JPH01171000A - Voice synthesis system - Google Patents

Abstract

PURPOSE:To generate efficient phoneme data by generating the correlative characteristic curve between pitch and energy in the case of generating the phoneme data, and correcting an energy value by the characteristic curve with the pitch and normalizing phoneme data. CONSTITUTION:When a voice is synthesized by using phonemes as basic units, the respective phonemes are sectioned by a rising state, a stationary state, and a falling state and a phoneme time constant, continuance, pitch, a pitch time constant, energy, an energy time constant, and a sound source are determined by the sections to generate phoneme data. At this time, the correlative characteristic curve between the pitch P and energy E is generated preliminarily and the energy value E is corrected with the pitch P according to the characteristic curve to normalize the phoneme data (a). Consequently, the phoneme data can be efficiently generated.

Description

DETAILED DESCRIPTION OF THE INVENTION A. Field of Industrial Application The present invention relates to a regular speech synthesis method for synthesizing speech using phoneme data.

B1 Summary of the Invention The present invention synthesizes speech using phonemes as basic units.

Create phoneme parameters by classifying into steady and falling sounds, and determining the duration, pitch and energy that determine the pitch and loudness of each of these divisions, the sound source, and the time constants of these phonemes, pitch, and energy. The present invention relates to a speech synthesis method characterized in that the energy value is corrected by the pitch to normalize the phoneme data.

C1 Conventional technology Electronic devices that artificially synthesize and output speech have recently become low-cost LSIs for speech recognition and speech synthesis of one or several chips thanks to speech information processing and semiconductor large-scale integrated circuit technology. Various methods have been proposed depending on the purpose of use and constraints. For this speech synthesis,
There are recording and editing methods that record raw human voices and combine them appropriately to edit them into sentences, and two methods that do not use the human voice directly but extract only the parameters of the human voice. There is a method of artificially creating a speech signal by controlling the parameters during the speech synthesis process.

There is a PARCOR method which is widely used because it allows high-quality synthesized sounds to be obtained using this parameter method.

When handling audio using an electronic computer, the audio waveform is sampled at certain intervals, the audio signal value at each sampling point is converted from analog to digital, and the resulting values are displayed as codes of 0 and 1. In order to record faithfully to analog signals, it is necessary to increase the number of bits, but voice synthesis signals require a large amount of memory.

Therefore, various highly efficient encoding methods are being researched and developed in order to reduce the amount of information as much as possible.

As one of the methods, for information of one audio signal,
There is a delta modulation method, which uses a minimum of 1 bit. In this method, one bit is used to determine whether the next audio signal value is higher or lower than the current value, and if it is higher, it is given the code "B", and if it is lower, it is given the code "0", and the audio signal is encoded. In actual system configuration, a fixed amplitude step amount (delta) is determined, and the difference between the audio value obtained by conventional encoding and the input audio signal is determined to prevent errors from accumulating. Encoding is performed on the residual signal.

This kind of configuration is called predictive coding, and is based on the linear prediction method (
There are two methods: the Percoll method (which uses a partial autocorrelation function called a Percoll coefficient instead of the prediction coefficient of the linear prediction method).

D6 Problems to be Solved by the Invention As mentioned above, the method using predictive coding has the problem that it is difficult to make articulatory connections corresponding to the joints between sounds.

For example, in pronunciation from a vowel to a consonant to a vowel, the sound at the joint between the vowels is cut off in the process of going from the steady state of the vowel to the consonant sound through the transitional sound, and then through the transitional sound of the vowel to the steady sound of the vowel. It does not give a natural feeling when heard by humans.

Furthermore, in the case of musical instrument sound synthesis, the joints between scales are important, but since the synthesis method differs from the principle of sound generation in an actual musical instrument, it still lacks a natural feel, especially in reverberant sounds. In order to approximate natural sounds in both of these systems, there are problems such as the need for a large number of electronic components such as memory and arithmetic units, making the device expensive.

Means for Solving Problem E1 H. For 4 Curtains Therefore, the inventor of the present application believes that human sounds and musical tones of musical instruments are produced by changes in the shape of the human mouth and the length and cross-sectional area of the acoustic tube. We analyzed the traveling wave phenomenon that represents the transmission of sound waves in acoustic tubes using acoustic tube equivalent circuits, and focused on the fact that the cross-sectional area of the acoustic tube is inversely proportional to the surge impedance.We simulated the cross-sectional area by changing the surge impedance. change,
Created a speech synthesis method that enables smooth articulatory coupling by continuously changing surge impedance, making it easier to synthesize sounds similar to human speech and improving the naturalness of speech. and filed a patent application earlier (
Japanese Patent Application No. 62-91705 (hereinafter referred to as the earlier application).

When creating speech synthesis data by extracting parameters from natural speech based on the invention of this earlier application, there are variations in the pitch and strength (energy) of the natural sound, making it difficult to hear with the ear. It is necessary to normalize the intensity of the sound so that it sounds the same intensity as when it was heard. When normalizing, you can extract the actual waveform of each natural sound and compare its height and pitch, or you can output the sound and listen to it with your ears, and then adjust the pitch and energy to match the sound level. Creating synthetic data. This work requires a lot of time and skill. Therefore, as a result of various experiments, the present invention focused on the fact that there is a correlation between energy and pitch, and that where the pitch is large, the energy is also large.The present invention creates a correlation characteristic curve between pitch and energy, and from this curve, energy values are determined by pitch. This corrects the audio data and normalizes it to improve work efficiency.

F. Example First, the invention of the earlier application, which is the basis of the present application, will be explained in detail.

The change in the cross-sectional area of the vocal tract during vocal production is caused by, for example, when making the ``a'' sound, the back of the throat is narrow and the lips are open, and the exhaled air that is forced out of the lungs causes the vocal cords to open and close intermittently to absorb the exhaled air, resulting in the voice being distorted. The sound waves that are repeatedly reflected in the path (acoustic tube) come out as the sound waveform of "a". When the throat is wider and the tip of the lips are narrower, the sound waveform of "i" is output.

In this way, the frequency is determined by the shape of the mouth, and if the shape of the mouth is simulated, "a" or "i" will be uttered.

The shape of the mouth can be simulated by the cross-sectional area of the sound tube, and changes in the cross-sectional area of the sound tube can be simulated by changes in surge admittance. Therefore, the shape of the mouth can be simulated by changing the surge admittance. Since changes in surge admittance can be varied very easily in electrical circuits, various sounds can be synthesized using electrical signals. Figure 2 (a) shows the cross-sectional area AI.

A! ... A simulates the vocal tract by connecting acoustic tubes with different cross-sectional areas. In the same figure (a), the acoustic impedance is replaced with an LC circuit of an electric circuit, and each acoustic tube is made into one LC line, and the whole is a concentrated line.
-1 electric circuit. Figure 2 (2) is a traveling wave equivalent model diagram, showing the acoustic impedance 21. Z! ... Z7 is inversely proportional to the cross-sectional area of the acoustic tube (acoustic admittance is proportional), and is proportional to the speed C' of the sound wave and the air density ρ. In addition, in the same figure, Zg is the sound source impedance, and ZL
indicates radiation impedance, and arrows between blocks indicate forward waves and backward waves.

If you want to produce the sound "a" now, you can create the sound "a" by creating the mouth shape of "a" at the cross-sectional area of the acoustic tube corresponding to the tip of the lips and applying impulse P intermittently.
If you want to produce the sound "a" and then "i", you can make "i" by narrowing the cross-sectional area of the acoustic tube corresponding to the tip of the lips and giving it the shape of the mouth of "i". can get.

When the impulse P is applied continuously and intermittently to change the entire cross-sectional area to resemble the mouth of "i", the vocal tract is simulated by n acoustic tubes shown in Figure 2, so these By moving each cross-sectional area from "A" to "A", the shape of the mouth will change continuously to "A". Changing the cross-sectional area of the acoustic tube is done by gradually changing the surge impedance.

Therefore, since the cross-sectional area can be continuously changed, it is of course possible to obtain the steady-state sounds "a" and "i", but since the impedance can also be continuously varied, the sounds in between, i.e. You can get the sounds between the sounds. Therefore, articulatory combination similar to human pronunciation is performed smoothly without any sound breaks.

Next, considering the propagation speed of a sound wave, this is similar to the transient phenomenon when an impulse is applied to an electric line with length e and LC.

That is, when the line having LC as shown in FIG. 3 is equivalently represented, it becomes as shown in FIG. 4. Here is the surge impedance Z seen from both ends. i Zot is Z, ,=JL/C,
Zot=JL/C.

If we consider the traveling wave that has arrived from the other party as an equivalent current source, then 1+=it(L-T)+Vt(t f)Zo.

1! = i l(t − r ) + V l(
t −r )OI, and if there are n delay circuit blocks Z in the middle, the current is output after n hours. In other words, what occurred in the left circuit reached the right circuit one hour later.

I2 is the current i, + 'V, (t-t) on the feed pipe side
Z(11) However, in digital calculation, since the voltage or current is subdivided, V+ and Vt represent the voltage at measurement time t, and τ represents the elapsed time.

In Figure 4, if an impulse is applied to the L and C circuits, τ
After some time, it will come out to the output tube side. And what is reached before τ time equivalently represents that it has also been reached at the other party. Setting the line length C to 1 can be easily calculated by normalizing the delay block n to l. e
When dividing into 3ci, the delay block n should be set to 3 blocks.

Figure 2 (a) shows that the human vocal tract is approximately 17 cm long for a male.
If simulated with 17 acoustic tubes in 1CjI increments, the waveform entering from A will take 170 μsec to emerge from the An side if half-cycle current is divided into lO and its Δt is 10 μsec.

Therefore, the arithmetic processing unit performs arithmetic processing corresponding to the change in the cross-sectional area of the acoustic tubes AI-An, and the necessary A
1 to An, a memory that has values of impedances Z, ~Z, as a table, a calculation means that calculates the current values of each part of the equivalent circuit, and this equivalent circuit calculates the current values of adjacent equivalent circuits. If the arithmetic processing is performed using the arithmetic means for calculating the current value using the electric current, an audio signal will be obtained, and if the output is D/A converted and output to the speaker, it will be output as audio from the speaker.

Next, an example will be described in which speech is synthesized from character input signals according to rules using the acoustic tube model described above.

FIG. 5 is a block explanatory diagram for explaining one embodiment of the present invention, where l is a Japanese language processing unit that receives as input a sentence written in a mixture of kanji and kana, and associates this with dictionary 2. Separation of one clause and two sentences.

Natural language analysis is performed for morphological classification, and then accent processing is performed, phonetic conversion is performed, and intonation is added to create sentence processing data. 3 is a syllable processing unit which has syllable parameters and performs syllable processing on document-processed data.

The syllable parameters give the pitch, strength (energy), and duration of each syllable of 110 to 140 consonants (about 110 are sufficient for normally spoken words). For example, in the case of "cherry blossoms", as shown in Figure 6, S
Pitch P, energy E1, time T are normalized for each syllable of A, KU, and RA. 4 is a phoneme processing unit which has phoneme parameters with a parameter interpolation function. The phoneme parameters are determined at the rising part 08 of the sound for each phoneme. Steady part 02. The falling part 03 is divided into sections, and each section has a phoneme (cross-sectional area).
The time constant, duration, pitch, pitch time constant, energy, energy time constant, and sound source are normalized to form a block of data for each section. Taking the above-mentioned "cherry blossoms" as an example, as shown in FIG. 7, rSJ, rAJ +, rKJ.

If the rising part O3 of each phoneme of rUJ, rRJ, rAJ is Dog, Tr, P r, DP l+
Et.

D E +, G + data units are formed. These data units are provided corresponding to the cross-sectional areas A, .about.A1 of the acoustic tube model shown in FIG. 1, respectively. That is, if the cross-sectional area A of the acoustic tube model is 17, 6×17=102 data units are prepared for each syllable.

Each of the above time constants specifies how to move from the final value of the previous segment to the target value corresponding to each of that segment. Time T is a duration time, and each of the above processes is performed within this time T. In addition, the sound sources G, , G, , G3 vary depending on the time T for each segment in the consonant part, but provide a sound source with a pulse train of approximately 300 pulses and approximately 50 pulses in the vowel part.

6 is an acoustic tube model section that performs control to simulate changes in the cross-sectional area of the acoustic tube, and inputs its output to the speech synthesis waveform section 7;
A voice synthesis waveform unit 7 converts the digital signal into an analog signal and outputs it as voice from a speaker 8.

The above phoneme data is created by extracting parameters from natural speech. That is, the above energy, pitch, @
! Various parameters such as duration and cross-sectional area of sound tube are derived from basic audio data. At this time, the parameter values obtained from the basic speech data by inverse filter analysis are not used as they are, but are normalized and modified many times while checking the state of articulatory combination, etc., to create phoneme parameters.

At this time, if you try to normalize the intensity of the sound based on the energy value, energy is generally proportional to the height of the waveform, so if you want to make the sound stronger, increase the height of the waveform. However, the energy value affects not only the size of the waveform, but also the pitch. That is, in the case of waveforms of the same height, if the width of the pitch is small, the energy value will be small, and if the pitch is wide, the energy value needs to be large.

Therefore, it is difficult to normalize the intensity of sound simply based on the energy value.

Therefore, the present invention focuses on the fact that, from the sound data, the energy is also large where the pitch of the vowel is large.As shown by the dotted line in FIG. , a correlation characteristic curve of the average value of the dispersion is created as shown in Fig. 1 (a). Then, from this curve, the energy value is determined mechanically according to the pitch.

That is, depending on the type of sound, if the pitch is P, the energy is E, and if the energy is E from the waveform extracted from another sound, even if the pitch is P, then the energy is E,'. to correct. Note that since the pitch can be easily measured from the audio waveform, this correction can be easily performed. Also, this curve does not change depending on the type of vowel, but it changes depending on the person who makes the sound.

In the above embodiment, the creation of phoneme data in the case of the acoustic tube cross-sectional area has been described, but it can also be used to create phoneme data using the conventional Percoll method. At this time, the Percoll coefficient is used as the phoneme time constant.

H3 Effect of the Invention The present invention, when creating phoneme data as described above, creates a correlation characteristic curve between energy values and pitch, and corrects the energy value according to the pitch from this characteristic curve to normalize the phoneme data. Therefore, there is no need to normalize the energy values of parameters obtained from basic audio data by inverse filter analysis, etc., or to repeatedly make corrections while listening to the condition of articulatory combination etc., as in the past. Since the energy value can be normalized from the pitch to the pitch-energy correlation characteristic curve, phoneme data can be efficiently normalized without requiring any skill.

[Brief explanation of the drawing]

Fig. 1 is a pitch-to-energy correlation characteristic curve diagram for explaining the present invention in detail, Fig. 2 is an electric circuit equivalent model diagram of an acoustic tube, and Fig. 3 is an electric circuit diagram electrically simulating sound propagation. , Fig. 4 shows an equivalent circuit diagram of Fig. 3, Fig. 4 shows a block wiring diagram for synthesizing speech from character input signals to explain the present invention, and Fig. 4 shows a meter explanatory diagram. Processing unit, 3... Syllable processing unit, 4...
Phoneme processing section, 5... Parameter interpolation section, 6... Acoustic tube model section, 7... Speech synthesis waveform section, 8... Speaker. Pi1shiteP p, p2 Pi1νsuP Figure 5 1 Miya! eA section Fig. 6 Fig. 7

Claims (1)

[Claims] In a device that synthesizes speech using phonemes as basic units, each phoneme is classified into rising, steady, and falling, and for each of these classifications, the phoneme time constant, duration, pitch, pitch time constant, energy, energy time constant, When determining a sound source and creating phoneme data, the energy value is used to create a correlation characteristic curve between the pitch and energy, and the energy value is corrected according to the pitch from this characteristic curve to normalize the phoneme data. A speech synthesis method that uses