JPH0363696A - Text voice synthesizer - Google Patents

Abstract

PURPOSE:To generate a synthesized voice of high quality with good understanding and high naturalness by finding the understanding difficulty of each word of an input character string, controlling the rhythm of the synthesized voice according to the found difficulty and controlling the syllable clearness. CONSTITUTION:A KANJI (Chinese character) and KANA (Japanese syllabary) mixed sentence which is inputted from an input part 10 is divided into respective words by a word division processing part 21 in a character string analysis part 20 by referring to a Japanese dictionary 14a and the understanding difficulty of each word is calculated. When a word divided by the processing part 21 is a word divided matching the storage contents of the dictionary 14a, a processing part 22 performs identical type word selection processing and outputs the result to a rhythm processing part 23. When not, on the other hand, a KANJI dictionary 14b is referred to and the word is converted by a processing part 24 into rendering, which is given an accent by a processing part 25 and outputted to the processing part 23. The processing part 23 calculates a rhythm parameter by using a memory 15 according to the understanding difficulty and performs rhythm limitation processing. A generation part 26 receives the parameter and performs the control processing of syllable clearness according to the understanding difficulty by referring to a dictionary 16, and a voice is synthesized 12 and outputted 13.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a text-to-speech synthesizer that synthesizes and outputs speech based on input of a character string.

[Prior art] In text-to-speech synthesis, speech is synthesized by extracting phonological sequences and prosodic information from Japanese orthography (Kanji, Kana, and Kana), and generating speech parameters according to predetermined rules based on these extracted contents. What is done is done.

When performing such regular speech synthesis, it is important not only to ensure the clarity and intelligibility of the synthesized speech, but also to improve its naturalness in order to improve the quality of the synthesized speech. However, it is difficult to improve clarity, comprehensibility, and naturalness at the same time. Therefore, it is generally recommended to first ensure a certain degree of clarity throughout the text to be synthesized, and then try to improve naturalness. It is being done. That is, the syllable intelligibility of the synthesized speech remains constant regardless of the content of the text.

[Lesson 1 that the invention attempts to solve! According to the conventional text-to-speech synthesis method described above,
Even words that can be easily understood from the flow of the text are synthesized to maintain a certain degree of clarity, so they are easier to hear than natural speech, making them feel less natural.I! In some cases, it may be difficult to hear words that are used infrequently, such as unfamiliar words, proper nouns, or numerals.

In addition, according to the conventional text-to-speech synthesis method, the synthesized sound is in one tone, so the listener must constantly listen to the synthesized sound with tension, and the problem is that the listener gets tired easily if listening for a long time. was there.

In other words, in the case of human speech, by appropriately changing the method of speech in consideration of the difficulty of the text, the level of understanding of the listener, etc., it is possible to generate speech that accurately conveys information and maintains naturalness. However, conventional text-to-speech synthesis methods have not been able to generate such speech.

Therefore, an object of the present invention is to provide a text-to-speech synthesis device that can generate synthesized speech with high intelligibility and naturalness.

[Means for Solving the Problems] A feature of the present invention that is distantly related to the above-mentioned object is that an input character string is parsed to generate voice parameters, and voice is synthesized based on the generated voice parameters. A text-to-speech synthesis device, comprising: a character string analysis unit that determines the comprehension difficulty level of each word in the character string; and a prosody processing unit that controls the prosody of the synthesized speech according to the determined comprehension difficulty level of each word. The present invention further includes a speech parameter generation section that controls the syllable intelligibility of the synthesized speech according to the obtained comprehension difficulty level of each word.

The character string analysis unit calculates the degree of comprehension difficulty based on the degree of familiarity of the word obtained by referring to a Japanese dictionary that also stores the frequency of use of each word in the character string, the part of speech of the word, the number of occurrences, and the interval of occurrence. It is desirable to have a comprehension difficulty calculation function.

It is preferable that the prosody processing section has a prosody control function that changes the fundamental frequency, power, and time length of the synthesized speech according to the degree of difficulty in understanding the word.

The speech parameter generation unit is a vowel constant part control that changes the ratio of the vowel constant part to the vowel transient part so as to improve syllable intelligibility by lengthening the vowel constant part for words that are difficult to understand. It is desirable to have this function.

The speech parameter generation unit may include a transient section control function that increases the allowable amount of temporal change in the speech parameter for words that are difficult to understand, thereby changing the allowable amount so as to improve syllable intelligibility. desirable.

It is desirable that the speech parameter generation unit has a devoicing determination function that does not devoice vowels of words that are difficult to understand.

[Example] Embodiments of the present invention will be described in detail below using the drawings.

FIG. 2 is a block diagram schematically showing the configuration of an embodiment of the present invention.

In the figure, numeral 10 is an input section into which the kanji-kana-original sentence to be synthesized is input, 11 is a control section, 12 is a speech synthesis section that synthesizes speech according to speech parameters and outputs it, and 13 is a speech synthesis section. 12 is an output unit that outputs a synthesized speech signal; 14 is a memory for a Japanese dictionary and a kanji dictionary; 15 is a memory for a Japanese dictionary and a kanji dictionary;
Reference numeral 16 indicates a prosody control memory, and 16 indicates a speech data dictionary memory. These input section 10, control section 11,
The speech synthesis section 12 and the memories 14, 15, and 16 are connected to each other via a bus 17.

The control unit 11 is mainly composed of a programmed computer, and generates audio parameters from input data provided from the input unit 10 using memories 14, 15, and 16, as will be described later.

FIG. 3 is a block diagram specifically showing the functional configuration of this control section 11 in detail.

The kanji/kana collocation sentence given from the input section 10 is applied to the word division processing section 21 in the character string analysis section 20 .

The word division processing unit 21 divides each word into a kanji-kana-kori sentence by referring to the Japanese dictionary 14a and using a known longest-match method or a clause-minimization method that selects words such that the number of clauses in the sentence is minimized. The Japanese dictionary 14a, which is divided into 0 Japanese dictionary 14a, stores in advance the general usage frequency, part of speech, pronunciation, accent, etc. for each word.

The word division processing unit 21 further calculates the ease of understanding of the divided words, that is, the degree of difficulty of understanding. As pre-processing for calculating the comprehension difficulty level, n words that have appeared in the past are stored.

Storage locations f(n) to f(1) are prepared as n word arrays, and after sequential shifting as described below, the current word is stored in f(1).

f(n) ←f(n-1) f(n-1)-f(n-2) f(2) ←f(1) However, for n, b(w) in Table 1 below is b( w)
= 0, and in the example in Table 1, n = 100.

The comprehensibility difficulty level calculation process is executed according to the program shown in FIG. First, in step S1, the usage frequency of the corresponding word is determined by referring to the Japanese dictionary 14a which stores the numerical usage frequency in five levels for each word, and this is stored as the comprehension difficulty level. However, if the word is not found in the Japanese dictionary 14a, the comprehension difficulty level "is set to the lowest value" = 1.As for the comprehension difficulty level r, the larger the value, the higher the frequency and the easier it is to understand.

Next, in step S2, the Japanese dictionary 14a is used to check whether the part of speech of the corresponding word is a numeral or a proper noun, and if it is a numeral or a proper noun, the comprehension difficulty level is set to 1. When storing it in 14a,
As for proper nouns, it may be possible to deal with this by reducing the frequency of use a little.

In the next step S3, it is determined whether the corresponding word is included in the words of nfil that appeared in the past.
2) to f(n). However, as described above, n is the value of W such that b(w) in Table 1 becomes b(w)=0, and in the example in Table 1, n is the old value of 100. If it is included in the past n words, proceed to step S4;
If not included, the process advances to step S5.

In step S4, the number of the corresponding word in the word array is checked, and the number n=2 is set as W. Furthermore, b(w) corresponding to this W is determined from Table 1, and the determined b(w) is added to ``.

In step S5, it is determined whether the word follows,
If there is a subsequent one, the process advances to step S1, and if there is no subsequent one, this calculation process is ended.

Table 1 If the word divided by the word division processing unit 21 is a word divided by matching with the contents stored in the Japanese dictionary 14a, the word pronunciation accent processing unit 22 performs isomorphic word selection processing. This process is well known and is used to distinguish between words that have the same character but are pronounced with different pronunciations and accents. The processed words are output to the prosody processing section 23.

On the other hand, unknown words that are divided without being matched with the contents stored in the Japanese dictionary 14a are processed as follows using a known method. That is, among the unknown words, unknown words in kanji are once converted into pronunciations in the unknown word pronunciation processing unit 24 with reference to the kanji dictionary 14b in which the pronunciations of each kanji character are stored in advance, and then converted into pronunciations in the unknown word pronunciation processing unit 24. 25. Among the unknown words, unknown words in plain name and katakana are output as they are to the unknown word accent processing section 25 with the plain name and katakana as readings. In the unknown word accent processing section 25, a process of generating an accent from the pronunciation is performed using a predetermined rule, and the accented unknown word is output to the prosody processing section 23.

The prosody processing unit 23 uses the accents of each word obtained from the word reading accent processing unit 22 or the unknown word accent processing unit 25 to set segment accents when words are chained, set phrases, and pauses between exhalation paragraphs. The setting is performed using a known method.

Furthermore, in this prosody processing section 23, calculation of prosody parameters according to the comprehension difficulty level and prosody control processing are performed using the prosody control memory 15.This processing will be explained below using the program shown in FIG. do.

Note that the prosodic parameters FO1pw and ``■'' described here are those that control the average fundamental frequency, average power, and average duration of the entire text, respectively, and these prosodic parameters have already been calculated using a known method. shall be.

In step 81G, each prosodic parameter Fo, Pw
, Tr is multiplied by the function S(''). That is, Fo is Fo−
3(r), pw to Pw-3(r) ■" to ■"・S(
).However, the function S(``) is a function as shown in Table 2, for example, with the comprehension difficulty r as a variable.The functions shown in Table 2 change depending on the type of prosodic parameter. However, they agree that each prosodic parameter is kept the same or increased depending on the difficulty of understanding.

Table 2 Next, in step 811, the vowel length Iv and the consonant length 1c are calculated from the following equations.

Iv=IV/C/-1c(v,I) -TrIC snow 1
/c/ ・Ic (c, I) ・■"However, Iv/c
/ is the basic length of the vowel by preceding consonant, I/C/ is the specific length of the consonant,
Ic (v, I), Ic (c, I) are the mora position coefficients for each phoneme, T' is the basic time length, ■ is the vowel, C is the consonant, and I is the mora position.The vowel interval is the time axis The inside is divided into a transient part 1v1, a steady part 1v2, and a transient part 1v3.The steady part Iv2 may be calculated by any known method, but it is possible to control the syllable intelligibility described later. Corrected in processing.

In the next step Si2, the pitch pattern F(t) is calculated as follows.

I n(F(t))-j n(Fmin)+Ap-Gp
(t-To)+Aa-Fo-(Ga(t-TI)-Ga
(t-72))Gl)(t)=a+t-exEl
(-a -t)Ga(t)=1-(1+β-t)-e
xp(-β-13, where ^D-GE) (t-To) is a phrase component, ^a-F.

-(Ga(t-Tl)-Ga(t-72)) is the accent component, Fm1n is the lower critical value, Ap, ^q is the phrase component, the amplitude of the accent component, TO1■1 is the phrase component, the start of the accent component A command time point, 12 is a command time point to end an accent component, α and β are a phrase component, a falling coefficient of an accent component, and t is time.

In the next step 313, the power pattern (pH) is calculated in the same manner as the calculation of the pitch pattern F(t) described above. However, if Fo in step 812 is changed to Pw, F
This is done by replacing min with Pm1n (lower limit critical value).

By controlling prosody as described above, words that are difficult to understand have a longer time length and a higher power, for example.

As a result, it becomes easier to listen to, as will be described later.

Vowel length Iv, consonant length 1c calculated by prosody processing unit 23,
The pitch pattern F(t) and the power pattern p(t) are applied to the audio parameter main generator 26. This voice parameter main generator 26 uses a voice data dictionary 16 as a synthesis unit.
A synthesis unit corresponding to the pronunciation of each word is searched by referring to , and these interpolations and synthesis are performed according to the above information from the prosody processing unit 23, and finally the time series of speech parameters for speech synthesis is obtained. can get.

Further, in this voice parameter main generating section 26, a control process of syllable intelligibility is performed in accordance with the comprehension difficulty level r.

FIG. 5 shows an example of this syllable intelligibility control processing method. In this example, for words that are difficult to understand, the ratio of the vowel constant part to the vowel transient part is changed so as to improve syllable clarity by lengthening the vowel constant part.

The prosody processing unit 23 calculates the duration IVl of the transition part of the vowel part.

The steady-state time length 1v2 and the transient time length 1v3 are calculated in advance, and each section is corrected using the program shown in FIG.

First, in step 820, the transient portion time lengths 1vl and Iv3 are corrected as follows using a function S(') as shown in Table 2, for example, with the comprehension difficulty level '' as a variable.

Ivl ←Ivl-1v2 ・(S(r)-1)/2I
v3←Iv3-1v2 .(S(r)-1)/2 In the next step 821, the steady-state time length Iv2 is similarly corrected as follows.

Iv2 +Iv2-5(r) Next, in step S22, the audio parameter with the transition time length 1v1 is linearly interpolated to the closest target parameter of the preceding sound, and the audio parameter with the transition time length 1v3 is interpolated with the target parameter of the following sound. linearly interpolated to the nearest target parameter.

Audio parameters are generated by the above processing.

FIG. 6 is a diagram explaining the effect of the above-mentioned control processing method, and shows the temporal change in the acoustic parameters of the vowel chain, that is, the first formant and second formant of the vowel from /a/ to /i/.
It represents the temporal transition of formants. Compared to the conventional control processing method shown in the same figure (^), according to the above-mentioned control processing method shown in the same figure (B), the time length of the stationary part of each vowel becomes longer and the vowel part becomes clearer. This improves syllable intelligibility.

FIG. 7 shows another example of the syllable intelligibility control processing method. This example is a method of generating speech parameters by increasing the allowable amount of temporal change in acoustic parameters for words that are difficult to understand, thereby changing the allowable amount so as to improve syllable intelligibility.

First, in step 330, a target parameter is extracted for each phoneme according to the phoneme duration calculated by the prosody processing section 23 while referring to the speech data dictionary 16 as a unit for synthesis.

For the following explanation, it is assumed that, for example, the vowel series @/a (/ is synthesized. Also, the center position of the target parameter of each vowel is tl, t2 (each is the center of the vowel interval),
The magnitude of each first parameter is Fl(i), F2(
i), where i indicates the number of the target parameter.

In the next step 831, the permissible amount τ(i) of the temporal change in normal utterance is corrected as follows using the function S(') as shown in Table 2, for example, with the comprehension difficulty level '' as a variable.

τ(i)←τ)・S(“) Next, in step 832, for each phoneme, a position t that is the temporal center between the center position IFt2 of the subsequent target parameter and the center position tl of this phoneme is determined.

is calculated from the following formula.

tO snow t1+(t2-tl)/2 In the next step 533, the magnitude of the target parameter Fo(i) at to is calculated from the following equation.

Fo(i)=(Fl(i)+F2(i))/2 In the next step 334, the temporal change of the target parameter F(t,i) between tl and 12 is expressed as Fl(i) and F2(
In relation to i), it is calculated using the following formula.

However, t represents time.

When Fl(i) < F2(i), F(t, 1) - FHi) tl < t < (Fl(i) - Fo(i))/τ(i
)+t.

F(t, 1)=Fo(i)+r(i)−(t−t
o) (Fl(i)-Fo(i))/r (i)+to≦
t≦(F2(i)−Fo(i))/τ(i)+t.

F(t, 1)=F2(i) (F2(i)−Fo(i))/r(i)+to<t
When < t2F1(i) >F2(i), F(t, 1) = FHi) tl< t < (Fo(t)-Fl(i))/r (
i)+t.

F(t, 1) = Fo(i)-r(i) ・(t-to
)(Fo(i)−Fl(i))/τ(i)+to≦t≦
(Fo(i)-F2(i))/τ(i)+t.

F(t, 1)=F2(i) (Fo(i)−F2(i))/r(i)+to<t
< Voice parameters are generated by the processing above t2.

FIG. 8 is a diagram for explaining the effect of the above-described control processing method, and shows the temporal transition of the first formant and second formant of the vowel from /a/ to /i/. The same figure (A
According to the conventional control processing method shown in ), there is a large restriction on the temporal change in the vowel transition part, and the change is gradual. However, according to the above-described control processing method shown in FIG.
For example, in words that are difficult to understand, the temporal change becomes large, the transient part becomes shorter and the steady part becomes longer, resulting in improved syllable intelligibility.

As a processing method for controlling syllable intelligibility according to the comprehension difficulty level in the voice parameter main generating section 26, a devoicing determination process in which the vowel of a word with a high comprehension difficulty level is not devoiced will be described below.

In general, noble vowels /i/ and /u/ sandwiched between voiceless consonants that are not in the accent core are often devoiced according to the rules because devoicing them increases naturalness.However, words that are difficult to understand Therefore, if the intelligibility level is less than 2, for example, semi-devoicing processing that shortens the duration or reduces the power without devoicing is performed. By providing the voice parameter main generating section 26 with a devoicing determination function that performs this, syllable intelligibility can be improved.

The procedure of this devoicing determination process will be explained below using FIG. 9.

First, in step S40, it is determined whether the target vowel is a high tongue vowel (/i 8 to U/), and if it is not a high tongue vowel, it is determined that it is voiced, and the process proceeds to step S44, where no devoicing processing is performed and the vowel is high. If it is a tongue vowel, the process advances to the next step 541.

In step 841, it is determined whether the target vowel is sandwiched between voiceless consonants, and if it is not sandwiched, it is determined that it is voiced and the process proceeds to step 344, where no devoicing processing is performed. Proceed to Ste Nob 342.

In step 842, it is determined whether or not the target vowel has an accent nucleus, and if it does, the process proceeds to step 846 in which quasi-devoicing processing is performed to shorten the duration or reduce the power of the accent. If it does not have a nucleus, proceed to the next step 343.

In step S43, it is determined whether the comprehension difficulty level ``of the target vowel is 2 or less, and if it is 2 or less, the process proceeds to step 846 in which the above-mentioned semi-devoicing process is performed.

If it is 3 or more, it is determined that there is no voice and the process proceeds to step 845.
Perform silent processing.

By performing the above processing in the voice parameter main generation section 26, voice parameters for synthesis are generated and applied to the voice synthesis section 12.

The speech synthesis section 12 synthesizes a signal corresponding to an actual synthesized speech waveform based on the speech parameters using a known method, and outputs the synthesized signal to the output section 13.

Next, the operation of this embodiment will be explained with reference to a case where the sentence "My name is Izuko," is actually synthesized.

When this sentence is input from the input unit 10, the word division processing unit 21 divides it as shown in (A) of Table 3, and the Japanese dictionary 14a further divides the part of speech and part of speech for each word as shown in (B) of Table 3. Usage frequency can be obtained. The word division processing unit 21 further calculates the comprehension difficulty level of each divided word as shown in (C) of Table 3.

In the word accent processing unit 22, (0
) are given the pronunciations and accents shown below.

In the prosody processing unit 23, the fundamental frequency of the part of ``Izuko Maeda'' with a low comprehension difficulty level is increased, the power is increased,
Prosodic control is performed to lengthen the duration, making this part much easier to listen to.

In the speech parameter generation unit 26, the syllable intelligibility control process described above causes the ``e'' in ``maede'' to have a higher proportion of the constant part of the vowel than the ``e'' in the name j. this is,
This is because the ``e'' in ``maede'' has a lower comprehension difficulty than the ``e'' in ``name.''Furthermore, in conventional processing, the ``tsu'' in ``itsuko'' is devoiced from the phonetic sequence. According to the present invention, the ``tsu'' in ``itsuko'' whose comprehension difficulty level is ``low'' is not devoiced by the above-described control processing.

Through the above processing, the part "Maede Izuko" is uttered more clearly than other parts, making it easier to hear proper nouns that are normally difficult to understand. Conversely, the power and fundamental frequency of words such as "desu" at the end of a sentence, which can be easily audible, are lower, which reduces the burden on the listener's ears and makes them less tiring to listen to than with conventional synthesized speech. This makes it possible to create difficult-to-understand synthesized speech.

The text-to-speech synthesis device of the present invention is not limited to the embodiments described above.

For example, the character string analysis unit may have any other configuration as long as it determines the comprehension difficulty level of each word in the character string, and the prosody processing unit also determines the comprehension difficulty of each word. Any other configuration may be used as long as it controls the prosody of the synthesized speech according to the difficulty level.
Similarly, any other configuration may be used as long as the speech parameter generation unit also controls the syllable intelligibility of the synthesized speech according to the obtained comprehension difficulty level of each word.

Furthermore, the configuration of the speech parameter generation unit does not necessarily have to include all of the configurations of the above-mentioned embodiments, but may include at least one configuration for controlling the syllable intelligibility of the synthesized voice according to the comprehension difficulty level of each word. good.

The speech synthesis section may be of any type as long as it performs speech synthesis using speech parameters.

[Effects of the Invention] As described in detail above, according to the text-to-speech synthesis device of the present invention, the character string analysis unit calculates the comprehension difficulty level of each word of the input character string, and the It is equipped with a prosody processing section that controls the prosody of the synthesized speech according to the degree of difficulty of understanding the synthesized speech, and a speech parameter generation section that controls the syllable intelligibility of the synthesized speech according to the obtained understanding difficulty of each word. It is possible to generate highly natural and high quality synthetic speech.

Table 3

[Brief explanation of drawings]

FIG. 1 is a flowchart of a prosodic control processing program in an embodiment of the present invention, FIG. 2 is a block diagram schematically showing the configuration of the above embodiment, and FIG. 3 is a block diagram showing the detailed functional configuration of the control section. 4 is a flowchart of a program for calculating comprehension difficulty, FIG. 5 is a flowchart of a program for an example of syllable intelligibility control processing, and FIG.
7 is a flowchart of another example of a program for controlling syllable intelligibility; FIG. 8 is a diagram illustrating the effect of the control process in FIG. 7. , FIG. 9 is a flowchart of a program for devoicing determination processing. 10... Input unit, 11... Group/control unit, 12...
...Speech synthesis section, 13...Output section, 14
・・・・・・Memory for Japanese dictionary and Kanji dictionary, 14a
...Japanese dictionary, 14b...Kanji dictionary, 15...Memory for prosody control, 16...
・Memory for voice data dictionary, 17...Bus, 2
1...Word division processing unit, 22...Word reading accent processing unit, 23...Prosody processing unit,
24. Old: unknown word reading processing unit, 25: unknown word accent processing unit, evening generation unit. 26...Voice balance diagram Figure 6 /a/ /1/ Figure 8

Claims (1)

[Claims] A text-to-speech synthesizer that parses an input character string to generate voice parameters, and synthesizes voice based on the generated voice parameters, the character string determining the degree of difficulty of understanding each word in the character string. an analysis unit, a prosody processing unit that controls the prosody of the synthesized speech according to the obtained comprehension difficulty of each word, and a speech that controls the syllable intelligibility of the synthesized speech according to the obtained comprehension difficulty of each word. A text-to-speech synthesis device comprising: a parameter generation section.