JPS63199399A - Voice synthesizer - 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 [Industrial Application Minute Hand] The present invention relates to a speech synthesis device, and particularly to a speech synthesis device that changes the speech rate by thinning out or duplicating feature parameters for speech synthesis.

[Prior Art] An audio signal is almost stationary within a certain period of time. Conventionally, focusing on this point, the audio signal is analyzed at fixed time intervals, each section is expressed by a set of characteristic parameters based on the analysis results, these are stored in advance, and these are used during speech synthesis. A method is known in which feature parameters are extracted at regular intervals and sequentially synthesized. This method is of practical use because the synthesis operation is extremely simple and there is little deterioration in sound quality. Specifically, the − set of feature parameters corresponds to audio having a certain length of time. Therefore, the duration of the synthesized speech can be increased or decreased by appropriately thinning out or duplicating the set of feature parameters. Conventionally, attempts have been made to change the speaking speed using this method. However, the plosive consonant (
k, t, p.

b, d, g, r, etc.) are short, so at most 1
Only one or two sets of feature parameters are combined. Therefore, in the conventional method, if a set of feature parameters that are thinned out or used repeatedly happens to correspond to a plosive consonant, the intelligibility of speech is significantly impaired.

[Problems to be Solved by the Invention] The present invention eliminates the drawbacks of the prior art described above,
The purpose is to provide a speech synthesizer that does not impair the clarity of synthesized speech even when the speech rate is changed.

- [Means for solving the problems] In order to achieve the above object, the speech synthesis device of the present invention has the following features:
a storage means for storing feature parameters corresponding to a voice of a predetermined length of time and at least information on whether speech rate control is possible in correspondence with each of the feature parameters, and the content of the enable/disable information allows speed control during speech synthesis; The outline of the present invention is to include a speed control means for thinning out or duplicating feature parameters only for feature parameters.

Further, in order to achieve the above object, the speech synthesis device of the present invention has a storage means for storing feature parameters corresponding to speech of a predetermined length of time and multi-valued information regarding whether or not speech rate control is possible, which is associated with at least each of the feature parameters. a threshold value setting means for setting a threshold value according to a speaking speed; and a speed for thinning out or duplicating feature parameters only for feature parameters whose multi-valued information is smaller than the threshold value during speech synthesis. The outline of the system is to include a control means.

Preferably, the storage means stores maximum multivalued information corresponding to a feature parameter indicating the point of rupture of the plosive consonant,
Its aspect is to store multivalued information that decreases in accordance with subsequent feature parameters.

Preferably, one aspect of the speed control means is that when the multivalued information has a predetermined sign, the characteristic parameter is not reused unconditionally.

Preferably, one aspect of the threshold setting means is to set a higher threshold as the speaking speed is faster or slower than the standard speed.

[Operation] In this configuration, the storage means stores characteristic parameters corresponding to a voice of a predetermined length of time and at least information (for example, binary information) on whether speech rate control is possible or not associated with each of the characteristic parameters. During speech synthesis, the speed control means thins out or uses duplicate feature parameters only for those feature parameters whose content of the availability information is speed controllable.

Further, in such a configuration, the storage means stores feature parameters corresponding to a voice of a predetermined length of time and multi-valued information regarding whether or not speech rate control is possible, which is associated with at least each of the feature parameters. Preferably, the storage means stores maximum multi-value information corresponding to a feature parameter indicating the point at which the plosive consonant ruptures, and stores decreasing multi-value information corresponding to subsequent feature parameters. The threshold value setting means sets the threshold value according to the speech rate (for example, a speech rate command from the outside).

Preferably, the threshold setting means sets a higher threshold as the speaking speed is faster or slower than the standard speed. The speed control means thins out or uses duplicate feature parameters only for feature parameters whose multi-valued information is smaller than the threshold value during speech synthesis. Preferably, when the multivalued information has a predetermined sign, the speed control means unconditionally does not use the characteristic parameter repeatedly.

[Description of Embodiments] Embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[First Embodiment] FIG. 1 is a block diagram of a speech synthesis apparatus according to a first embodiment of the present invention. In the figure, 1 is an input terminal,
A voice command, a voice speed command, etc. sent from a host side (not shown) are input. 2 is a central processing unit (CPU),
The synthesized voice is uttered and its speed is controlled according to the input utterance command and utterance speed command. 2A is a memory (ROM) that stores the control program executed by the CPU 2;
For example, the control program of the first embodiment shown in FIG. 6 or the control program of the second embodiment shown in FIG. 10 is stored. Furthermore, 3 is a first storage device that stores a set of voice characteristic parameters along with speed control enable/disable information, and 4 is a CPU.
2 is an auxiliary storage device used, 5 is a PARCOR type voice synthesizer, 6 is a D/A converter, 7 is an amplifier, and 8 is a speaker for audio output.

Figures 2 (A) to (C) are “Mitai” uttered by the same man.
Figure 2 (A
) is the audio waveform when speaking carefully, Figure 2 (B) is the audio waveform when speaking at approximately 1.5 times the speed, and Figure 2 (c) is the audio waveform when speaking at approximately 1.5 times the speed.
) shows the audio waveform when speaking at approximately twice the speed.

Figures 3 (A) to (C) are diagrams in which a part of each voice waveform in Figures 2 (A) to (C) is enlarged at the same magnification in the time axis direction, and shows the beginning of the voice "Yu". It shows. On the scale below the audio waveform, one division is a frame with a length of 10 milliseconds, and each frame expresses the audio waveform in that section using a set of characteristic parameters. For example, frame (a+) in FIG. 3A shows the characteristics of the rupture point of the consonant "t". As can be seen by comparing this with frame (b+) in Figure 3 (B) or frame (C1) in Figure 3 (C), the characteristic of the rupture of the consonant "t" is that the utterance rate changes. has not changed much. Therefore, when changing the R voice speed, if the characteristic frames at the point of burst are thinned out or used repeatedly, the characteristics will change significantly and the clarity of the voice will be impaired.

This point also applies to other plosive consonants (k, p, d, g, r, etc.). Therefore, in the first embodiment, when storing the speech waveform in the form of feature parameters analyzed frame by frame, information on whether or not speed control is possible is added to each frame. For frames that should not be thinned out or duplicated, the content of the permission information is set to “no”.

FIG. 4 is a diagram showing the structure of a set of availability information and feature parameters in the first embodiment. When the male voice "Mitai" is analyzed using frames of 10 milliseconds each, the total number of frames of the feature parameter set is N (1 to N). The set of characteristic parameters for each frame consists of pitch PJ (i is the frame number), amplifier A, and PARCOR coefficient.

Also, information e on whether or not speed control is possible is attached to each frame.

When the content of availability information e is “0′°, speed control (thinning,
Duplicate use) is possible, and when it is "1", speed control is not possible.

FIG. 5(A) is a diagram showing the relationship between the speed command V and the frame thinning or duplication cycle m in the first embodiment. In the figure, the content of the speed command (■) is 0° when the speed is standard. In this case, the CPU 2 outputs all the characteristic parameter sets shown in FIG. 4 as they are. The speed command V, which is faster than the standard, is not expressed as a positive integer from 1 to 4 degrees. In this case, C
PU2 executes the calculation m=6-1vl to find the period m,
In addition, since the sign of the speed command V is positive, control is performed to determine whether thinning is possible or not at the period m@. That is, the content of the speed control availability information e is checked every m frames, and if the content of e is '0' (possible), the PARCO of the feature parameter set of that frame is
Transfer to the R-type speech synthesizer 5 is thinned out. A speed command V that is slower than the standard is expressed as a negative integer from -1 to -4. In this case, the CPU 2 executes the calculation m=6-IvI to obtain the period m, and since the sign of the speed command V is negative, it controls whether or not duplicate use is possible for each period m. That is, the content of the speed control availability information e is checked every m frames, and if the content of e is
``(Acceptable), the set of feature parameters of that frame is used redundantly and transferred to the PARCOR type speech synthesizer 5.

FIG. 6 is a flowchart showing the speed control procedure of the first embodiment. When a voice command and a speed command V are input from the input terminal 1 shown in FIG. 1, the process shown in FIG. 6 is started. In FIG. 6, a variable j indicates a frame count value (frame number) and takes a value from 1 to N. Variable (period counter)
n counts the period m for thinning out or redundant use, and takes a value from 0 to m-1. The flag f indicates the completion status of thinning or duplicate usage within one cycle, and is 1
At the beginning of a cycle, it is reset to '0'' along with the cycle counter n, and when thinning out or duplication occurs, it is set to 1''. Further, in the case of duplication processing, the flag f is primarily set to the numerical value ``''-1'' as an index for using the same feature parameter twice.

<Initial Processing> In step S1, the calculation m=6-1vl is performed to obtain the period m. In step S2, the frame number j is set to a numerical value of 1 to enable access to the audio parameters (including availability information e) from frame (1). In step S3, the period counter n and flag f are reset. In step S4, the contents of the speed command ■ are checked.

<Standard speed utterance> If the content of the speed command (■) is 0°' in step S4, the standard speed is uttered. The flow proceeds to step Sll, where the set of feature parameters of the frame j is PARCO
It is transferred to the R-type speech synthesizer 5. Further, the PARCOR type speech synthesizer 5 synthesizes the transferred characteristic parameter set into speech information, the D/A converter 6 converts the synthesized speech information into an analog signal, and the amplifier 7 amplifies the analog signal. , the speaker 8 outputs synthesized speech.

Meanwhile, during this time, the CPU 2 checks the flag f in step S12, and since it is not .degree.-1.degree., the process proceeds to step S13, where the frame number j is incremented by 1. In step S15, the process waits for approximately one frame time length (10.9 seconds), and in step S16, it is checked whether the frame number j has reached the total number of frames N. If N has been reached, the output of the total number of frames N has been completed, and the process advances to step S19 to end the process. Also, while N has not been reached, step S1
Proceed to step 7 and increment the cycle counter n by 1.

In step 318, it is checked whether n<m, and if n<m, it is in the middle of one cycle, so the process returns to step S4 and the next frame is read. Also, if n<m, then n=
m (start of the next cycle), the process returns to step S3 and resets the cycle counter n and flag f. In this way, when speaking at the standard speed, a set of feature parameters with a total number of frames N is unconditionally output.

Voicing faster than the standard speed> If the content of the speed command V is not "o" in step S4, it means that the utterance is faster or slower than the standard speed. When the sign of the speed command V is positive, the utterance is faster than the standard speed, and the following thinning process is performed. Now, after passing through step S3, the timing of the cycle counter n=o and the flag f-=O (the beginning of one cycle) is the timing to check whether or not the thinning process is possible. In step S5, the contents of the flag f are checked. Since the content of the flag f is initially °゛O゛', the process proceeds to step $6, and the speed control enable/disable information eJ of the frame is read out. In step S7, the contents of eJ are
Check whether °0”°. If the content of the availability information eJ is ′
If it is 0, the frame is a frame for which speed control is enabled, and the flow advances to step S8, where it is checked whether the sign of the speed command ■ is positive. Now, since the sign of the speed command V is positive, the process proceeds to step SIO and sets the flag f to 1" to declare that the thinning process has been completed. In the determination at step S12, the content of the flag f is negative. Since it is not, the process proceeds to step S13 and increments frame number j by 1.The thinning process is thus completed by updating the frame number by one without performing the process of step Sll.In step S15, a predetermined time wait is performed. The waiting time in this case is not the length of one frame.Then, when the process returns to step S5 next time, the content of the flag f is not 0''.In other words, from step S5 to Proceed to Sll, read out the set of feature parameters one after another in the same way as described for standard rate utterance, and use PAR
It is transferred to the COR type speech synthesizer 5. In this way, the content of the acceptability information e,1 of the first frame is checked every period m, and if it is acceptable, the set of feature parameters of that frame is thinned out.

However, if the content of the availability information eJ is "°1pa" in the determination in step S7, the transfer of the feature parameter set of the frame is not thinned out.The flow advances to step Sll, and the feature parameter set of the frame is PARCOR Therefore, in this frame processing, the flag f is not set to 1', so the determination in the next step S5 also satisfies the flag f=o.Then, in step S6, the flag f=o is satisfied. The content of the availability information e Jul is checked, and if it is '0°', thinning processing is performed for this frame.In this way,
In the case of speech that is faster than the standard speed, thinning control is executed every period m, and if the feature parameter set of the frame cannot be thinned out, the next frame is thinned out according to the contents of the availability information e. As a result, faster-than-standard pronunciation is always faithfully achieved, and the frame at the point of the important plosive consonant is not lost.

Speech Slower than Standard Speed> If the sign of the speed command V is negative in the determination in step S8, this means that the speech is slower than the standard speed, and the following duplicate usage process is performed. Similarly, after passing through step S3, the timing of the cycle counter n=o and the flag f=o is the timing to check whether or not duplicate use processing is possible. Step S5
Now, check the contents of flag f. Since the content of the flag f is initially 0°, the process proceeds to step S6, and the speed control enable/disable information eJ of the frame is read out.

In step S7, it is checked whether the contents of "e" are PA0'''. If the contents of the permission information eJ are °゛0゛°, the frame in question is a frame for which speed control is allowed, and the flow goes to step S8. Proceed to check whether the sign of the speed command V is positive or not.Now, the sign of the speed command V is negative, so step S
Proceeding to step 9, the value "-t" is primarily set in the flag f as an index for using the same feature parameter twice. In step Sll, the first set of feature parameters is P
It is transferred to the ARCOR type speech synthesizer 5. Step S12
In this determination, since the content of the flag f is -1°, the process proceeds to step S14, where the flag f is set to 1".

This is a declaration to the effect that a set of feature parameters has been transferred an extra time due to overlapping use. Furthermore, by skipping the process in step S13, the frame number is not updated. That is, the feature parameter of this frame number is used twice.

In this way, in the determination of the next step S5, since the content of the flag f is °1°°, the frame number j is updated and each feature parameter set is used for PARCOR type speech synthesis until the corresponding one cycle is completed. Transfer to device 5.

However, if the content of the permission information e is 1° in the determination in step S7, the set of feature parameters of the frame is not redundantly transferred. The flow advances to step Sll, where the set of feature parameters of the frame is transferred to the PARCOR type speech synthesizer 5, and the frame number is updated by one at step S13. Therefore, in this frame processing, the flag f is not set to "1", so the next step S
The determination of 5 also satisfies the flag f=o. Then, in step S6, the contents of the next frame's availability information eJ and 1 are checked, and if this is "o", the process of overlapping use is performed for this frame.

In this way, in the case of utterances that are slower than the standard speed, control for overlapping use is executed every cycle m, and if the feature parameter set of the frame cannot be used overlappingly, the next frame's availability information e is By duplicating the set of feature parameters of the next frame according to the content, utterances slower than the standard rate are always faithfully achieved, and frames at the point of plosive consonants, which are important, are not duplicated.

FIG. 7(A) is a diagram showing the processing results for four types of speed commands V for the availability information e of the first embodiment, from frame (i-2) of the start part of the voice "Yu" to frame (
The processing results for 8 frames up to (i+5) are shown. In the figure, the "x" mark indicates a frame that has been thinned out, and the ◎" mark indicates a frame that has been used repeatedly. Also, the position of frame (i-2) is exactly a multiple of one cycle m at any speed. Assume.

In the case of speed command v=2, by calculation m=6-121,
The period m=4. Therefore, the availability information e for the first frame (i-2) and the fourth frame (i +2)
In this case, since they are all 'o' (acceptable), they are also thinned out.

When the speed command v=3, the period m=3 according to calculation m-6-131. Therefore, the first frame (i-2) and
Then, the contents of the acceptability information e of the third frame (i+1) etc. are checked, and in this case they are all O'°, so
They will be thinned out.

When the speed command v=4, the period m=2 according to the calculation mmilo-141. Therefore, the first frame (i-2) and
The second frame (i) and the second frame (i
+2) etc., etc., can be checked for internal customers with availability information e. In this case, frame (i-2) and frame (i+2) are both "0'" and are therefore thinned out. However, for frame (i), the content of availability information e is
°゛ (impossible), the feature parameter set of the frame is not thinned out, and the content of the acceptability information e of the next frame (i+1) is checked, and in this case, the content is “0°
°, so it is thinned out.

In this way, the average speech rate is not affected, and the characteristic parameters of frame N) indicating the plosive point of the plosive consonant "t" are transferred to the synthesizer 5 as they are without being thinned out, so that the intelligibility is improved. A certain voice is synthesized.

In the case of speed command V--4, the period m=2 due to calculation m=6-141. Therefore, the first frame (i-2) and
The second frame (i) and the second frame (i
+2), etc., is checked. Also, since the sign of V is negative, frames are used repeatedly. That is, in this case as well, frame (i-2) and frame (i
+2) are both 0°, so both are used repeatedly. However, since the content of the acceptability information e for frame (i) is °'1'' (unavailable), the feature parameter set of the frame is not used repeatedly, and the acceptability information e of the next frame (i+1) is The content is checked and in this case the content is 0'', so it is used redundantly. In this case as well, the average speech rate is not affected, and the characteristic parameters of frame (i) indicating the plosive point of the plosive consonant "t" are transferred to the synthesizer 5 only once without being used twice. As a result, clear speech is synthesized without duplication of plosive sounds.

FIG. 8(A) is a diagram showing the processing results of four types of speed commands V for the availability information e of the first embodiment together with the waveform of the original voice, and shows the frame (i-
The processing results for 8 frames from frame 2) to frame (i+5) are shown. Similar to FIG. 7(A),
The "x" mark indicates a frame that has been thinned out, and the "◎" mark indicates a frame that has been used repeatedly. As is clear from Figure 8 (A), the frame (
The signal i) is not subject to thinning out or repeated use, regardless of the speaking rate (2).

[Second Embodiment] The block diagram of the second embodiment is the same as that in FIG. The feature of the second embodiment is that the 1-bit enable/disable information e added to each frame in the first embodiment is multi-valued and used, thereby thinning out frames or using duplicate frames according to the size of the speed command V. The object of the present invention is to adaptively synthesize and output more natural and clear speech even when the speech rate is changed.

Returning to Figure 3 (A), this time we will look at the frame (al) at the time of the plosive point of the voiceless plosive consonant “t” and the next frame (a1+
+). As mentioned above, even if the speed command V is changed,
No significant changes were observed in frames (a,), (bI), and (0 plate) at the time of rupture. However, if we focus on the next frame (aI++), it is almost unchanged from the 1.5 times faster frame (b+++), but when compared with the twice faster frame (c1+1),
No more frames are found that represent the characteristics of frame (al+1). As the utterance becomes faster, the consonant "t"
” to the following vowel “a” becomes shorter, and this point is similar to other plosive consonants (k, p, b, d, g,
The same applies to the case of r, etc.).

Therefore, in the second embodiment, when changing the speaking speed V, when speaking faster than the standard, the frame at the point of rupture is not thinned out, and the thinning method is devised for the frame at the long sound joining part to the following vowel part, That is, by adaptively changing the thinning method according to the magnitude of the speed command V, a sound that is close to natural is synthesized and output. In addition, in speech that is slower than the standard, it is known that if the duration of a plosive consonant is increased beyond a certain level, the phonology of the consonant part is lost. Code information that only prohibits frame thinning is also added and processed to prevent changes in the phonology of plosive consonants due to repeated use of frames.

FIG. 9 is a diagram showing the structure of a set of availability information and feature parameters in the second embodiment. In the figure, the set of frame numbers and characteristic parameters are the same as those in FIG. 4, but the speed control enable/disable information e2 is different. The availability information e2 of the second embodiment is multivalued as shown in the figure, and "o"
It is expressed as a positive or negative integer including .

The content of the availability information e2 is such that when the absolute value is less than a predetermined threshold value t determined according to the speed command V, thinning examination or duplicate use of the frame is allowed; In some cases, the set of feature parameters of the frame is used to output audio as is.

Also, when a negative sign is attached to the content of the availability information e2,
Always excluded from duplicate use.

That is, if the speed command V is slower than the standard, the availability information e
The above processing is performed only on frames in which the content of 2 is not negative.

In Figure 9, for example, the maximum absolute value of 181 is given to frame (i) at the time of the plosive point of the voiceless plosive consonant "t", and the following three frames of the long conjunctive part leading to the following vowel "a" are given the maximum absolute value of 181. The absolute value for each is 131.121. It gives lit. If such a slope characteristic is given, the speed command V
When V is close to the standard speed (threshold t is low), only frames close to the vowel stationary part are subject to thinning out or duplication, and as the speed command V deviates from the standard speed (threshold t increases), the frame approaches the rupture point. Even frames are subject to thinning and duplicate use. At that time, the frame (i) of the consonant part
A negative sign is given to frame (i+1) and repeated use is unconditionally prohibited to prevent phoneme change.

FIG. 5(B) is a diagram showing the relationship between the speaking rate v1 threshold and the period m of thinning or overlapping use in the second embodiment. Similarly, the content of the speed command V is to set the standard speaking speed to 0.
", and if it is faster than the standard speed, a positive integer "1~4°°
When the speed is slower than the standard speed, it is expressed as a negative integer "-1 to -
It is expressed in 4°'. Then, the values of the threshold value and the period m are calculated using the following calculations (1) and (2) using the contents of the speed command V.
) to be determined.

t= l v l-1...(1) m=6-lvl...(2) Therefore,
If the speed command V is the standard 0°, the threshold value t=-1 according to operation (1), so in this case, the absolute value of the acceptability information e2 cannot take a value less than the threshold value. Therefore, frames are not always thinned out or used repeatedly, and feature parameter sets of all frames are synthesized and output as they are.

In this way, when the voice command and the speed command V are input manually from the input terminal 1, the CPU 2 executes calculations (1) and (2) to find the threshold value t and the period m. If it is a positive integer, the content of the availability information e2 is checked every m frames, and if the absolute value is less than or equal to the threshold, the feature parameter set of the frame is thinned out. Also speed command■
If is a negative integer, the contents of the availability information e2 are checked every m frames, and when the sign of e2 is not negative and is equal to or less than the threshold value, the feature parameter set of the frame is used redundantly.

FIG. 10 is a flowchart showing the speed control procedure of the second embodiment. Note that the same steps as those in FIG. 6 are given the same step numbers, and the description thereof will be omitted.

Initial Processing> When a voice command and a speed command V are input manually from input terminal 1, they are inputted to step 5100. In step 5100, the threshold value t=lvl-1 is determined according to calculations (1) and (2) above.
and find the period m=6-1vl.

<Standard speed utterance> If the content of the speed command V is 0°' in step 5101, the standard speed is uttered. The flow is step 510
Proceed to step 5, le2. It is determined whether +I>t. By the way, since the threshold value t=Iol-1=-1 at the standard speed, 1e2Jl>t is definitely satisfied. Therefore, step 5106 is executed throughout all frames to perform standard rate speech.

Speech faster than standard speed> If the content of the speed command V is a positive integer in step 5101, the speech is faster than the standard speed.

Similarly, step S3 is passed and the period counter n=o
The timing of the flag f=o (the beginning of one cycle) is the timing to check whether or not thinning processing is possible. Step S6
Then, the availability information e2J is extracted, and in step 5105 it is determined whether 1e2JI>t. If 1e2tl>
When t is satisfied, the process proceeds to step 5IO6 without thinning out the frame. In this case, the flag f is not set to ``1'', so step 510 is performed for the next frame as well.
In step 5, it is determined whether 1e2Jl>t. Also 1e2
When Jl>t is satisfied, proceed to step 5IO7,
The frame is thinned out, the flag f is set to 1', and the completion of the thinning process is declared.

<Voice Slower Than Standard Speed> If the sign of the speed command V is negative in the determination in step 5101, this means that the voice is slower than the standard speed, and the following duplicate usage process is performed. Similarly, after passing through step S3, the timing of the cycle counter n=o and the flag f=o is the timing to check whether or not the duplicate use process is possible. In step S6, the availability information e2j is extracted, and in step 5102
Then, it is determined whether e2J<O. When e2j<O, the frame is determined to be a frame for which duplicate use is prohibited, and duplicate use is not performed unconditionally. The flow proceeds to step 5106, where the set of feature parameters of the frame is transferred, and step S1
Proceed to step 3 and update the frame number.

Further, when e2j<0, control based on the threshold value is followed. That is, in step 5103, it is determined whether e 2J>t. When e2j>t, it is determined that the frame cannot be used in duplicate, and the frame is not used in duplicate. If eB>t is not satisfied, the process proceeds to step 5104, where the flag f is set to °°-1°°, allowing the frame to be used repeatedly.

FIG. 7(B) is a diagram showing the processing results of four types of speed commands V for the availability information e2 of the second embodiment. ) shows the processing results for eight frames. Similarly, "x" marks represent thinned out frames, and ◎ marks represent duplicated frames.

It is also assumed that the position of frame (i-2) is exactly at a multiple of one period m at any speed.

First, when the speed command v=2, since the content of ■ is positive, it becomes a target of thinning control. Threshold value t=1 and period m=4 are determined by calculation (1) and calculation (2). Therefore, each first frame is frame (i-2) and frame (i+2).
The absolute value of the corresponding availability information e2 is compared with the threshold value t. In frame ('i -2) l e2+-21
=O and the threshold value t=1 or less, so it becomes a target of thinning.

However, in frame (i+2), since fe2cross+21=2 is larger than the threshold value t=1, paddy thinning is not performed. In the next frame (i+3), l e2r+3 l =
1, the threshold value t=1 or less, and thinning is performed. In this way, when the speed command v=2, the feature parameter sets of frame (i-2) and frame (i+3) are thinned out.

Next, when the speed command v=3, since the content of ■ is positive, it becomes a target of thinning control. Threshold value t=21 period m=3 is determined by operation (1) and operation (2). Therefore, each first frame is frame (i-2), frame (i+1
) and frame (i+4). Frame (i-2)
and le2. and frame (i +4), respectively. −2 l=
Since l e21+4 1 =O and less than the threshold value t=2, it becomes a target of thinning. However, frame (i+1
), I e2141 1 = 3, which is greater than the threshold t = 2, so no thinning is performed. Next frame (i+2
), when l e21+2 l =2, the threshold value t=2 or less, and thinning is performed. In this way, speed command v=3
In this case, the feature parameter sets of frame (f-2), frame (i+2), and frame (i+4) are thinned out.

Next, when the speed command v=4, since the content of ■ is positive, it becomes a target of thinning control. By calculations (1) and (2), the threshold value t=32 cycles m=2 is determined. Therefore, each first frame is frame (i-2), frame (i)
, frame (i+2) and frame (i+4). In frame (i-2), frame (i+2) and frame (i+4), l e21-2 l
=”l e2142 1 =2. l e21+a
Since l = O and both are less than or equal to the threshold value t = 3, they become targets for thinning. However, in frame (i) le2
Since +1=8 is larger than the threshold t=3, no thinning is performed.In the next frame (i+1), l e2++
+ l =3, the threshold value t=3 or less, and thinning is performed. In this way, when the speed command v=4, the frame (
i-2), frame (i+1).

The feature parameter sets of frame (i+2) and frame (i+4) are thinned out.

Finally, in the case of speed command v=-4, since the content of ■ is negative, it is a control target for duplicate use. Operation (1) and operation (
2), the threshold value t=3 and the period m=2 are found. Therefore,
Each first frame is frame (i-2), frame (
i), frame (i+2) and frame (i+4).

Therefore, the value of each corresponding availability information e2 is checked, and if it is not negative, it is compared with the threshold value t. frame (i-2),
In frame (i+2) and frame (i+4), the value of the availability information e2 is not negative.

And l e2+-2l =O1le2μ21=2,
Since Ie2I and 1=0, both are less than the threshold value t=3 and are subject to repeated use. However, frame (i
), since e2H=-8, the value of the availability information e2 is negative, so no duplicate use is performed. Further, in the remaining frame (i+1) within the same period, e2+++=3, which is negative, so no duplicate use is performed. In this way, the speed command v
=-4, frame (i-2), frame (i+
2) and frame (i +4).

FIG. 8(B) is a diagram showing the processing results of four types of speed commands V for the availability information e2 of the second embodiment together with the waveform of the original voice, and shows the 111-minute frame (i
The processing results for eight frames from frame (i+5) to frame (i+5) are shown. Similar to Fig. 7(B), the "x" mark represents the thinned out frame, and the par ◎°
The mark indicates a frame that has been used repeatedly. As is clear from Figure 8 (B), when speaking faster than the standard rate, the frame at the time of the rupture of the voiceless plosive consonant "t" is always preserved, and furthermore, the frame at the time of the plosive point and the constant part of the vowel "a" is connected. Frame (i+1) and frame (i+2) of the articulatory connection part
and frame (i+3) as the speaking speed V increases,
Since the frames are adaptively thinned out in order from the frames closest to the vowel stationary part, the synthesized and output speech can maintain its clarity and naturalness regardless of the speaking speed V. In addition, when speaking at a slower rate than the standard rate, the voiceless plosive consonant "t"
Frame (i) and frame (i+1), which exhibit the characteristics of can be kept.

In addition, in the above-mentioned embodiment, a PARCOR coefficient and a PARCO
Although an R-type speech synthesizer is used, it is clear that any synthesis method can be used as long as it expresses speech of one fixed length of time using one set of parameters.

In addition, in the second embodiment, the threshold value t for thinning and overlapping use is
is given in the form of a linear equation with the speed command V as a variable, but it is clear that each speed command V can be given by independent means.

Furthermore, in the second embodiment, information e on whether or not speed control is possible for the articulatory connection part between the plosive consonant part and the following vowel stationary part.
Although the effect of No. 2 has been described, it is clear that the effect of the present invention is not limited thereto, and can be applied to any part of the synthesized and output speech.

[Effects of the Invention] As described above, according to the present invention, the thinning out and redundant use of characteristic parameters, which were conventionally performed simply mechanically, are loaded with information on whether or not speed control is possible, and the magnitude of the speed command V is changed. Since feature parameters are thinned out or duplicated accordingly, it is possible to synthesize clear and natural speech without phonological changes or omissions.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a speech synthesis device according to the first embodiment of the present invention, and Figs. 2 (A) to (C,) show the speech waveforms of ``tai'', a part of ``mitai'', uttered by the same man. Figures 3 (A) to (C) are enlarged views of a portion of each audio waveform in Figures 2 (A) to (C) at the same magnification in the time axis direction, and Figure 4 is a diagram showing the results of the first implementation. FIG. 5(A) is a diagram showing the relationship between the speed command V and the frame thinning or duplication cycle m in the first embodiment; FIG. (B) is a diagram showing the relationship between the speaking rate V, the threshold value, and the period m of thinning or overlapping use in the second embodiment; FIG. 6 is a flowchart showing the speed control procedure of the first embodiment; ) is a diagram showing the processing results for the four types of speed commands V for the acceptability information e of the first embodiment, and FIG. 7(B) is the processing result for the four types of speed commands V for the acceptability information e2 of the second embodiment. FIG. 8(A) is a diagram showing the processing results of four types of speed commands V for the acceptability information e of the first embodiment together with the waveform of the original voice. FIG. 8(B) is a diagram showing the processing results of the second embodiment. A diagram showing the processing results of four types of speed commands V for the example availability information e2 together with the waveform of the original voice. FIG. 9 is a diagram showing the structure of the availability information and feature parameter set of the second embodiment. FIG. 10 is a flowchart showing the speed control procedure of the second embodiment. In the figure, 1...input terminal, 2... central processing unit (cp
U), 3... first storage device, 4... auxiliary storage device,
5...PARCOR type speech synthesizer, 6...D/A converter, 7...amplifier, 8...speaker.

Claims (5)

[Claims] (1) In a speech synthesis device that changes the speech rate by thinning out or duplicating the feature parameters to be synthesized, the feature parameters corresponding to a predetermined length of speech and the speech rate control that corresponds to at least each of the feature parameters. A storage means for storing acceptance/disapproval information, and during speech synthesis,
A speech synthesis device characterized by comprising speed control means for thinning out or duplicating feature parameters only for those feature parameters whose content of the availability information is speed controllable. (2) In a speech synthesis device that changes the speech rate by thinning out or duplicating the feature parameters to be synthesized, is it possible to control the speech rate so that the feature parameters corresponding to a predetermined length of speech correspond to at least each of the feature parameters? a storage means for storing multivalued information; a threshold setting means for setting a threshold according to the speaking speed;
A speech synthesis device comprising a speed control means for thinning out or duplicating feature parameters only for feature parameters whose contents of the multi-valued information are smaller than the threshold value. (3) The storage means is characterized in that it stores maximum multi-value information corresponding to a feature parameter indicating the point of rupture of a plosive consonant, and stores decreasing multi-value information corresponding to subsequent feature parameters. A speech synthesis device according to claim 2. (4) The speech synthesis device according to claim 2, wherein the speed control means does not unconditionally use the feature parameter repeatedly when the multi-valued information has a predetermined sign. (5) The speech synthesis device according to claim 2, wherein the threshold setting means sets a higher threshold as the speaking speed is faster or slower than the standard speed.