech (TTS) synthesis system is the artificial
production of human system. This paper reviews recent
research advances in field of speech synthesis with related to
statistical parametric approach to speech synthesis based on
HMM. In this approach, Hidden Markov Model based Text to
speech synthesis (HTS) is reviewed in brief. The HTS is
based on the generation of an optimal parameter sequence
from subword HMMs. The quality of HTS framework relies
on the accurate description of the phoneset. The most
attractive part of HTS system is the prosodic characteristics of
the voice can be modified by simply varying the HMM
parameters, thus reducing the large storage requirement.
Keywords
TTS, speech corpus, Marathi phonemes.
1. INTRODUCTION
The primary task of Text-to-speech (TTS) synthesis is to
translate input text into intelligible and natural sounding
speech. The TTS components involves the two phases [1], the
front end- which analyses the text, creates possible
pronunciations for each word in the context with grapheme to
phoneme conversion. The back end generates the speech
waveform along with the prosody of the sentence to be
spoken. The evaluation of TTS system is based on 3
attributes: Accuracy, Intelligibility and naturalness. The HTS
system provides the frequency spectrum (Vocal tract),
fundamental frequency (vocal source) and duration (Prosody)
of speech, which are commonly modeled simultaneously by
HMMs. Speech waveforms are generated from HMMs
themselves based on maximum likelihood criterion [2]. The
figure 1 shows architecture of TTS system [3].
Figure1. Architecture of TTS system
The architecture of TTS system is divided into four modules
as follows:
1. Text analysis: The normalization of the text wherein the
numbers and symbols become words and an abbreviation
are replaced by their whole words or phrases etc. The
linguistic analysis which means syntactic and semantic
analysis and aims at understanding the context of the text
is also performed in this step. The statistical methods are
used to find the most probable meaning of the utterances.
This is significant because the pronunciation of a word
may depend on its meaning and on the context.
2. Phonetic Analysis: This converts the orthographical
symbols into phonological ones using a phonetic
alphabet.
3. Prosodic Analysis: Prosody contains the rhythm of
speech, stress patterns and intonation. The naturalness in
speech is attributed to certain properties of the speech
signal related to audible changes in pitch, loudness and
syllabic length, collectively called prosody. Acoustically,
these changes correspond to the variations in the
fundamental frequency (F0), amplitude and duration of
speech units [1].
4. Speech Synthesis: This block finally generates the
speech signal. This can be achieved either based on
parametric representation, in which phoneme realizations
are produced by machine, or by selecting speech units
from a database. The resulting short units of speech are
joined together to produce the final speech signal.
2. LITERATURE SURVEY
All Paul Taylor explained the basic concepts of speech and
signal processing in “Text-to-Speech Synthesis” [5]. Basics of
speech synthesis and speech synthesis methods are discussed
in “An Introduction to Text-to-Speech Synthesis” by Thierry
Dutoit [1]. Text to speech system organization, functions of
each module and conversion of text which is given as input in
to speech is clearly explained in this book. Different speech
synthesis methods, which are used for development of
synthesis system, are explained in “Review of methods of
Speech Synthesis” [6]. The process of text normalization is
understood from “Normalization of non standard words” by
Christopher Richards. In this book, conversion of non-
standard words (NSWs) in to standard words is explained
clearly with examples. Non-standard words are the words
which are not found in the dictionary. NSWs are tokens that
need to be expanded into an appropriate orthographic form
before the text-to-phoneme module [7]. The concept of
prosody is explained briefly in “Text to speech synthesis with
prosody feature” by M. B. Chandak. This book explains how
International Journal of Computer Applications (0975 – 8887)
Volume 130 – No.3, November 2015
36
the prosody is predicted from the text which is given as input.
In linguistics, prosody includes the intonation, rhythm and
lexical stress in speech. The prosodic features of a unit of
speech can be grouped into syllable, word, phrase or clause
level features are largely determined by the naturalness of the
prosody generated during synthesis. The correct prosody also
has an important role in the intelligibility of synthetic speech.
Prosody also conveys paralinguistic information to the user
such as joy or anger. In speech synthesis system, intonation
and other prosodic aspects must be generated from the plain
textual input. Paper by Ramani Boothalingam (2013)
presented the comparison of the performance of Unit
Selection based Synthesis (USS) and HMM based speech
synthesizer. Difference between the two major speech
synthesis techniques namely unit selection based synthesis
and HMM based speech synthesis is explained clearly in this
paper [9]. Unit selection systems usually select from a finite
set of units in the speech database and try to find the best path
through the given set of units. When there are no examples of
units that would be relatively close to the target units, the
situation can be viewed either as lacking in the database
coverage or that desired sentence to be synthesized is not in
the domain of the TTS system. To achieve good quality
synthesis, the speech unit database should have good unit
coverage. To obtain various voice characteristics in TTS
systems based on the selection and concatenation of acoustical
units, a large amount of speech data is needed. It is very
difficult to collect and segment large amount of speech data
for different languages.. These features are manifested as
duration, F0 and intensity [8]. Prosodic units do not need to
necessarily correspond to grammatical units. In HMM based
speech synthesis system, the parameters of speech are
modeled simultaneously by HMMs. During the actual
synthesis, speech waveforms are generated from HMMs
themselves based on maximum likelihood criteria. The
advantage of using HMM based speech synthesis technique
for developing TTS system is discussed briefly in “An
Overview of Nitech HMM-based Speech Synthesis System
for Blizzard Challenge 2005” by Heiga Zen and Tomoki
Toda. The concept of accuracy measurement is outlined. The
parameters on which the accuracy of a TTS system depends
are discussed [10]. The concept of HMM‟s explained clearly
in “Text-to-speech synthesis” by Paul Taylor. This book
explains the concepts of Hidden Markov Model and how they
are used in synthesis process. HMM themselves are quite
general models and although developed for speech
recognition have been used for many tasks in speech and
language technology and are now in fact one of the
fundamental techniques in speech technology applications.
Text to Speech(TTS) is a system in which sequence of words
are taken as input and converts them to speech. In conversion
process of speech synthesis method, vowels and consonants
are most important in marathi language [1]. Each phonemes
are made of combination of consonants and vowels. There are
different concatenation methods like unit selection,diaphone
or domain specific method for speech synthesis. In all these
methods the voices are sampled from real recordedspeech and
speech synthesis is handled by computers. Here researchers
main focus is to designing a database of phonemes in such a
way that it speedup searching and retrieving process in
marathi TTS.
3. HIDDEN MARKOV MODEL (HMM)
BASED SPEECH SYNTHESIS
In the early 1970s, Lenny Baum of Princeton University
invented a mathematical approach to recognize speech called
Hidden markov model (HMM). The Hidden markov model
(HMM) [11] [12] [13] is a doubly stochastic process that
produces a sequence of operations. Table 1 compares the
HMM based speech synthesis system and Unit selection.
Table1. Compares the HMM based speech synthesis and
Unit selection
HMM based
Unit selection
Statistics based
Multi template based
Clustering( Use of
HMM)
Clustering( possible use of
HMM)
Multiple tree
Single tree
Advantage:
Smooth , stable
Advantage:
High quality at waveform level
Disadvantage:
Vocoded speech
(buzzy)
Disadvantage: Discontinuity or
miss
Small run time data
Large run time data
Various voices
Fixed voices
3.1 Speech Synthesis and Development for
Indian languages
India is a multilingual society with 1652 dialects/native
languages. Speech technologies can play a very important role
in development of applications for common people. Prior
1990s, Indian speech synthesizers were research synthesizers,
generating small segments of speech in non-real time and the
progress was very slow. Speech synthesizers were not
developed for commercial purpose. In the 90s, Government of
India had funded Indian language projects generously,
through Technology Development for Indian Languages
(TDIL) and other schemes [42].
3.2 Current Research projects for speech
synthesis in India:
Some of the institutions in India are engaged in speech
synthesis. The IIT Madras has worked on a novel scheme
where the “unit” is a character of written “text‟. The Tata
Institute of Fundamental Research (TIFR), Mumbai has
reported unlimited continuous speech synthesizer using
formant synthesis technique. Whereas TIFR [14] and Central
Electronics Engineering Research Institute (CEERI) [15]
worked with formant synthesis, ISI, Kolkata [16], Indian
Institute of Information Technology (IIIT), Hyderabad [17],
and center for Development of Advanced Computing
(CDAC), Pune and Kolkata developed concatenation-based
synthesizers. Between the concatenation and formant
synthesizers, the quality obtained so far is comparable. Speech
synthesizers based on Festival has been developed in
languages including Hindi, Bangla, Kannada, Marathi and
Tamil.
International Journal of Computer Applications (0975 – 8887)
Volume 130 – No.3, November 2015
37
3.3 Speech Corpora developed by the
LDC-IL
Linguistic Data Consortium for Indian Languages (LDCIL) is
the Consortium responsible to create the database and shall
provide forum for the researchers all over the world to
develop speech application using the collected data in various
domains. The LDC-IL has collected Speech databases in
various Indian languages, the details are described in [18].
The research that has been carried out is mostly for text to
speech synthesis which uses phoneme/syllables concatenation
on isolated words and is either based either on concatenative
or formant synthesis techniques. The need of the hour is to
work on the continuous speech and apply latest techniques
such as Hidden Markov Models for development of T-T-S for
general purpose or limited domain to achieve true application
potentials of speech synthesis. Although Indian language
speech synthesis has come up a long way, the amount of work
for Indian languages in speech domain has not yet reached to
a critical level to be used as real communication tool, as that
in other languages of developed countries.
3.4 Current development of HMM –based
speech synthesis system for Marathi (HTS)
Marathi language consists of 33 consonants and 12 vowels.
The monophone HMM’s designed to build the phoneset for
the language [19]. The lexicon describes the set of words
known by the system and their pronunciation. In HMM, based
speech synthesis, the speech parameters of a speech unit such
as spectrum, fundamental frequency (f0) and phoneme
duration are statistically generated by using HMM based on
maximum likelihood criterion. The most attractive part of
HTS system is that its voice characteristics, speaking styles,
or emotions can easily be modified by transforming HMM
parameters using various techniques such as adaptation,
interpolation, Eigen voice, Or multiple Regression [20]. HTS
i.e. HMM based text to speech synthesis system is an open
source tool which provides a research and development
platform for statistical parametric speech synthesis [21]. The
HMM-based speech synthesis system (HTS) has been
developed by the HTS working group as an extension of the
HMM toolkit (HTK). The source code of HTS is released as a
patch for HTK. The first version 1.0 HTS was first released in
December 2002. After an interval of three years, HTS version
2.0 was released in December 2006 with major update and
inclusion of number of new features, such as introduction of
global mean and variance calculation tool, for large databases
the previous version often suffered from numerical
errors.HTS version 2.0.1 was a bug –fixed version and the
latest version, HTS version 2.1, was released in July
2008.This version includes important features; Hidden semi-
markov models (HSMMs) [22][23], the speech parameter
generation algorithm considering global variance (GV) [24],
advanced adaptation techniques [25], and stable version of run
time synthesis engine API. The HTS version 2.1, with the
STRAIGHT analysis/synthesis techniques [26], provides the
ability to construct the state-of-art HMM based speech
synthesis systems developed for the past Blizzard Challenge
events [27][28].
3.5 Architecture of typical HMM based
speech synthesis system
The main advantages of the referred HMM –based synthesis
techniques when compared with unit selection and
concatenation method is the fact that the voice alteration can
be performed without large databases, being at par with
quality with unit selection and concatenation ones. The figure
2 shows the architecture of HMM based speech synthesis
system [29]. In the training part, spectrum and excitation
parameters are extracted from speech database and modeled
by context dependent HMMs. In the synthesis part, context
dependent HMMs are concatenated according to the text to be
synthesized. Then spectrum and excitation parameters are
generated from the HMM by using a speech parameter
generation algorithm. Finally, the excitation generation
module and synthesis filter module synthesize speech
waveform using the generated excitation and spectrum
parameters. The training part performs the maximum
likelihood estimation by using the Expectation Maximization
(EM) algorithm [30]. In this process, spectrum (e.g., mel-
cepstral coefficients) [31] and their delta and delta-delta
coefficients and excitation (e.g., log F0 and its dynamic
features) parameters are extracted from a database of natural
speech and modeled by a set of multi-stream [32] context-
dependent HMMs (phonetic, linguistic, and prosodic contexts
being taken into account).
Figure2. Typical architecture of HMM based speech
synthesis system
In the temporal structure of speech, each HMM has its state-
duration distribution namely, the Gaussian distribution [33]
and the Gamma distribution [34]. They are estimated from
statistical variables obtained at the last iteration of the
forward-backward algorithm. As they have their own context
dependency, each of spectrum, excitation, and duration is
clustered individually by using phonetic decision trees [35].
Hence, the system can model the spectrum, excitation, and
duration in a unified framework. In the synthesis part, a given
word sequence is converted into a context dependent label
sequence, and then the utterance HMM is constructed by
concatenating the context-dependent HMMs according to the
label sequence. Then, various kinds of speech parameter
generation algorithm [36] [37] have been used to generate the
spectrum and excitation parameters HMM. Finally, the
excitation generation module and synthesis filter module
filter, such as Mel log spectrum approximation (MLSA) filter
[38] synthesize speech waveform using the generated
excitation and spectrum parameters.
International Journal of Computer Applications (0975 – 8887)
Volume 130 – No.3, November 2015
38
4. CONS AND PROS OF HMM BASED
SPEECH SYNTHESIS SYSTEM
4.1 Advantages:
The main advantage of statistical parametric synthesis is that
it can synthesize speech with various voice characteristics
such as speaker individualities, speaking styles, and emotions
etc. The combination of unit-selection and voice-conversion
(VC) techniques [39] can alleviate this problem but high-
quality voice conversion is still difficult. However, we can
easily change voice characteristics, speaking styles, and
emotions in statistical parametric synthesis by transforming its
model parameters. There are three major techniques to
achieve this, namely adaptation, interpolation, and Eigen
voices.
4.2 Disadvantages:
Although the operation and advantages of statistical parameter
speech synthesis is impressive, a few disadvantages are
associated with it. First, the parameters must be automatically
derivable from databases of natural speech. Second the
parameters must give rise to high quality synthesis; finally,
the parameters must be predictable from text; the synthesis
quality is intelligible but nowhere close to natural speech.
5. CONCLUSION
Synthetic speech is in progress from last few decades. The
study presented an overview of speech synthesis-past progress
and current trends. The three basic methods for synthesis are
the formant, concatenative, and articulatory synthesis. The
formant synthesis is based on the modeling of the resonances
in the vocal tract and is perhaps the most commonly used
during last decades. However, the concatenative synthesis
which is based on playing prerecorded samples from natural
speech is more popular. In theory, the most accurate method is
articulatory synthesis which models the human speech
production system directly, but it is also the most difficult
approach. Currently, the statistical parametric speech
synthesis has been the most rigorously studied approach for
speech synthesis. We can see that statistical parametric
synthesis offers a wide range of techniques to improve spoken
output. Its more complex models, when compared to unit-
selection synthesis, allow for general solutions, without
necessarily requiring recorded speech in any phonetic or
prosodic contexts. The unit-selection synthesis requires very
large databases to cover examples of all required prosodic,
phonetic, and stylistic variations which are difficult to collect
and store. In contrast, statistical parametric synthesis enables
models to be combined and adapted and thus does not require
instances of any possible combinations of contexts.
Additionally, T-T-S systems are limited by several factors that
present new challenges to researchers. They are 1) The
available speech data are not perfectly clean 2) The recording
conditions are not consistent & 3) Phonetic balance of
material is not ideal. Means to rapidly adapt the system using
as little data as a few sentences would appear to be an
interesting research direction. It is seen that synthesis quality
of statistical parametric speech synthesis is fully
understandable but has “processed quality” to it [43]. Control
over voice quality (naturalness, intelligibility) is important for
speech synthesis applications and is a challenge to the
researchers. As described in this review, unit selection and
statistical parametric synthesis approaches have their own
advantages and drawbacks. However, by proper combination
of the two approaches, a third approach could be generated
which can retain the advantages of the HMM based and
corpus based synthesis with an objective to generate synthetic
speech very close to the natural speech. It is suggested that a
more detailed evaluation and analysis, plus integration of
HMM based segmentation and labeling for building database
and HMM based search for selecting best suitable units shall
aid in using the better features of the two methods.
6. REFERENCES
[1] T. Dutoit, “An Introduction to Text-to-Speech
Synthesis”, Kluwer Academic Publishers, 1997.
[2] Black, A. Zen, H., Tokuda, K. “Statistical Parametric
Synthesis”, in proc. ICASSP, Honululu, USA,2007.
[3] X.Huang, A.Acero, H.-W. Hon, “Spoken Language
Processing”, Prentice Hall PTR, 2001.
[4] D. Jurafsky and J.H. Martin, “Speech and Language
Processing”, Pearson Education, 2000.
[5] Paul Taylor, “Text to Speech Synthesis”, University of
Cambridge, pp.442-446.
[6] Newton, “Review of methods of Speech Synthesis”,
M.Tech Credit Seminar Report, Electronic Systems
Group, November, 2011, pp. 1-15
[7] Christopher Richards, “Normalization of non-standard
words”. Computer Speech and Language (2001),
pp.287–333
[8] M.B.Chandak, Dr.R.V.Dharaskar and
Dr.V.M.Thakre,”Text to Speech with Prosody Feature:
Implementation of Emotion in Speech Output using
Forward Parsing”, International Journal of Computer
science and Security, Volume (4), Issue (3)
[9] Ramani Boothalingam,V Sherlin Solomi, Anushiya
Rachel Gladston,S Lilly Christina, “Development and
Evaluation of Unit Selection and HMM-Based Speech
Synthesis Systems for Tamil”, 978-1-4673-5952-8/13,
IEEE 2013 National Conference
[10] Heiga Zen, Tomoki Toda and Keiichi Tokuda. “The
Nitech-NAIST HMM-based speech synthesis system for
the Blizzard Challenge 2006”, INTERSPEECH 2005.
[11] J. Ferguson, Ed., “Hidden Markov Models for speech”
IDA, Princeton, NJ, 1980
[12] L.R. Rabiner, “A tutorial on hidden markov models and
selected applications in speech recognition” Proc. IEEE,
77(2), pp.257-286, 1989
[13] L.R.Rabiner and B.H. Juang, “Fundamentals of speech
recognition”, Prentice-Hall, Englewood Cliff,New
Jersey,1993.
[14] Furtado X A & Sen A, “Synthesis of unlimited speech in
Indian Languages using formant-based rules”’
Sadhana,1996,pp 345-362 .
[15] Agrawal S S & Stevens K, “Towards synthesis of Hindi
consonants using KLSYN88”, Proc ICSLP92, Canada,
1992, pp.177-180 .
[16] Dan T K, Datta A K & Mukherjee, B, “Speech synthesis
using signal concatenation”, J ASI, vol. XVIII (3&4),
1995, pp 141-145 .
[17] Kishore S. P., Kumar R & Sanghal R, “A data driven
synthesis approach for Indian language using syllable as
basic unit”, Proc ICON 2002, Mumbai, 2002 .
International Journal of Computer Applications (0975 – 8887)
Volume 130 – No.3, November 2015
39
[18] Agrawal S. S. 2010, “Recent Developments in Speech
Corpora in Indian Languages: Country Report of India”,
O-COCOSDA, Nepal.
[19] B. Ramani, S.Lilly Christina, G Anushiya Rachel, V
Sherlin Solomi,Mahesh Kumar Nandwana, Anusha
Prakash,, Aswin Shanmugam S, Raghava Krishnan, S P
Kishore, K Samudravijaya, P Vijayalakshmi, T
Nagarajan and Hema A Murthy.” A Common Attribute
based Unified HTS framework for Speech Synthesis in
Indian Languages”. 8th ISCA Speech Synthesis
Workshop. August 31 – September 2, 2013,Barcelona,
Spain
[20] Zen HeigaNose, Takashi Yamagishi, Junichi Sako, Shinji
Masuko, Takashi, Black, Alan W.” The HMM-based
speech synthesis system (HTS) version 2.0”. 6th ISCA
Workshop on Speech Synthesis, Bonn, Germany, August
22-24, 2007.
[21] K. Tokuda , H. Zen, J. Yamagishi, T. Masuko, S. Sako,
T. Toda, A.W. Black, T. Nose , and K. Oura, “The HMM
based synthesis system(HTS)” http://hts.sp.nitech.ac.jp/.
[22] H. Zen, K. Tokuda, T. Masuko, T. Kobayashi and T.
Kitamura,“A hidden semi-Markov model-based speech
synthesis system.” IEICE Trans. Inf.Syst., E90-D
(5):825–834, 2007.
[23] J. Yamagishi and T. Kobayashi. Average-voice based
speech synthesis using HSMM-based speaker adaptation
and adaptive training. IEICE Trans. Inf. Syst., E90-D
(2):533–543, 2007.
[24] J. Yamagishi, T. Kobayashi, Y. Nakano, K. Ogata, and J.
Isogai, “Analysis of speaker adaptation algorithms for
HMM-based speech synthesis and a constrained
SMAPLR adaptation algorithm”, IEEE Trans. Audio
Speech Lang. Process., 17(1), pp.66–83, 2009.
[25] H. Kawahara, I. Masuda-Katsuse, and A.de Cheveign´e,
“Restructuring speech representations using a pitch-
adaptive time-frequency smoothing and an
instantaneous-frequency-based F0 extraction: possible
role of a repetitive structure in sounds”, Speech Comm.,
27:187–207, 1999.
[26] H. Zen, T. Toda, M. Nakamura, and K. Tokuda, “Details
of Nitech HMM based speech synthesis system for the
Blizzard Challenge 2005. IEICE Trans. Inf. Syst., E90-
D(1):325–333, Jan. 2007.
[27] H. Zen, T. Toda, and K. Tokuda, “The Nitech-NAIST
HMM-based speech synthesis system for the Blizzard
Challenge 2006”, In Blizzard Challenge Workshop,
2006.
[28] J. Yamagishi, T. Nose, H. Zen, Z.-H. Ling, T. Toda, K.
Tokuda, S. King, and S. Renals,“A robust speaker-
adaptive HMM-based text-to-speech synthesis”, IEEE
Trans. Audio Speech Lang. Process., 2009. (accept for
publication).
[29] T.Yoshimura, K.Tokuda, T. Masuko, T. Kobayashi and
T. Kitamura,“Simultaneous Modeling of Spectrum, Pitch
and Duration in HMM-Based Speech Synthesis”In Proc.
of ICASSP 2000, vol 3, pp.1315-1318, June 2000.
[30] Dempster, A., Laird, N., Rubin, D., 1977,“ Maximum
likelihood from incomplete data via the EM algorithm”,
Journal of Royal Statistics Society 39, 1–38.
[31] Fukada,T., Tokuda, K., Kobayashi, T., Imai, S., 1992,
“An adaptive algorithm for mel-cepstral analysis of
speech”, In Proc. ICASSP. pp. 137–140.
[32] Young, S., Evermann, G., Gales, M., Hain, T., Kershaw,
D., Liu, X.-Y., Moore, G., Odell, J., Ollason, D., Povey,
D., Valtchev, V., Woodland, P., 2006,“The Hidden
Markov Model Toolkit (HTK) version 3.4.
http://htk.eng.cam.ac.uk/.
[33] Yoshimura, T., Tokuda, K., Masuko, T., Kobayashi, T.,
Kitamura, T. 1998, “Duration modeling for HMM-based
speech synthesis”, In Proc. ICSLP. pp. 29–32.
[34] Ishimatsu, Y., Tokuda, K., Masuko, T., Kobayashi, T.,
Kitamura, T., 2001,“Investigation of state duration model
based on gamma distribution for HMM based speech
synthesis”, In Tech. Rep. of IEICE. vol. 101 of SP 2001-
81. pp. 57–62, (In Japanese).
[35] Odell, J., 1995,“The use of context in large vocabulary
speech recognition”, Ph.D. thesis, University of
Cambridge.
[36] Tokuda, K., Yoshimura, T., Masuko, T., Kobayashi, T.,
Kitamura, T., 2000,“Speech parameter generation
algorithms for HMM-based speech synthesis”In Proc.
ICASSP. pp. 1315–1318.
[37] Tachiwa, W., Furui, S., “A study of speech synthesis
using HMMs” In: Proc. Spring Meeting of ASJ. pp. 239–
240,(In Japanese),1999.
[38] Imai, S., Sumita, K., Furuichi, C., “Mel log spectrum
approximation (MLSA) filter for speech synthesis”,
Electronics and Communications in Japan 66 (2), 10–18,
1983 .
[39] Stylianou, Y., Cap´pe,O., Moulines, E., 1998,
“Continuous probabilistic transform for voice
conversion”, IEEE Trans. Speech Audio Process. 6 (2),
131–142.
[40] Sangramsing Kayte , Kavita Waghmare , Dr. Bharti
Gawali "Marathi Speech Synthesis: A review"
International Journal on Recent and Innovation Trends in
Computing and Communication ISSN: 2321-8169
Volume: 3 Issue: 6
[41] Monica Mundada, Bharti Gawali, Sangramsing Kayte
"Recognition and classification of speech and its related
fluency disorders" Monica Mundada et al, /
(IJCSIT)International Journal of Computer Science and
Information Technologies, Vol. 5 (5) , 2014, 6764-6767
[42] Monica Mundada, Sangramsing Kayte, Dr. Bharti
Gawali "Classification of Fluent and Dysfluent Speech
Using KNN Classifier" International Journal of
Advanced Research in Computer Science and Software
Engineering Volume 4, Issue 9, September 2014
[43] Sangramsing Kayte, Monica Mundada "Study of Marathi
Phones for Synthesis of Marathi Speech from Text"
International Journal of Emerging Research in
Management &Technology ISSN: 2278-9359 (Volume-
4, Issue-10) October 2015.
IJCATM : www.ijcaonline.org
