CONCATENATIVE TEXT-TO-SPEECH SYNTHESIS OF
TWO-SYLLABLE FILIPINO WORDS
Lourdes T. Tupas, Rowena Cristina L. Guevara, Ph.D., and Melvin Co
Digital Signal Processing Laboratory
Department of Electrical and Electronics Engineering
University of the Philippines, Diliman
ABSTRACT
In concatenative-based speech synthesizers, one of the most important problems is proper union of speech units to achieve an intelligible and natural-sounding synthetic speech. For that purpose, speech units need to be processed and concatenated so that discontinuities at concatenation points are minimized. Another possible solution to this is by using a larger speech unit to decrease the number of concatenation points. In this project, which utilized two-syllable Filipino words, the speech unit is syllable. Characterization of these Filipino words is done to differentiate words of the same spelling but of different meanings. This characterization took note of the pitch, duration of utterance of each syllable in the word, and the first three formant frequencies. A digital signal processing
(DSP) block is also implemented. It accepts two-syllable text and outputs all the possible utterances of that word; this block is the text-to-speech synthesizer.
A two-interval forced choice test was conducted to evaluate the level of naturalness of the synthesized speech. Words of the same spelling but of different meanings are distinguished using the prosody and intelligibility test.

1. INTRODUCTION
Speech is the easiest way to communicate among human beings. The analysis and synthesis of speech signals are complex processes. It is not surprising that market-driven interests appeared just recently. Among the latest technology that makes use of speech signals are speech coding, synthesis, and recognition.
Text-to-speech (TTS) synthesis is the process of converting text to speech output. It is used in applications where it is inconvenient or impossible to record a human speech. A TTS system, when coupled with an Optical
Character Recognition (OCR) system, is a very powerful

tool for blind people in accessing written documents.
Machines can be an invaluable support in voice handicaps: with the help of an especially designed keyboard and a fast sentence-assembling program, synthetic speech can be produced in a few seconds to remedy these impediments [3]. In some cases, oral information is more efficient than written messages. TTS systems make it possible to access textual information over the telephone. These, among other things, motivated the researcher to develop a TTS system that will generate synthetic speech from a database of spoken Filipino syllables. A text-to-speech system is typically composed of two basic modules: a Natural Language Processing module and the Digital Signal Processing module. This research focuses only on the latter, which is responsible for speech unit manipulation and concatenation and production of synthesized Filipino speech.
An input text is converted into speech by concatenating pre-recorded speech units. For this case, the speech unit used is the syllable.
Section 2 will discuss the construction of the database. In section 3, a brief discussion about the characterization tool is presented. Section 4 describes the text-to-speech system. Evaluation results for the twointerval forced choice test and for the prosody and intelligibility test are in section 5. A discussion of the problems encountered in implementing the design is on section 6. Finally, concluding remarks are given in section
7.
2. DATABASE CONSTRUCTION
Database construction includes the proper recording and labeling of the speech corpus. The 528 two-syllable
Filipino words used as the training data were produced using 98 Filipino syllables, with each syllable having at most two letters only. Twenty persons recorded the words and the syllables. Using the Transcriber, each of the 20 wave files produced in the recording was properly labeled. The Transcriber is a tool that tags a segment of a sound file. Its output, the transcription file, gives information on the different segments tagged and the start and end synchronization times of these segments. The transcription was done in a way that two segments will tag a word, i.e., the two syllables consisting the word. This is to automatically compute from the transcription file the durations of utterance of the syllables in a word.
The speech sound files together with the transcription files produced make up the whole database.

Using the Praat program, a powerful tool used in speech synthesis, the computation of necessary values like frequency and time components had been made easier and more accurate. A DOS-based application called sendpraat.exe was used to make a praat script written in MatLab® be evaluated automatically into the praat program.
Text

3. CHARACTERIZATION OF TWO-SYLLABLE
FILIPINO WORDS
The basic features of speech used to characterize the words are pitch frequency (f0), duration of utterance and the first three formant frequencies (f1, f2, and f3). The main purpose of this characterization is to differentiate
Filipino words of the same spelling but of different meanings, e.g., words basa can convey two different meanings when pronounced differently: “read” or “wet.”
The manner by which each syllable is pronounced makes the two words different.
Characterization of such words in terms of pitch, duration of utterance, and the formant frequencies of each syllable of the word were implemented. These features extracted are also the vital elements used in generating the synthetic speech.
The duration of utterance is automatically computed from the synchronization times provided by the transcription file.
One of the most difficult tasks in speech processing is the determination of the pitch. It has been claimed that there is no single pitch determination algorithm that will be suitable and reliable for all applications [5]. This is mainly due to the irregularity and non-stationary natures of speech. The pitch determination algorithm (PDA) proposed by [5] was used in this research.
Using the real cepstrum of the speech signal, the first three formant frequencies are estimated. Short-time analysis signals of 10-15 ms are extracted using a
Hanning window. This window is enough to represent the spectral properties of the signal [7]. The estimation discussed in [7] is used here.
These features extracted from the speech signals were stored in a database and were later used in the speech synthesizer. 4. TEXT-TO-SPEECH SYNTHESIS SYSTEM
The implemented synthetic speech generator is composed of five processing blocks: pitch, time, waveform, and envelope modifications, and the concatenation and smoothing technique application.

Text Segmentation and
Representation

Isolated
Syllable
Sounds

Waveform Modification

Pitch Contour Modification

Time-alignment

Envelope Modification

Duration
Pitch
Formant
Frequencie

Concatenate Speech Units and Apply
Smoothing Technique

Spoken Word
Figure 1: Block Diagram of the Speech Synthesizer

The speech synthesizer utilizes a pair of speech signals. The first, denoted as source signal, is the naturally spoken word and the other, denoted as target signal, is the speech produced by concatenating isolated syllable sounds. The prosodic features of the source are used to modify that of the target. These features are the pitch, duration, intensity, and LPC coefficients.
The naturally spoken speech (syllable) segments of the input word are fetched from the database. These are the source signals. Each of these signals is characterized according to the following parameters: pitch, duration of utterance, intensity, and LPC coefficients. The first two parameters are available in the database of features extracted previously.
Figure 1 shows the generalized block diagram of the speech synthesizer.

4.1. Text Segmentation and Representation
The text segmentation tool serves as a preprocessing block of the speech synthesizer. Here, the input word is
“read” and segmented into its syllabic components. Each syllable is represented with its corresponding syllable sound. This sound will be the target signal.
4.2. Waveform Modification
The idea that speech is produced by passing air from the lungs through the vocal tract motivated this modification.
Based on informal listening tests, the speech produced in redefining the formant frequencies of the target signal is generally inferior to the speech generated when the LPC coefficients are modified. Because of this, the implemented speech synthesizer utilized the LPC coefficients instead of the formant frequencies.
The LPC autocorrelation method is used to determine the coefficients that will be applied to the excitation signal of the target. (The excitation signal or the residue is defined as the signal remained when the LPC coefficients are extracted.)
4.3. Pitch and Time Modification
The pitch of the target signal is modified by first defining a path from the source signal pitch mark to its corresponding pitch mark in the target signal. The source signal time axis has a direct path to a certain virtual signal.
The target signal time axis has a direct path to this certain virtual signal, as well, so that a one-to-one correspondence between the source and the target signals is realizable. Using this one-to-one property, the pitch of the target signal at a certain pitch mark is modified by replacing it with the pitch of the source signal at the pitch mark that corresponds to it.
The source and the target signals are time-aligned with each other using the Dynamic Time Warping
Method.
4.4. Envelope Modification
To further improve the quality of the generated speech, the intensity (envelope) of the target signal has been modified exactly the way the envelope of the source looks. This was done by dividing the envelope of the target signal with itself and multiplying to it the envelope of the source.

4.5. Concatenation of Speech Units and the Final
Output
Simple concatenation of the syllables will make the final output sound irritating because of the clicks that can be heard due to the discontinuity between the concatenated signals. To avoid this problem, a considerable length of the signals has been overlap-added. To ensure that the overlap between the two syllables is continuous, a smoothing technique is implemented. This is done by obtaining the correlation coefficients of the last 12.5ms of the first syllable and the first 12.5ms of the second syllable. The length that produced the highest correlation coefficient is chosen and the overlap-add process is done accordingly. Finally, the synthetic speech is played. If there are two utterances available for the word, steps 1 to 5 are repeated to generate and play the other utterance.
5. EVALUATION
Two types of tests were conducted. The first is to evaluate the level of naturalness of the synthetic speech generated and the other is to differentiate words of the same spelling but of different utterances (and/or meanings).
5.1. Two-interval Forced Choice Test
The two-interval forced test is a preference type of test used to evaluate the naturalness of the synthesized speech.
A pair of speech sounds of a word (100ms apart) is played to the listener once. One of the speech sounds (called an interval) is the naturally spoken and the other is the synthesized. The listener then chooses the interval that he prefers more.
Four sessions were conducted for each listener. The first is for trial: to familiarize the listener with the type of the test. The remaining three are the actual tests wherein
60 pairs of speech in each session are used.
Twelve pairs, repeated five times in random, are used in every session. This sums to a total of 180 pairs for all three test sessions. The result is computed by counting the number of times the listener chose the naturally spoken interval. To be considered successful, the number of times that the listener chooses the naturally spoken word must equal the number of times that he chooses the synthesized, i.e., a rate of 50%. This signifies that the listener cannot distinguish the difference between the two intervals.
Twenty-four listeners participated the test. The mean score is 82% with a standard deviation of 20%. This means that 82% of the time, a listener preferred the naturally spoken speech to the synthesized.
5.2. Prosody and Intelligibility Test

The main goal of this research is to differentiate words of the same spelling but of different utterances. This evaluation measure is used to check intelligibility of the speech and the correctness of its prosody.
Twenty-five words were played to a listener. For each word, a pair of meanings is provided. The listener is asked to note the meaning that the speech conveys based on his understanding. The number of times that the listener gets the correct meaning is the listener’s score.
Twenty-five listeners participated the test. The mean score is 94%. This means that the synthesized speech is considerably successful in sending the correct utterance of the word.
6. DISCUSSIONS
The naturalness of the generated speech when isolated from the original sound is considerably at a good level.
However, based on the results of the preference test, the naturally spoken word is still preferred. It is possible that the listener took cues from the slight noise that can be heard in the synthesized speech. This noise was produced during the time-alignment tool.
The prosodic features that were applied to the target signals were purely dependent on that of the pre-recorded naturally spoken words. This is also one of the factors that caused the slight noise in the synthesized speech. this happens when the recorded speech is not properly articulated. Prosodic features of the slight pauses due to breathing are also extracted and copied into the synthesized speech.
7. CONCLUDING REMARKS
The concatenative text-to-speech synthesizer was able to generate a natural-sounding speech such that words of the same spelling but of different meanings are distinguishable. Based on the prosody test, the meaning conveyed by the synthetic speech is sent 84.4% correct to the listener.
The speech synthesizer implemented followed five steps in generating synthesized speech. The text segmentation tool segments the input text into its syllabic components. Each syllable is represented with its corresponding syllable sounds called the target signals.
The waveform modification tool, the pitch contour modification tool, and the time-alignment tool are used to completely change the prosody of the target signals.

Finally, the syllable sounds modified are concatenated and thus, the synthetic speech is generated.
The factors that contribute to the success of the synthesizer can be attributed partially to the proper recording and labeling of the database, the speech unit used, which is the syllable, the combined power of the techniques used such as the PSOLA method, the dynamic time-warping, envelope detection, and waveform modification. 8. REFERENCES
[1] “A Short Review of Text-to-Speech Technology, Including a
List of Commercially Available Products and Freeware
Packages,” http://ai.about.com/library/weekly/aa020801a.htm.
[2] J. Deller, J. Hansen, and J. Proakis, Discrete Time
Processing of Speech Signals, New York: McMillam Publishing
Company, 1993.
[3] T. Dutoit, “A Short Introduction to text-to-Speech http://tcts.fpms.ac.be/synthesis/intrtts.html Synthesis,”
(December 1999).
[4] J. Deller, J. Hansen, and J. Proakis, Discrete Time
Processing of Speech Signals, New York: McMillam Publishing
Company, 1993.
[5] Y. Medan, E. Yair, and D. Chazan, “Super Resolution Pitch
Determination of Speech Signals,” IEEE Trans. Signal
Processing, vol.39, no. 1, pp. 40-48, Jan. 1991.
[6] E. Moulines and F. Charpentier, “Non-parametric Technique for Pitch-scale and Time-scale Modification of Speech,” Speech
Communication 16, pp. 175 – 200, 1995.
[7] W. D. Oliver “Singing Tree Implementation: Technical and
Artistic Issues,” http://feynman.stanford.edu/people/Oliver_www/singhtml/node3 3.html (June 1997).
[8] G. Pelton, Voice Processing, New York: McGraw-Hill, 1993.
[9] Tassa and J. S. Lienard, “A New Approach to Evaluation of
Vocal
Effort by the
PSOLA
Method,” http://www.essex.ac.uk/web-sls/papers/00-01/00-01.html. [10] W. Howitt, “Linear Predictive Coding (LPC).” http://asylum.sf.ca.us/pub/u/howitt/lpc.tutorial.html (October
1995).
[11] M.W. Kadous, “Dynamic Time Warping,” http://www.cse.unsw.edu.au/~waleed/phd/tr9806/node14.html (October 1998).