JPS63253996A - Sentence-voice converter - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed Description of the Invention] [Summary] The present invention provides a sentence-to-speech conversion device that converts a sentence into speech and outputs the same content by fixing the time length of each phoneme of synthesized speech. In order to solve the problem that the length of time is the same every time, resulting in unnatural synthesized speech, we synthesized speech by varying the length of time given to the synthesis unit symbol corresponding to each phoneme that is uttered. By adopting a structure that
The length of time is varied each time to output highly natural synthesized speech that is closer to human speech.

C. Industrial Application Field] The present invention relates to a sentence-to-speech conversion device that synthesizes and outputs speech by varying the time length given to each synthesis unit symbol corresponding to a sentence.

This text-to-speech conversion device takes any text as input, converts it into speech, and outputs it, and is used in many fields such as translation phones, computer speech output, and reading machines for the blind. be.

[Conventional technology]

The flow of conventional processing for synthesizing text into speech and outputting it will be described with reference to FIGS. 10 and 11.

In FIG. 10, the composition unit symbol generation unit 11 applies the conversion rule 12 to the input sentence, and converts it into a composition unit symbol string representing the basic unit of composition. This converted composite unit symbol string is generated by the parameter time series generator 1
3, the parameter rule 14 is applied, and the parameter rule 14 is converted into a parameter time series to be given to the speech synthesis unit 15.

This parameter time series is voice synthesized by the voice synthesis section 15 and output as voice.

Next, the process flow of the parameter time series generating section 13 will be explained in detail using FIG. 11.

In FIG. 11, the time length setting unit 13-1 applies the time length rule 13-2 to the notified combination unit symbol string to allocate a time length to each combination unit symbol, and
The parameter setting unit 13-3, which adjusts the time length in consideration of the difference in time length depending on the phoneme environment, uses the adjusted time length to set the parameter value corresponding to the synthesis unit symbol string, and further Perform interpolation to generate parameter time series.

At this time, in order to enhance the naturalness of the voice, the parameter time series generated in this way is processed to create smooth connections between parameters and give "fluctuations" similar to those seen in real speech. By inputting to section 15, speech is synthesized and output.

[Problem 3 to be solved by the invention: When a person repeatedly utters the same word (sentence), the uttered voice changes its physical characteristics (amplitude, pinch, formant frequency, utterance length) each time it is uttered. etc.) are changing.

In conventional sentence-to-speech conversion devices as shown in FIGS. 1O and 11, the parameter setting section 13-3 may connect parameters smoothly or add fluctuations to parameter values. However, since the time length remains fixed, there is a problem in that when the same content is uttered, an unnatural sound with the same length of time is synthesized and output every time.

[Means for solving problems]

In order to solve the above-mentioned problems, the present invention provides a time length setting section 2 that assigns a time length by applying time length rule 3 to a composite unit symbol generated in association with a sentence, and a time length setting section 2 that allocates a time length by applying a time length rule 3. A time length fluctuation addition unit 4 adds fluctuation to the time length assigned by the time length fluctuation rule 5 using the time length fluctuation rule 5, and to the time length to which fluctuation is added by the time length fluctuation addition unit 4, After setting parameter values by applying parameter rule 7, a parameter setting section 6 is provided which generates a parameter time series, and the parameter time series generated by the parameter setting section 6 is supplied to a speech synthesis section to generate speech. I am trying to synthesize and output.

FIG. 1 shows a basic configuration diagram of the present invention. In FIG. 1, a parameter time series generation section 1 receives a composite unit symbol as input, generates and outputs a parameter time series. This is the 10th
It corresponds to the parameter time series generation unit 13 in the figure, and outputs a parameter time series generated by adding fluctuation to the time length.

The time length setting section 2 applies the time length rule 3 to assign a time length to the input composite unit symbol.

The time length fluctuation adding unit 4 applies a time length fluctuation rule 5 to the assigned time length to add fluctuation.

The parameter setting section 6 applies the parameter rule 7 to set parameter values according to the synthesis unit symbol string and the time length to which fluctuations are added, and generates a parameter time series to be given to the speech synthesis section. .

[Effect]

Next, the operation will be explained.

In FIG. 1, when composite unit symbols are manually entered into the time length setting section 2, time length rule 3 is applied and a time length is set for each composite unit symbol.

When these time lengths are input to the time length fluctuation adding section 4, the time length fluctuation rule 5 is applied, and a different value of fluctuation is added to these time lengths each time. When the time length to which this fluctuation is added and the synthesis unit symbol are input to the parameter setting unit 6, the parameter rule 7 is applied, the parameter value is set, and further interpolation is performed to generate a parameter time series.

By inputting this parameter time series to the speech synthesis section, speech is synthesized and output.

As described above, a parameter time series is generated by further applying time length fluctuation rule 5 to the time length generated using time length rule 3 and adding fluctuations, for example, fluctuations generated using random numbers, By performing speech synthesis and outputting, even if the same content is uttered, natural speech with a different length of time is output each time.

〔Example〕

Next, the configuration and operation of one embodiment of the present invention will be explained in detail using FIGS. 2 to 9.

In FIG. 2, a time length setting unit 2 applies a time length rule 3 to the input composition unit symbol to assign a time length.

The time length adjustment unit 2-1 adjusts the time length to the assigned time length.
The time length adjustment rule 2-2 is applied to adjust the difference in time length depending on the environment of each composite unit symbol.

The time length fluctuation adding section 4 adds fluctuation to the adjusted time length by applying the time length fluctuation rule 5, taking into account that the width of fluctuation differs for each composite unit symbol.

The super-segmental element parameter setting unit 4-1 applies the super-segmental element parameter rule 4-2 to the time length to which the fluctuation is added, and sets the ultra-segmental elements (phonetically meaningful speech) such as amplitude and pitch. It goes beyond the segmental element, which is the smallest unit of speech, and sets parameters for a wide range of speech features such as syllables, phrases, and sentences.

The parameter setting unit 6-1 sets a parameter value by applying parameter rule 7 to the supersyllabic element.

The parameter interpolation unit 6-2 performs interpolation between set parameter values, such as linear interpolation, quadratic curve interpolation, etc.
It generates a parameter time series to be given to the speech synthesis section.

Next, the flow of the process in the configuration shown in FIG. 2 will be sequentially explained using the rules shown in FIGS. 3 to 7.

FIG. 3 shows an example of a time length rule. This shows the rule applied by the time length setting unit 2 to the input synthesized speech symbol. This rule 1 is "The time length of AE is 2.
00m5".

Here, “AE” is the phonetic symbol for the English word “cat” (
``a!'' in (kaat) (corresponding to the compound unit symbol)
, DU represents an interval, and 5tart and end represent time variable names used in processing.

The time length setting unit 2 converts this rule 1 into a composite unit symbol “a!”
By applying to the time variable 5tart=
o, and a time variable and-200 (ms).

FIG. 4 shows an example of a time length adjustment rule.
Indicating the rules applied by the time length adjustment unit 2-1,
Rule 2 states that "the duration of the vowel preceding a voiced fricative is extended by 120%," and rule 3 states that "the duration of the vowel preceding a voiceless stop is shortened to 75%." Here, the characters in [ ] represent the features (acoustic characteristics) of the composite unit symbol, and must be set in advance for each composite unit symbol. For (vowel), the unit of composition is a vowel, (+
voc) indicates voiced, (-voc) indicates voiceless, (fric) indicates fricative, and (stop) indicates stop consonant.
The first term from the left of the rule, for example [νo@el], represents the composite unit symbol or its feature to which the rule is applied, and the second term, for example (Do 120%), represents the result after applying the rule. , the third term, for example “-(+vo
c fric ) ” represents the environment (composite unit symbol or its feature) of the target □composite unit symbol, and the same applies hereafter.0 Time length adjustment unit 2-1 applies these rules 2 or 3 to the composite unit symbol. By doing this, the time variable end becomes 1.2 times or 0.75 as shown on the right side.
be multiplied.

FIG. 5 shows an example of the time length fluctuation rule. This rule 4 shows the rule applied by the time length fluctuation adding section 4 to the time length after being adjusted by the time length adjustment rule 2-2. Rule 5 is ``The duration of vowels fluctuates within 20% of that range'', and Rule 5 is ``The duration of fricatives fluctuates within 10% of that range''.
Here, RND is a random number, such as a uniform random number or a random number filtered so that the spectrum is 1/f, and has a value from -1 to +1. When the 0 time length fluctuation addition unit 4 representing 0 is applied to, for example, the composite unit symbol AE, the calculation shown in the right column of the rule in Figure 5 is executed.For example, if the random number takes the values -0, 3, In this case, the time variable end=200 set in Rule 1 is set to the time variable end=188 calculated by the following formula.

end-200+200X0.2X (0°3)=18
8......+11 or more rules are applied to complete the processing for the time length.Next, the processing for setting parameter values will be described.

FIG. 6 shows an example of a supersegmental parameter rule. This shows the rule applied by the supersegment element parameter setting unit 4-1 to the time length to which fluctuation is added.This rule 6 is "The amplitude between the start point and end point of a vowel is OdB",
Rule 7 is ``The amplitude 30 ms after the start of the vowel is 6.
0 dB, and the amplitude 3 Qms before the end point is 55 dB," and Rule 8 means "The amplitude 70 m5 after the start point of the vowel is 70 dB." This means that the black circle in Figure 8 (a) This means that the parameter value at the position indicated by is set. Further, Rule 9 indicates that "the pitch at the start point of a vowel is 100 Hz", and Rule 10 indicates that the pitch at the end point of one vowel is 0.9 times the pitch of the previously set starting point.

As a result, the position parameters are set using the black circles in FIG. 8(b). Here, the time variable 5tart and time variable end used have the values set in the previous stage. Further, AV represents an amplitude value, FO represents a pinch frequency, and $ represents a variable for holding a previously set parameter value and for later use.

FIG. 7 shows an example of a parameter rule, which shows the rule applied by the parameter setting unit 6. This rule 11 is 'The first formant of the starting point and ending point of AE is 600.
Hz, the second formant is 1600Hz, and the third formant is 2300Hz.''

As described above, parameter values such as amplitude, pitch frequency, first formant, second formant, and third formant are set as shown in FIG. 9 for the time length to which fluctuation is added. And, FIG. 2 parameter interpolation unit 6-
2 interpolates the discontinuously set parameter values as shown in FIG. 9 so that they have values at each parameter specification interval required by the speech synthesis section. This interpolation is performed by linear interpolation,
It is performed using quadratic curve interpolation, interpolation using a critical braking quadratic system, etc. When performing this interpolation, fluctuations may be added to the parameter values. The parameter time series generated by interpolation in this way is used for speech synthesis. By synthesizing and outputting the voices input into the section, it is possible to output highly natural voices whose duration changes each time they are uttered.

As already mentioned, FIG. 8(a) schematically represents the parameter values set according to rules 6 to 8 in FIG.
FIG. 8(b) schematically represents the parameter values set according to rule 9.10 of FIG. 6.

FIG. 9 shows the parameter values set according to the rules 1 to 3.6 to 11 described above. Note that this figure does not show the fluctuation applied according to rule 4.5, but the If rule 4.5 applies, then T
200 m5 of IMER is added as a fluctuation by 20% (when rule 4 is applied) or 10% (when rule 5 is applied) of the random number.

〔Effect of the invention〕

As explained above, according to the present invention, since a configuration is adopted in which speech is synthesized by varying the time length given to the synthesis unit symbol corresponding to each phoneme to be uttered, even if the utterance content is the same, By changing the length of time each time, it is possible to generate highly natural synthesized speech that is closer to human speech.

[Brief explanation of drawings]

Figure 1 is a diagram showing the principle of the present invention, Figure 2 is a diagram showing the configuration of one embodiment of the present invention, Figure 3 is an example of a time length rule, Figure 4 is an example of a time length adjustment rule, and Figure 5 is a time length fluctuation diagram. Rule example, Figure 6 is an example of a hypersegmental parameter rule, Figure 7 is an example of a parameter rule, Figure 8 is an example of applying a hypersegmental parameter rule, and Figure 9 is an example of a hypersegmental parameter rule.
The figure shows an example of parameter values, FIG. 10 is a block diagram of a sentence-to-speech conversion device, and FIG. 11 is a block diagram of a conventional parameter time series generating section. In the figure, l is a parameter time series generation part, 2 is a time length setting part, 3 is a time length rule, 4 is a time length fluctuation adding part, 5 is a time length fluctuation rule, 6 is a parameter setting part, and 7 is a parameter rule. represent.

Claims (1)

[Claims] In a sentence-to-speech conversion device that converts a sentence into speech and outputs it, a time length rule (3) is applied to a composition unit symbol generated in association with a sentence to assign a time length. and a time length fluctuation adding section (4) that adds fluctuation by applying the time length fluctuation rule (5) to the time length allocated by the time length setting part (2). and a parameter setting unit (6) that applies parameter rules (7) to set parameter values for the time length to which fluctuations have been added by the time length fluctuation addition unit (4), and then generates a parameter time series. A sentence-to-speech conversion device comprising: a parameter time series generated by the parameter setting unit (6) is supplied to a speech synthesis unit to synthesize and output speech.