JPH06236197A - Pitch pattern generation device - Google Patents

Abstract

PURPOSE:To attain generating a pitch pattern appliable to a device synthesizing an optional voice like regular synthesis and being natural more than usual. CONSTITUTION:This device is constituted so that a unit pattern cut out from the pitch pattern of human vocalization with a proper unit is stored in a pattern storage part 1, and the unit pattern suited to an input attribute is retrieved from the pattern storage part 1 based on the prescribed input attribute in a pattern retrieval part 2, and the unit pattern obtained by the retrieval is arranged on a time base in a pattern arrangement part 3, and the unit pattern arranged on the time base is deformed and the pitch pattern is generated in a pattern deformation part 4, and since the pitch pattern vocalized by a human is used as it is, the pitch pattern capable of expressing natural change and fluctuation, nuance, etc., being hardly generated with a rule is generated.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pitch pattern generator used for a regular speech synthesizer or the like to generate a time pattern of a pitch frequency, that is, a pitch pattern.

[0002]

2. Description of the Related Art Conventionally, a synthetic sound produced by a regular speech synthesizer is mechanical and monotonous, which is obviously different from human speech, lacks naturalness, and gives a feeling of strangeness to a listener. In order to obtain a synthetic sound close to natural voice, it has been recently shown in, for example, Japanese Patent Application Laid-Open No. 2-197897 (hereinafter referred to as Conventional Example 1) or Japanese Patent Application Laid-Open No. 3-139996 (hereinafter referred to as Conventional Example 2). Such a device has been proposed.

That is, in the conventional example 1, a pitch frequency increase pattern is stored for each phoneme type in a mora whose pitch frequency increases according to the accent type of a word or a phrase, and the accent pattern of the word or the phrase is stored. The pitch frequency falling pattern is stored for each type of phoneme in the mora whose pitch frequency drops accordingly, and is stored according to the type of phoneme in the mora whose pitch frequency rises according to the accent type of the word or phrase to be synthesized. The pitch frequency rising pattern is read, the pitch frequency falling pattern is read according to the accent type of the word or phrase to be synthesized, and the stored pitch frequency falling pattern is read according to the phoneme type in the mora. Below the rising frequency pattern and the read pitch frequency A pitch pattern generation device has been proposed that connects a pattern to generate a pitch pattern for an entire word or phrase. In this device, a pitch pattern close to that of natural speech is generated in a mora where the pitch frequency level changes. Intended to do.

Further, in the second conventional example, in a voice editing / synthesizing device in which voices are stored in units of prosodic words and pitch patterns can be arbitrarily controlled, only accent components are stored in pitches when storing each prosodic word. Incidentally, there has been proposed a voice edit synthesizer that generates a pitch component corresponding to the sentence at the time of sentence synthesis and also generates a pitch pattern by superimposing an accent component on the phrase component. , It is intended to generate a pitch pattern in which intonation as a whole sentence becomes natural.

[0005]

As described above, the conventional example 1
Also, any of the devices shown in Conventional Example 2 is intended to obtain a synthetic sound close to natural speech. However, in the pitch pattern generation device of the first conventional example, the pattern to be used is limited to only the portion where the pitch rises or falls according to the accent type, so the unnaturalness of other portions cannot be eliminated. In addition, since this pattern is classified only by the type of phoneme at the position where the pitch frequency changes from "high" to "low" or vice versa, the same pattern is always used when changing with the same phoneme. There is a problem that unnaturalness remains because it is used as a monotonous voice and the surrounding phonological environment and the presence or absence of prominence are not taken into consideration.

Further, since the device of the second conventional example relates to the voice edit synthesizer, it is necessary to store the accent components corresponding to all the prosodic words to be used. Since it consists of a combination of words and clauses, it is practically impossible to comprehensively store all prosodic words. In addition, if the position in the sentence, the context, the intonation, the prominence, the speech rate, etc. change, it cannot be used as it is. Therefore, there is a problem that it is difficult to apply it to a device that synthesizes an arbitrary voice such as rule synthesis.

The present invention solves such a problem and is also suitable for application to a device for synthesizing an arbitrary voice such as rule synthesizing, and a pitch pattern for generating a more natural pitch pattern than the conventional one. It is intended to provide a generator.

[0008]

In order to achieve the above-mentioned object, the invention according to claim 1 is a pattern storing means for storing a unit pattern cut out from a pitch pattern of human utterance in appropriate units, A pattern search means for searching a unit pattern matching the input attribute based on a predetermined input attribute, a pattern arranging means for arranging the unit pattern obtained by the search on a time axis, and a unit arranged on the time axis. And a pattern deforming means for deforming the pattern. As described above, by using the pitch pattern uttered by a human as it is, it is possible to generate a pitch pattern capable of expressing a natural change, fluctuation, nuance, etc., which is difficult to generate by the rule.

In the invention according to claim 2, the pitch pattern is important when a unit pattern including one ascending part and one descending part of an accent phrase is cut out and used as a unit pattern. It is possible to generate a pitch pattern in which patterns of accent phrases, which are various units, are almost preserved. When a unit pattern including one ascending part or one descending part of an accent phrase is used as a unit pattern, a pitch pattern can be generated by a simple connection process. Also, when using a unit pattern that includes the flat part of the accent phrase,
It is possible to save fluctuations in the stationary part and generate a natural pitch pattern. Also, when the characteristic pitch change at the end of the phrase is used as a unit pattern, various intonations that are difficult to generate with rules can be expressed. In addition, when a pattern showing the pitch change rate in a specific range is used as a unit pattern, it is possible to store the outline of the changing part and the fluctuation of the steady part, which is difficult to generate with rules. A natural pitch pattern can be generated. Further, when a pattern showing a pitch curvature in a specific range is used as a unit pattern, the other portion can be interpolated by a straight line, so that the processing amount can be reduced.

According to the third aspect of the present invention, when the vector quantized unit pattern is stored and used, the capacity of the pattern storage means can be greatly reduced, and the cost of the device can be reduced. It is possible to reduce the size and increase the processing speed.

Further, in the invention according to claim 4, when a unit pattern is searched from a phonological environment over a plurality of moras in the vicinity of a pitch change portion, the reproducibility of the local change of the pattern due to the phonological environment is improved, and the unit pattern is more precise. It is possible to generate various pitch patterns. Further, when a unit pattern is searched for from prominence information, inflection of synthetic speech, which has been apt to be monotonous in the past, is conspicuous, and intelligibility and naturalness can be improved. Also,
When searching for a unit pattern from an intonation style, it is possible to accurately reproduce an intonation that is difficult to express by rule generation.

Further, according to the invention of claim 5, a unit pattern is searched by a neural network, whereby a pattern close to the learned human utterance pattern can be reproduced and the naturalness is improved. Can be made.

In the invention according to claim 6, when the unit pattern is arranged by the position of the accent phrase boundary by the pattern arranging means, the rising start point of the accent phrase can be almost reproduced and the intelligibility is improved. Can be made. Further, when the unit pattern is arranged by using the position of the accent nucleus, the descent start point of the accent phrase can be almost reproduced, and the intelligibility can be improved. Further, when the unit pattern is arranged by using the boundary position such as the phoneme, syllable, or mora at the end of the phrase, the intonation can be effectively expressed and the naturalness can be improved. Further, by arranging the unit patterns using the phonological environment information in the vicinity of the pitch change portion, the time structure of the pitch pattern can be reproduced more precisely and the naturalness can be improved.

In the invention according to claim 7, the pattern deforming means is adapted to expand and contract the flat part of the accent phrase in the time axis direction, and the pitch changing part is not deformed in the time axis direction. The time structure of the pitch changing portion can be preserved, and a stable and natural pitch pattern can be generated. Moreover, the rising part and the descending part of the accent phrase are deformed in the pitch axis direction, and the flat part is not deformed in the pitch axis direction.
The height, fluctuation, etc. of the constant pitch portion can be preserved, and the pattern expression including prominence etc. can be realized and the naturalness can be improved.

Further, in the invention described in claims 8 and 9, pattern connecting means for connecting the unit patterns transformed by the pattern transforming means to generate a series of phrase patterns is further provided. In the method, when interpolating between unit patterns with a straight line,
The amount of processing can be reduced, speeding up and cost reduction can be achieved, and when interpolating between unit patterns with a cubic curve, an approximate shape similar to a human pitch change can be expressed, which is natural. A pitch pattern can be generated.

The invention according to claims 10 to 16 is characterized in that a pitch pattern characteristically appearing in a specific linguistic situation is used as the unit pattern stored in the pattern storage means. As a result, it is possible to heuristically search for a pattern from linguistic information that is input information for regular speech synthesis, and it is possible to realize efficient and effective pitch pattern generation.

Further, in the seventeenth and eighteenth aspects of the invention, the unit pattern approximated to the broken line is stored in the pattern storage means. At this time, the unit pattern approximated to the broken line is: The data amount of the unit pattern is reduced by being expressed by the expression data of the pitch of the starting point and each section of each polygonal line and the inclination of the polygonal line and stored in the pattern storage means, thereby reducing the memory amount and the processing amount. It is possible to reduce the cost, reduce the cost, and increase the speed. Furthermore, the unit pattern can be shifted in the pitch direction only by increasing or decreasing the starting point pitch, and the expansion and contraction in the time direction can be easily realized by increasing or decreasing the section length, and the unit pattern can be deformed by a simple process. Therefore, high speed and high functionality can be achieved.

According to the nineteenth aspect of the invention, the slope of each of the polygonal lines is expressed by a step value as a time required for the pitch period to change by a unit length. In this case, the pattern searching means When the expression data of the unit pattern stored in the pattern storage means is searched and the unit pattern is reproduced based on the expression data, if the step value is a positive value, the pitch period is increased by a unit length and the step value is When the value is negative, the pitch period is reduced by a unit length. As a result, it is possible to reproduce the unit pattern approximated to the broken line and obtain the actual pitch pattern with a very simple structure such as a counter and a comparator.

According to the twentieth aspect of the invention, the pattern deforming means applies a constant bias to the step value representing the inclination of each polygonal line to smoothly deform the unit pattern in the pitch direction. As a result, the pattern modification processing required for connection between unit patterns and provision of prominence can be realized with a very simple configuration of only the adder.

In the twenty-first aspect of the invention, the pattern transforming means treats the step value by the complement notation of "2", and when the absolute value of the step value as a result of applying the bias exceeds the maximum value, This is the maximum absolute value of the opposite sign. As a result, it is possible to easily and continuously perform the pattern transformation process from the upward-sloping pattern to the downward-sloping pattern, or vice versa.

According to the twenty-second aspect of the invention, when the correspondence between the pitch change amount and the bias is obtained in advance, the pattern deforming means uses the correspondence to deform the unit pattern in the pitch direction. To do. This allows
Bias required for connection between unit patterns and provision of prominence can be obtained very easily without requiring multiplication and division with a large amount of processing, and the processing speed can be increased.

Further, in the invention of claim 23, there is further provided a rule storage means for storing a rule including a condition in association with a unit pattern stored in the pattern storage means, and pattern search is performed. The means is
By checking the condition of the rule stored in the rule storage means for each input language information unit such as mora, a unit pattern satisfying the rule is retrieved from the pattern storage means. By making the unit pattern search process a rule search, maintenance such as addition, deletion, and change of the unit pattern becomes easy.

According to the twenty-fourth aspect of the present invention, the rule condition includes a range of values of a plurality of variables representing the characteristics of the linguistic information unit, and the pattern search means is included in the rule condition. For all variables, when the value of the input linguistic information unit is within the range, the rule is adopted to search the unit pattern. By including the range of the values of a plurality of variables representing the characteristics of the linguistic information unit in the rule condition, the rule search is performed.
It can be done by chin treatment. Further, it becomes easy to convert the rule into external data, and the maintainability of the rule itself can be improved.

According to the twenty-fifth aspect of the invention, the position from the beginning or end of the phrase of the linguistic information unit is used as the variable included in the condition of the rule, or the phrase containing the linguistic information unit is adjacent to the phrase. A boundary type with a phrase is used, accent information of a phrase including the linguistic information unit is used, or phonological information of the linguistic information unit is used, which is included in the rule condition. By using the position from the beginning or end of the linguistic information unit as a variable, express the property of the pitch pattern that changes continuously depending on the position within the phrase, such as gradually descending from the beginning to the end of the phrase. It becomes possible to do. In addition, by using the boundary type of the phrase including the linguistic information unit with the adjacent phrase, it is possible to produce a feeling of beginning and end of the sentence by lowering the end of the sentence more than the reading point position, and to various intonation styles such as question sentences. It becomes possible to express the characteristics of the pitch pattern at the phrase boundary position such as correspondence. Also, by using accent information of phrases including linguistic information units, accent type features such as head height becoming a mountain tilted to the left rather than middle height type, and features depending on accent level such as prominence and secondary accent. It becomes possible to express the characteristics of the pitch pattern. Further, by using the phoneme information in units of linguistic information, it is possible to represent the characteristics of the local pitch pattern seen in a specific phoneme, the influence of the phoneme duration, and the like.

According to the twenty-sixth aspect of the present invention, one rule is associated with a plurality of unit patterns, and when a plurality of unit patterns satisfying the rule are searched for, a unit is given a fluctuation. It is designed to select a pattern. As a result, it is possible to avoid a uniform and mechanical pitch pattern and obtain a more natural pitch pattern.

[0026]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of an embodiment of a pitch pattern generation device according to the present invention. The pitch pattern generation device according to the present embodiment stores a pattern storage unit 1 in which a unit pattern cut out in a predetermined unit from a pitch pattern of natural voice is stored, and a unit pattern conforming to the input attribute based on a predetermined input attribute. A pattern search unit 2 for searching,
A pattern arranging unit 3 for arranging the unit pattern obtained by the search on the time axis, a pattern deforming unit 4 for deforming the unit pattern arranged on the time axis, and a series of phrases by connecting the deformed unit pattern. And a pattern connection unit 5 for generating the pattern.

Here, the unit pattern stored in the pattern storage unit 1 is cut out from a pitch pattern of natural voice (that is, a human voice) in a manner as shown in FIG. 2, FIG. 3, FIG. 4 or FIG. Done. That is, in FIG. 2, the accent phrase A of the pitched pattern PP which is uttered by the sentence
From the starting point of the rising part AU of K, through the flat part AF, to the descending part AD
Unit pattern UNP with the unit DM ₀ up to the end point of
It is designed to cut out. In addition, in FIG. 3, the accent phrase A of the pitch pattern PP that is uttered is uttered.
The unit pattern UNP is cut out from the start point of the rising portion AU of K to the end point of the subsequent flat portion AF and from the start point of the falling portion AD to the end point of the following flat portion AF as units DM ₁ and DM ₂ , respectively. I am supposed to do it. Further, in FIG. 4, the unit pattern UNP is cut out using only the flat portion AF as the unit DM ₃ . Further, in FIG. 5, the pitch change portion having a characteristic that expresses the intonation at the end of the sentence is represented by the unit DM.
_{As 4} , the unit pattern UNP is cut out. 2 to 5, the flat portion AF of the accent phrase refers to both the peak portion from the end point of the pitch rising portion AU to the start point of the falling portion AD and the opposite valley portion. There is. The cutouts shown in FIGS. 2 to 5 are made in relation to attribute information such as the accent nucleus position, but in addition to this, although not shown, a pitch pattern showing a pitch change rate in a specific range. The unit pattern may be cut out in units of sections, or the unit pattern may be cut out in units of pitch pattern sections showing a curvature in a specific range.

In this way, the pattern storage unit 1 stores
Various unit patterns cut out from sentence speech by any one of the above methods are stored, or unit patterns obtained by vector-quantizing these unit patterns by a codebook created in advance are stored as a database. .

Further, the pattern search unit 2 searches the pattern storage unit 1 for a unit pattern matching a predetermined input attribute. As a method of this search, each unit pattern stored in the pattern storage unit 1 is searched. In addition, it is possible to add the attribute information when this is cut out and select the unit pattern having the attribute information most similar to the input attribute. Further, when the unit pattern is cut out and the unit pattern is vector-quantized by the code book created in advance and stored in the pattern storage unit 1, the code of the vector-quantized unit pattern is output and A method of searching for a unit pattern by inputting a predetermined input attribute to a neural network trained by teacher data in which the attribute of time is input can be considered.

The predetermined input attribute used in the pattern search section 2 is information that is considered to affect the pitch pattern among the input information given to the rule synthesizer, and is a typical one. Although there are accent types and the like, the following information is also added to the input attributes in this embodiment. That is, conventionally, it has been known that the types of phonemes in the vicinity before and after the pitch change causes a local change in pitch, and it is considered effective to use this information for pattern search. Add. Further, conventionally, it has been known that the position where prominence is added changes in the height of accent and the like, so it is considered effective to use this information for pattern retrieval, and this is added to the input attribute. Further, when selecting a pattern as shown in FIG. 5, information on the intonation style of the phrase is necessary, and it is considered effective to use this information for the pattern search, and this is added to the input attribute.

Further, the pattern arranging unit 3 uses the position of the accent phrase boundary on the time axis as input information, or the position of the accent nucleus on the time axis as input information, or the phoneme, syllable, or mora at the end of the phrase. The position on the time axis of the unit pattern is determined by using the position on the time axis of the boundary such as or the like as input information or the phoneme environment information in the vicinity of the pitch change portion as input information.

Further, the pattern transformation unit 4 matches the end point of each unit pattern with the start point of the next unit pattern, according to the type of the unit pattern used, etc., the rising portion AU and the flat portion in the accent phrase. Any one or two or all of the AF and the descending part AD are deformed so as to expand and contract in the time axis direction and / or the pitch axis direction.

The pattern connecting section 5 is adapted to interpolate and connect the unit patterns by a straight line or a cubic curve according to the type of the unit pattern used.

Next, the operation of the pitch pattern generator having such a configuration will be described. First, a unit pattern is cut out from a human voice and stored in the pattern storage unit 1 in advance. The unit pattern can be cut out as described above, for example, by the method shown in FIG. 2, FIG. 3, FIG. 4 or FIG.

That is, when the method shown in FIG. 2 is used, a unit pattern can be cut out from the starting point of the rising portion of the accent phrase to the ending point of the falling portion through the flat portion as a unit. As a similar type, it is possible to cut out the accent phrase as a unit as it is, and in this case as well, the pattern of the accent phrase can be stored almost as it is, as shown in FIG. Compared to this, the method has the advantage that phrases with different mora numbers with the same accent type can be treated in the same row.

When the method as shown in FIG. 3 is used, the unit from the start point of the ascending part of the accent phrase to the end point of the following flat part and from the start point of the descending part to the end point of the following flat part are units. The pattern can be cut out. It is also possible to cut out only the ascending part and the descending part as a unit similar to this,
In this case as well, the unit patterns can be arranged substantially only by deciding the position of the starting point of the unit pattern. However, the method shown in FIG. 3 is more natural than the method shown in FIG. There is an advantage that deterioration can be reduced.

When the method shown in FIG. 4 is used, the unit pattern can be cut out by using only the flat portion as a unit. In the document "Study on fluctuations of fundamental period and its model", Omuro Omuro, Hideki Kasuya, Journal of Acoustical Society of Japan, Vol.47, No.12, pp.928-934, 1991, it describes a part of stationary vowels. It has been reported that the slow fluctuation of the slab affects the naturalness, and if the method of FIG.
Such perceptually important fluctuations can be preserved.

When the method shown in FIG. 5 is used, a unit pattern can be cut out in units of characteristic pitch change portions that represent intonation at the end of a sentence. It should be noted that FIG. 5 shows an example of a rising pattern indicating a question. According to the method of FIG. 5, since the intonation pattern at the end of a sentence can be stored as it is, it is possible to express a subtle difference in nuance that is difficult to express by rules.

In the examples of FIGS. 2 to 5, it is premised that the unit pattern cut out in relation to the attribute information such as the accent nucleus position is used, but in addition to this, the pitch change rate in a specific range or It is also possible to use a unit pattern in which a pattern having a curvature is automatically cut out. The former has the advantage that the contour can be saved by storing the part where the pitch changes as a pattern, and the fluctuation can be saved by storing the stationary part as a pattern. Further, in the latter case, the curved portion is stored as a pattern, and the other portion can be interpolated by a straight line, so that the processing amount can be reduced. The unit pattern cut out by the method shown in FIGS. 2 to 5 or any other method may be stored as it is in the pattern storage unit 1, or may be coded by performing vector quantization or the like and stored in the pattern storage unit 1. You may.

In this way, after storing the unit pattern in the pattern storage unit 1, the actual pitch pattern generation process can be started. In this pitch pattern generation processing, a predetermined input attribute is first given to the pattern search unit 2 and a unit pattern suitable for this input attribute is searched from the pattern storage unit 1. Here, in this embodiment,
In addition to a typical accent type, at least one piece of information about the phoneme type information before and after the pitch change, prominence information, and phrase intonation style information is added to the input attribute. There is. When a unit pattern is searched using a phonological environment that spans multiple moras in the vicinity of the pitch change portion among various information of such input attributes, the reproducibility of local changes in the pattern due to the phonological environment is improved, and more detailed It is possible to generate various patterns. Further, when searching for a unit pattern using prominence information, it is possible to improve the intelligibility and naturalness of the synthetic speech, which has tended to be monotonous in the past. Further, when the unit pattern is searched using the intonation style, it is possible to accurately reproduce the intonation, which is difficult to express by the rule-based generation. Further, by searching the unit pattern by the neural network, a pattern close to the learned human utterance pattern can be reproduced, and the naturalness can be improved.

After the optimum unit pattern is searched in this way, the pattern placement section 3 mainly generates the pitch pattern of the searched unit pattern from the positions of the accent kernel and the accent phrase boundary on the time axis. The position on the time axis at, more specifically, the position of the starting point of the unit pattern is determined, and the unit pattern is arranged. More specifically, the pattern arranging unit 3 determines the position of the starting point of the unit pattern at the phoneme boundary position where the pitch level at the symbol level changes, in consideration of the shift that changes depending on the phoneme environment of that part. Here, the position at which the pitch level changes is a phrase boundary or core position for an accent phrase, but for a unit pattern as shown in FIG. The fact that the position is decided based on the point is the document ““ Perception of Japanese Questionable Sentences ”, Ichiro Miura, Miyoko Sugito, Proceedings of Autumn Meeting of ASJ, 1991, I, 2-6-9, p
Since it is known by "p255-256" etc., the boundary position of the phoneme, syllable, mora, etc. at the end of the phrase is used. FIG. 6 is a diagram showing the pattern placement process when the unit pattern shown in FIG. 3 is used.
The lower part shows the position of the phonological boundary and the pitch level at the symbol level, and the unit pattern is arranged at the upper part. In FIG. 6, the dotted line L1 represents the boundary between the ascending portion or the descending portion and the flat portion, and in this case, the position of the starting point of the unit pattern is determined as the thick line L2.

After the position of the unit pattern on the time axis is determined in this way, the pattern transforming section 4 transforms the unit pattern so that it is naturally embedded in the generated pitch pattern. That is, the pattern transformation unit 4 first performs the transformation on the time axis. FIG. 7 is a diagram showing how the unit pattern arranged as shown in FIG. 6 is deformed. That is, in FIG.
Since the unit pattern as shown in FIG. 3 is used, in FIG. 7, the expansion / contraction in the time axis direction is performed so that the start point of each unit pattern coincides with the end point of the previous unit pattern. At this time, it is considered appropriate to expand / contract in the time axis direction only at the flat part, from the physiological constraints of the vocalization mechanism. Here, the specific method of expansion / contraction is not particularly limited, and expansion / contraction also includes data truncation, repetition, and extrapolation.

In this way, after the expansion and contraction in the time axis direction, the deformation on the pitch axis is performed. FIG. 8 is a diagram showing a state in which the pitch pattern deformed in the time axis direction as in FIG. 7 is deformed on the pitch axis. That is, in this example, expansion and contraction in the pitch axis direction are performed so that the start point of each unit pattern coincides with the end point of the previous unit pattern. At this time, since it is appropriate that the ascending part and the descending part change depending on the front-rear connection, it is preferable to perform expansion and contraction in the pitch axis direction only at the ascending part and the descending part.
This makes it possible to preserve the outline and fluctuation of the flat portion. By such a method, the flat portion of the unit pattern also preserves its absolute height. Therefore, the pitch pattern can be reproduced as a unit pattern, including the influence of prominence and the influence of the position in the phrase.

In the above example, since the unit patterns shown in FIG. 3 are used, interpolation processing is not particularly required for connecting the patterns, but unit patterns other than the unit patterns shown in FIG. When used, some kind of interpolation processing is required in the pattern connection unit 5. At this time, when only the ascending part and the descending part are unit patterns, the space between them is linear, and when the flat part is unit pattern as shown in FIG. A pattern closer to a human pitch pattern can be generated.

By the way, in the unit connection type pitch pattern generating apparatus as described above, the problem of what kind of unit pattern is stored in the pattern storage section greatly affects the performance. Therefore, in the present invention, a pitch pattern characteristically appearing in a specific linguistic situation is used as a unit pattern stored in the pattern storage unit, thereby realizing efficient and effective pitch pattern generation. ing. Hereinafter, such a unit pattern will be described in detail.

The continuous voice has a plurality of phrases such as an exhalation paragraph and a sentence, but in normal utterance, the pitch pattern becomes lower as the position in the phrase approaches the end. Therefore, by preparing a plurality of patterns having different positions in the phrase, it is possible to give the above-mentioned phrase feeling without performing complicated processing such as separation / combination of the vocal components. At the end of a sentence, the feeling of the beginning and end of the sentence can be obtained by using a relatively low dedicated pattern regardless of the length of the sentence. In this way, the unit pattern 1
As an example, by using a pitch pattern characteristically appearing at a specific position in a series of phrases such as a breath paragraph and a sentence, it is possible to express a feeling of phrase, a feeling of beginning and end of a sentence, and the like.

The pattern outline of the accent phrase differs depending on the presence or absence of the accent nucleus, the position, and the like. The nucleus-free accent phrase pattern is gentler than that of the nucleus-free one, and the head-height one rises abruptly and then descends gently. Therefore, such a difference can be expressed by preparing a plurality of patterns with different accent types and using them properly. In this way, a more natural pitch pattern can be generated by using, as one of the unit patterns, a pitch pattern characteristically appearing in a specific accent type such as a head height type or a flat type.

Further, by using a relatively high dedicated pattern for the portion to which prominence is applied, it is possible to give a natural prominence feeling without unnatural pitch rise even at the beginning of a phrase. In addition, it is possible to express a secondary accent, which has been difficult in the past, by preparing patterns with different accent levels. In this way, as one of the unit patterns, by using the pitch pattern characteristically appearing at a specific accent level such as a part to which prominence is added or a part where a secondary accent nucleus appears, the prominence and the secondary accent It is possible to express and improve the naturalness.

Further, as one of the unit patterns, it is also possible to use a word such as an interjection, a final particle, an onomatopoeia, a mimetic word, a shout, etc. which is uttered with a unique pitch pattern, or the phrase pitch pattern is used as it is. In this case, the word uttered in a pitch pattern unique to the word can be faithfully reproduced, and the naturalness can be improved.

Further, when reading a number or reading a poem such as a Japanese poem, a constant pitch pattern appears periodically. Therefore, as one of the unit patterns,
By using a characteristic pitch pattern that appears periodically in a number string, a verse, etc., it is possible to deal with the peculiar utterance utterance as described above.

Also, in conversation, "er" and "yeah"
The word itself does not have much meaning, and it is often the case that the intention is expressed mainly by differences in prosody. Therefore, as one of the unit patterns, a characteristic pitch pattern that represents a specific intention mainly by a prosody independently of words or phonemes is used, whereby natural conversation can be efficiently output.

FIG. 9 is a diagram showing a specific example of processing for retrieving these unit patterns from input language information and synthesizing pitch patterns. In the example of FIG. 9, in the pattern storage unit 1, [P01] sentence-end flat plate type, [P03] sentence-end flat plate type,
[P41] Prominence, [P42] Secondary accent,
Is it [P51]? , [P61] number, [P71] surprisingly, seven types of unit patterns are stored in advance, and as input language information, "Eh? (Unexpected) / 920917? Again / menu number "(Prominence) / Enter." Is input, the pattern search unit 2 searches the pattern storage unit 1 for a unit pattern string corresponding to the input language information. As a result, "[P71] Eh? [P61] 92 [P6
1] 09 [P61] 17 [P51]? [P01]
It is possible to search for a unit pattern string of "P41", enter [P03] menu number again [P42]. Next, by arranging, deforming, and connecting the unit pattern strings thus searched in the pattern arranging section 3, the pattern deforming section 4, and the pattern connecting section 5, a natural pitch pattern as shown in the lower part of FIG. 9 is obtained. Can be generated.

As described above, by using the pitch pattern characteristically appearing in a specific linguistic situation as the unit pattern stored in the pattern storage unit 1,
A unit pattern can be heuristically searched from the input language information, and by arranging, transforming, and connecting the unit patterns, a natural pitch pattern can be efficiently and effectively generated by a relatively simple process. it can.

When the pitch pattern of natural voice is stored as a unit pattern as it is, a larger amount of data is required than when it is stored as a generation rule or model parameter, which increases the memory capacity. It is expected that real-time processing will be difficult due to the increase in processing amount. Therefore, some kind of ingenuity is required to avoid it.

The inventor of the present application, in order to prevent real-time processing from becoming difficult due to an increase in memory capacity and an increase in processing amount, a unit pattern cut out from a pitch pattern of natural voice in an appropriate unit as it is. In the form of
It was further devised to store the unit pattern in the pattern storage unit 1 by polygonal line approximation instead of storing it in the pattern storage unit 1.

FIG. 10 shows an example in which a unit pattern which is originally curved is approximated by a broken line. In the example of FIG. 10, one unit pattern is approximated to seven broken lines, and sections (segments) S1 to S for each broken line.
It is divided into 7. Here, the unit pattern can be expressed by the pitch of the starting point, the section length (segment length) of each section, and the step value indicating the slope of the polygonal line (line segment) of each section. The pattern can be stored in the pattern storage unit 1 in this expression format. The step value is the time required for the pitch cycle to change by one sample, that is, the pitch update cycle, and takes a positive (+) value when the slope of the polygonal line rises to the right as shown in FIG. , FIG. 11 (b)
As shown in, the negative (-) value is assumed.

As described above, the originally curved unit pattern is approximated by a polygonal line, and this is expressed by the pitch of the starting point, the section length of each section and the step value, and stored in the pattern storage unit 1. Unit pattern
The amount of data stored in the pattern storage unit 1 can be greatly reduced as compared with the case where the data is stored as it is. Further, the unit pattern can be shifted in the pitch direction only by increasing / decreasing the starting point pitch, and the expansion / contraction in the time axis direction can be performed by increasing / decreasing the section length, so that the pattern can be easily deformed. Can be done.

However, when the above expression is used, the pattern search unit 2 searches the pattern storage unit 1 for the data in the above expression form of a predetermined unit pattern and reads it. When this is done, it is necessary to reproduce (generate) this in the form of a unit pattern (that is, in the form of a polygonal line), but this unit pattern generation processing is realized by a simple configuration as shown in FIG. it can. That is, FIG.
In the circuit shown in (1), the step value of the section is input to reproduce (generate) a polygonal line in a certain section, and the absolute value of the input step value is obtained (that is, the step value is negative). When the absolute value of the step value is set and the sample clock at a predetermined time interval is counted and the count value of the sample clock is equal to the absolute value of the step value, A step counter 12 that outputs "1", a comparator 13 that determines whether the input step value is positive or negative,
The comparator 13 determines positive (+), the AND circuit 14 which outputs "1" when the step counter 12 outputs "1", and the comparator 13 determines negative (-), "1" from the step counter 12
AND circuit 15 that outputs "1" when is output
And the output "1" from the AND circuit 14 is counted in the positive (+) direction, the output "1" from the AND circuit 15 is counted in the negative (-) direction, and the count value is output as the pitch. This circuit is composed of a pitch counter 16 and a step value (pitch update cycle) by this circuit.
When is positive, the pitch period can be increased by the unit length, and when the step value is negative, the pitch period can be decreased by the unit length. In other words, the pitch of the starting point is set to the initial value of the pitch counter 16, and the pitch is output from the pitch counter 16 of the above circuit until the section length (segment length) is reached in the time axis direction in one section. By sequentially performing the sections, it is possible to reproduce (generate) a unit pattern approximated by a polygonal line. In this way, the unit pattern can be reproduced and the actual pitch pattern can be generated.

In the pattern arranging section 3, each unit pattern searched and generated by the pattern searching section 2 is arranged at a predetermined position, and in the pattern deforming section 4, a natural pitch pattern is arranged. In order to obtain the above, the processing for deforming each unit pattern arranged by the pattern arranging section 3 is performed. At this time, the inclination is changed to connect the adjacent unit patterns or to give prominence. Need to be adjusted. In the present invention, since the data representation as described above is adopted, this adjustment processing is performed by, for example, adding a constant bias to the step value of the entire transition section (see FIG. 10) by an adder (not shown). No)
It can be easily realized by adding.

FIG. 13 shows an example of such adjustment processing for one unit pattern. Referring to FIG. 13, by adding a constant bias value to each step value in each section of one unit pattern and changing the step value (by changing the slope of each polygonal line), the unit pattern is changed. Can be smoothly deformed in the pitch direction. For example, as in the example of FIG.
Step value (before transformation) ST of each section of unit pattern
By adding a negative bias value BIAS to _A and setting it as ST _B , the unit pattern UNP _A before deformation can be deformed like UNP _B. That is, the start point can be changed by a predetermined amount (bias amount) without changing the end point. In this way, the pattern modification processing required for connection between unit patterns and provision of prominence can be realized with a very simple configuration of only the adder.

In order to adjust the natural inclination in this way, it is necessary to continuously change from the upward slope to the horizontal and downward slope, which is the maximum step value. This can be achieved by continuously changing the value (positive maximum absolute value) and the minimum value (negative maximum absolute value). Such processing can be performed by setting the step value (pitch update cycle) to "2". When the absolute value of the step value of the result of adding the bias exceeds the maximum value,
This can be easily realized by setting the maximum absolute value of the opposite sign.

As described above, the pattern deforming section 4 can smoothly deform the unit pattern in the pitch direction by applying a constant bias to the unit value of each section of one unit pattern. In the section 4, the bias required for connection between adjacent unit patterns and provision of prominence must be calculated in real time. This calculation usually requires multiplication and division, which complicates the process and may become a bottleneck in performing the entire pitch pattern generation process in real time. To avoid this problem, the bias table as shown in FIG. 14 in which the correspondence between the pitch change amount (for example, the difference between the starting point pitches of the original pattern and the generated pattern or the starting point pitch of the generated pattern itself) and the bias is calculated in advance. Is provided for each unit pattern, and when one unit pattern is deformed in the pitch direction by a predetermined pitch change amount, a bias corresponding to the pitch change amount may be read from the bias table of this unit pattern and used. This eliminates the need to calculate the bias each time and does not hinder the real-time processing. Specifically, as shown in FIG. 15, the end point of each pattern is not changed, and the pitch change amount of the start point is calculated so that the start point of each pattern becomes the end point of the previous pattern, and this is dealt with. By reading the bias from the bias table as shown in FIG. 14 and using the bias to transform the original pattern into a generated pattern, connection with the previous pattern can be easily performed.

As a modified example of the present invention, in order to facilitate maintenance such as addition, deletion, and change of unit patterns in the pattern search unit 2, as shown in FIG. 16, unit patterns are input for each input language information unit. It is also possible to further provide a rule storage unit 7 that stores a plurality of rules for searching.

In this case, as the rule stored in the rule storage unit 7, a rule composed of a condition including a plurality of items and their value ranges and a unit number can be used. Here, the item has various attributes for each mora, and specifically, the position from the beginning of the phrase, the boundary type with the adjacent phrase, the accent type, the accent level, the phoneme, etc. are considered to be effective.

Next, a concrete example of the pattern search process in the pitch pattern generating apparatus having the structure shown in FIG. 16 will be described with reference to FIG.
This will be described with reference to FIGS. FIG. 17 is a diagram showing various unit patterns stored in the pattern storage unit 1. In the example of FIG. 17, four types of unit patterns are unit numbers “1”, “2”, “3”, “. 4 ”is stored. Further, FIG. 18 shows the rule storage unit 7.
2 is a diagram showing a configuration example of a rule stored in FIG.
In the example of No. 8, the rule uses four attributes as items: position Pos from accent phrase head, position Mora from accent phrase head, accent type Acc, position of accent nucleus Core, and four types of units. The numbers "1", "2",
A range of four attribute values is assigned to each of "3" and "4".

In FIG. 16, the pattern search unit 2
Referring to the rule storage unit 7 in FIG. 18, the pattern storage unit 1
The unit pattern corresponding to the input language information is searched from the four types of unit patterns stored in. Specifically, assuming that the input language information unit is a mora and the processing is performed in a mora unit, and when an input mora string as shown in FIG. 19 is input, first, for each mora,
Find the value of all items. Focusing on a certain mora, the pattern search unit 2 refers to the rule storage unit 7, and for all items included in the rule condition,
Check to see if the range contains a value for each item in this mora. When it is determined that the value of each item of this mora is included in the range of the item of a certain unit number, this mora satisfies the condition of the rule, and the pattern search unit 2 determines that the rule The unit pattern corresponding to the unit number can be searched from the pattern storage unit 1. If there is no rule satisfying the condition, it is determined that there is no unit pattern arranged at the mora position, and the process for the next mora is performed.

For example, in the input mora sequence of FIG.
The mora "ro" of "melon" has the item values "0", "4", "1", and "0", and the unit number of the rules stored in the rule storage unit 7 is " 4 ”is satisfied, so the pattern search unit 4
Can search the unit pattern of unit number “4” from the pattern storage unit 1.

As described above, by including the range of the values of a plurality of variables representing the characteristics of the linguistic information unit in the rule condition, the rule search can be performed by the routine processing.
Further, it becomes easy to convert the rule into external data, and the maintainability of the rule itself can be improved.

Further, by using the position from the beginning or end of the linguistic information unit as a variable included in the rule of the rule, the position in the phrase is gradually lowered from the beginning to the end of the phrase. This makes it possible to express the characteristics of the pitch pattern that changes continuously. In addition, by using the boundary type of the phrase including the linguistic information unit with the adjacent phrase, it is possible to produce a feeling of beginning and end of the sentence by lowering the end of the sentence more than the reading point position, and to various intonation styles such as question sentences. It becomes possible to express the characteristics of the pitch pattern at the phrase boundary position such as correspondence. Also, by using accent information of phrases including linguistic information units, accent type features such as head height becoming a mountain tilted to the left rather than middle height type, and features depending on accent level such as prominence and secondary accent. It becomes possible to express the characteristics of the pitch pattern. Further, by using the phoneme information in units of linguistic information, it is possible to represent the characteristics of the local pitch pattern seen in a specific phoneme, the influence of the phoneme duration, and the like.

In the above example, the rules and the unit patterns have a one-to-one correspondence, but one rule may have a plurality of unit numbers. In this case, one unit pattern can be selected by giving some "fluctuation" among them. This makes it possible to generate pitch patterns that are not exactly the same even from the same input, avoid a uniform mechanical pitch pattern, and obtain a more natural pitch pattern.

[0071]

As described above, according to the first aspect of the invention, the unit pattern cut out from the pitch pattern of the human utterance in an appropriate unit is stored in the pattern storage means, and the pattern search means, Based on a predetermined input attribute, a unit pattern matching the input attribute is searched from the pattern storage means, the pattern arranging means arranges the unit pattern obtained by the search on the time axis, and the pattern transforming means, the time axis. The unit pattern arranged above is transformed to generate a pitch pattern, and since the pitch pattern uttered by a human being is used as it is, natural changes, fluctuations, nuances, etc. which are difficult to generate by rules are generated. It is possible to generate a pitch pattern capable of expressing

Further, in the invention of claim 2, when a unit pattern including one ascending part and one descending part of an accent phrase is cut out and used as a unit pattern, the pitch pattern is important. It is possible to generate a pitch pattern in which patterns of accent phrases, which are various units, are almost preserved. When a unit pattern including one ascending part or one descending part of an accent phrase is used as a unit pattern, a pitch pattern can be generated by a simple connection process. Also, when using a unit pattern that includes the flat part of the accent phrase,
It is possible to save fluctuations in the stationary part and generate a natural pitch pattern. Also, when the characteristic pitch change at the end of the phrase is used as a unit pattern, various intonations that are difficult to generate with rules can be expressed. In addition, when a pattern showing the pitch change rate in a specific range is used as a unit pattern, it is possible to store the outline of the changing part and the fluctuation of the steady part, which is difficult to generate with rules. A natural pitch pattern can be generated. Further, when a pattern showing a pitch curvature in a specific range is used as a unit pattern, the other portion can be interpolated by a straight line, so that the processing amount can be reduced.

According to the third aspect of the present invention, the vector quantized unit pattern is stored in the pattern storage means, and when this is used, the capacity of the pattern storage means can be greatly reduced. The cost of the device can be reduced, the device can be downsized, and the processing speed can be increased.

According to the fourth aspect of the present invention, by retrieving the unit pattern from the phonological environment over a plurality of moras in the vicinity of the pitch change portion, the reproducibility of the local change of the pattern due to the phonological environment is improved, and A fine pitch pattern can be generated. Also, when searching for a unit pattern from information on prominence,
The inflection of synthetic speech, which used to tend to be monotonous in the past, tends to be distracting, and intelligibility and naturalness can be improved. Further, when searching for a unit pattern from an intonation style, it is possible to accurately reproduce an intonation that is difficult to express by the rule-based generation.

According to the fifth aspect of the present invention, by searching the unit pattern by the neural network, a pattern close to the learned human utterance pattern can be reproduced, and the naturalness can be improved. .

Further, in the invention according to claim 6, when the unit pattern is arranged by using the position of the accent phrase boundary, the rising start point of the accent phrase can be almost reproduced, and the intelligibility can be improved. Further, when the unit pattern is arranged by using the position of the accent nucleus, the descent start point of the accent phrase can be almost reproduced, and the intelligibility can be improved. Also, the phoneme at the end of the phrase,
When the unit patterns are arranged by using the boundary positions such as syllables and mora, intonation can be effectively expressed and the naturalness can be improved. Further, by arranging the unit patterns using the phonological environment information in the vicinity of the pitch change portion, the time structure of the pitch pattern can be reproduced more precisely and the naturalness can be improved.

According to the seventh aspect of the present invention, the flat portion of the accent phrase is expanded and contracted in the time axis direction, and the pitch change portion is not deformed in the time axis direction. Since the temporal structure can be preserved and a pattern that does not deviate from physiological constraints is guaranteed, a stable and natural pitch pattern can be generated. Also, the rising and falling parts of the accent phrase are deformed in the pitch axis direction, and the flat part is not deformed in the pitch axis direction, so that the height and fluctuation of the constant pitch part can be preserved. , It is possible to express patterns including prominence and improve the naturalness.

According to the inventions of claims 8 and 9,
By interpolating the unit patterns with a straight line, the processing amount can be reduced, and the speed and cost can be reduced. By interpolating the unit patterns with a cubic curve, the human pitch change can be prevented. You can express a similar outline,
A natural pitch pattern can be generated.

According to the tenth aspect of the invention, since the pitch pattern characteristically appearing in a specific linguistic situation is used as the unit pattern stored in the pattern storage unit, the input language information is used. A heuristic pattern search can be performed from, and a pitch pattern can be efficiently and effectively generated.

According to the eleventh aspect of the invention, as one of the unit patterns, an exhalation paragraph, a sentence, etc.
By using a pitch pattern that characteristically appears at a specific position in a series of phrases, the feeling of phrase can be achieved without complicated processing such as separation / synthesis of voice components.
It is possible to express the feeling of the beginning and end of a sentence and to simplify the processing.

According to the twelfth aspect of the present invention, as one of the unit patterns, by using a pitch pattern characteristically appearing in a specific accent type such as a head height type or a flat type, presence or absence of an accent nucleus is detected. The pattern outline of the accent phrase that differs depending on the position, etc. can be expressed by preparing a plurality of patterns with different accent types and using them properly, thereby improving the naturalness.

According to the thirteenth aspect of the present invention, as one of the unit patterns, the pitch characteristically appearing at a specific accent level such as a portion to which prominence is added and a portion where a secondary accent nucleus appears. By using a pattern, prominence and secondary accents can be expressed, and the naturalness can be improved.

According to the fourteenth aspect of the present invention, as one of the unit patterns, a word such as an interjection, a final particle, an onomatopoeic word, a mimetic word, or a shout, which is uttered in a pitch pattern peculiar to the word, is used? Alternatively, by using the pitch pattern of the phrase as it is, it is possible to faithfully reproduce a word uttered in a pitch pattern unique to the word, and improve the naturalness.

According to the fifteenth aspect of the present invention, as one of the unit patterns, a characteristic pitch pattern that appears periodically in a number string, a verse, etc. is used to read a number or a verse. It is possible to deal with peculiar turning voices such as when reading aloud, and the versatility of the device can be enhanced.

According to the sixteenth aspect of the present invention, as one of the unit patterns, a characteristic pitch pattern that expresses a specific intention mainly by prosody independently of words and phonemes is used, thereby making it unique to conversation. It is possible to efficiently and naturally output a conversation even when the intention is mainly expressed by a difference in prosody such as "er" and "yeah".

According to the seventeenth and eighteenth aspects of the present invention, the unit pattern approximated to the broken line is stored in the pattern storage means. At this time, the unit pattern approximated to the broken line is stored. Is expressed by expression data of the pitch of the starting point and each section of each polygonal line and the inclination of the polygonal line, and is stored in the pattern storage means, thereby reducing the data amount of the unit pattern, and reducing the memory amount and processing. The amount can be reduced, and the cost and speed can be reduced. Furthermore, the unit pattern can be shifted in the pitch direction only by increasing or decreasing the starting point pitch, and the expansion and contraction in the time direction can be easily realized by increasing or decreasing the section length, and the unit pattern can be deformed by a simple process. Therefore, high speed and high functionality can be achieved.

According to the nineteenth aspect of the present invention, the inclination of each of the polygonal lines is expressed by a step value as a time required for the pitch period to change by a unit length. In this case, the pattern search means. Searches the expression data of the unit pattern stored in the pattern storage means, and when the unit pattern is reproduced based on the expression data, if the step value is a positive value, the pitch period is increased by the unit length, and the step If the value is negative, the pitch period is reduced by the unit length.Therefore, the unit pattern of the polygonal line approximation is reproduced by a very simple structure such as a counter and a comparator, and the actual A pitch pattern can be obtained.

According to the twentieth aspect of the invention, the pattern deforming means applies a constant bias to the step value representing the inclination of each polygonal line, so that the unit pattern is smoothly deformed in the pitch direction. , The pattern modification processing required for connection between unit patterns and the addition of prominence can be realized with a very simple configuration of only the adder.

According to the twenty-first aspect of the invention, the pattern transforming means treats the step value by the complement notation of "2", and when the absolute value of the step value as a result of applying the bias exceeds the maximum value. Since this is the maximum absolute value of the opposite sign, the pattern rising to the right to the pattern descending to the right.
Pattern continuously and comfortably,
It is possible to easily realize the process of transforming the image.

According to the twenty-second aspect of the present invention, when the correspondence between the pitch change amount and the bias is obtained in advance, the pattern deforming means uses the correspondence in the pitch direction of the unit pattern. Since the transformation is performed, the bias required for connection between unit patterns and the addition of prominence can be calculated very easily without the need for multiplication and division, which requires a large amount of processing, and the processing speed can be increased. be able to.

Further, according to the invention of claim 23, there is further provided a rule storage means for storing a rule including a condition in association with the unit pattern stored in the pattern storage means, The pattern search means is configured to search the unit condition satisfying the rule from the pattern storage means by checking the condition of the rule stored in the rule storage means for each input language information unit such as mora, By making the unit pattern search process a rule search, maintenance such as addition, deletion, and change of the unit pattern becomes easy.

According to the twenty-fourth aspect of the present invention, the rule condition includes the range of the values of a plurality of variables representing the characteristics of the linguistic information unit, and the pattern search means sets the rule condition in the rule condition. For all variables included, when the value of the input linguistic information unit is within the range, the unit pattern is searched by adopting the rule, and the characteristics of the linguistic information unit are specified in the rule condition. By including the range of the values of a plurality of variables that represent, the rule search can be performed by the routine processing. Further, it becomes easy to convert the rule into external data, and the maintainability of the rule itself can be improved.

According to the twenty-fifth aspect of the invention, as the variable included in the condition of the rule, the position from the beginning or the end of the phrase of the language information unit is used, or the phrase including the language information unit is used. The boundary type of the adjacent phrase is used, or the accent information of the phrase including the linguistic information unit is used, or the phoneme information of the linguistic information unit is used. By using the position from the beginning or end of the linguistic information unit as a variable to be included, the pitch pattern changes gradually depending on the position within the phrase, such as a gradual decrease from the beginning to the end of the phrase. Can be expressed. In addition, by using the boundary type of the phrase including the linguistic information unit with the adjacent phrase, it is possible to produce a feeling of beginning and end of the sentence by lowering the end of the sentence more than the reading point position, and to various intonation styles such as question sentences. Pitch pattern at phrase boundaries, such as correspondence
It is possible to express the characteristics of the computer. Moreover, by using the accent information of the phrase including the linguistic information unit,
It is possible to express the characteristics of pitch patterns such as accent type features such as a head-height type that is tilted to the left rather than the middle-high type and accent level features such as prominence and secondary accents. Further, by using the phoneme information in units of linguistic information, it is possible to represent the characteristics of the local pitch pattern seen in a specific phoneme, the influence of the phoneme duration, and the like.

According to the invention of claim 26, 1
When a plurality of unit patterns are associated with one rule and a plurality of unit patterns satisfying the rule are searched for, one unit pattern is given with fluctuation, so that a uniform mechanical pitch pattern is selected. Therefore, it is possible to avoid the noise and obtain a more natural pitch pattern.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of a pitch pattern generation device according to the present invention.

FIG. 2 is a diagram for explaining an example of cutting out a unit pattern.

FIG. 3 is a diagram for explaining an example of cutting out a unit pattern.

FIG. 4 is a diagram for explaining an example of cutting out a unit pattern.

FIG. 5 is a diagram for explaining an example of cutting out a unit pattern.

FIG. 6 is a diagram for explaining how to determine the position of the unit pattern on the time axis.

FIG. 7 is a diagram for explaining how the unit pattern is deformed on the time axis.

FIG. 8 is a diagram for explaining how the unit pattern is deformed on the pitch axis.

FIG. 9 is a diagram showing a specific example of a process of retrieving a unit pattern from input language information and synthesizing a pitch pattern.

FIG. 10 is a diagram showing an example of polygonal line approximation of a unit pattern.

11A and 11B are diagrams for explaining a step value.

FIG. 12 is a diagram showing an example of a unit pattern generation circuit.

FIG. 13 is a diagram for explaining processing for adjusting the inclination of the unit pattern.

FIG. 14 is a diagram showing an example of a bias table.

FIG. 15 is a diagram for explaining pattern modification processing.

16 is a diagram showing a modification of the pitch pattern generation device shown in FIG.

FIG. 17 is a diagram showing various unit patterns stored in a pattern storage unit.

FIG. 18 is a diagram showing a configuration example of rules stored in a rule storage unit.

FIG. 19 is a diagram showing an example of an input mirror sequence.

[Explanation of symbols]

 1 pattern storage unit 2 pattern search unit 3 pattern arrangement unit 4 pattern transformation unit 5 pattern connection unit 7 rule storage unit

Claims (26)

[Claims] 1. A pattern storage means for storing a unit pattern cut out from a human utterance pitch pattern in appropriate units, and a unit adapted to the input attribute or input language information based on a predetermined input attribute or input language information. It has pattern search means for searching a pattern, pattern arrangement means for arranging the unit pattern obtained by the search on the time axis, and pattern deforming means for deforming the unit pattern arranged on the time axis. A pitch pattern generation device characterized by: 2. The pitch pattern generating device according to claim 1, wherein the unit pattern includes a pattern including an ascending part and an descending part of an accent phrase as one unit and an accent phrase. A pattern that is cut out as one unit that includes one ascending part or a descending part, a pattern that is cut out as a unit a pitch pattern section that includes a flat part of an accent phrase, and a characteristic pitch change at the end of a phrase. , A pattern cut out as a unit, and a pattern cut out as a unit of a pitch pattern section showing a rate of change in a specific range.
A pitch pattern generation device characterized in that at least one of patterns obtained by cutting out a pitch pattern section showing a curvature of a specific range as one unit is used. 3. The pitch pattern generation device according to claim 1, wherein the pattern storage means stores a cut-out unit pattern or a vector-quantized version of the cut-out unit pattern. A pitch pattern generation device characterized by the above. 4. The pitch pattern generation device according to claim 1, wherein the pattern search means has an input attribute,
At least one of phonological environment information over a plurality of moras near the pitch change portion, prominence information in or near the unit pattern, and intonation style information of the phrase including the unit pattern is used. A pitch pattern generation device characterized by: 5. The pitch pattern generating device according to claim 1, wherein the pattern searching means is adapted to determine an optimum unit pattern by a neural network from an input attribute. 6. The pitch pattern generating device according to claim 1, wherein the pattern arranging means arranges the position of the accent phrase boundary on the time axis, the position of the accent kernel on the time axis, the phoneme at the end of the phrase, the syllable, the mora, etc. The position of the unit pattern on the time axis is determined by using at least one of the position of the boundary of the time axis on the time axis and the phoneme environment information near the pitch change portion as input information. Generator. 7. The pitch pattern generating device according to claim 1, wherein the pattern modifying means expands or contracts the flat portion in the time axis direction when the unit pattern has flat portion boundary information in the accent phrase. When the unit pattern is modified and the unit pattern has boundary information of the rising part and the descending part in the accent phrase, the rising part and the descending part are expanded and contracted in the pitch axis direction. Pitch pattern generator. 8. The pitch pattern generating device according to claim 1, further comprising pattern connecting means for connecting the unit patterns deformed by the pattern deforming means to generate a series of phrase patterns. A pitch pattern generation device characterized by: 9. The pitch pattern generation device according to claim 8, wherein the pattern connection means connects by interpolating unit patterns with a straight line or connecting by interpolating unit patterns with a cubic curve. A pitch pattern generation device characterized by the above. 10. The pitch pattern generating device according to claim 1, wherein the unit pattern stored in the pattern storage means is a pitch pattern characteristically appearing in a specific linguistic situation. Pattern generator. 11. The pitch pattern generation device according to claim 10, wherein a pitch pattern characteristically appearing at a specific position in a series of phrases, such as an exhalation paragraph or a sentence, is used as one of the unit patterns. Characteristic pitch pattern generator. 12. The pitch pattern generating device according to claim 10, wherein a pitch pattern characteristically appearing in a particular accent type such as a head height type or a flat plate type is used as one of the unit patterns. Pitch pattern generator. 13. The pitch pattern generation device according to claim 10, wherein as one of the unit patterns, a characteristic pattern appears at a specific accent level such as a part to which prominence is added and a part where a secondary accent nucleus appears. A pitch pattern generation device characterized in that a pitch pattern is used. 14. The pitch pattern generation device according to claim 10, wherein a word uttered in a pitch pattern peculiar to the word is used as one of the unit patterns, such as interjections, final particles, onomatopoeia, mimetic words, and shouts. Or
Alternatively, the pitch pattern generation device is characterized in that the pitch pattern of the phrase is used as it is. 15. The pitch pattern generation device according to claim 10, wherein a characteristic pitch pattern periodically appearing in a number string, a verse, or the like is used as one of the unit patterns. . 16. The pitch pattern generation device according to claim 10, wherein a characteristic pitch pattern that represents a specific intention mainly by prosody independently of words or phonemes is used as one of the unit patterns. Pitch pattern generator. 17. The pitch pattern generation device according to claim 1, wherein the pattern storage means stores a unit pattern approximated to a broken line. 18. The pitch pattern generation device according to claim 17, wherein the polygonal line approximated unit pattern is expressed by expression data of a pitch of a start point, each section of each polygonal line, and an inclination of the polygonal line, and the pattern storage. A pitch pattern generation device characterized by being stored in a means. 19. The pitch pattern generation device according to claim 18, wherein the inclination of each polygonal line is represented by a step value as a time required for the pitch period to change by a unit length, and in this case, the pattern search. The means searches the expression data of the unit pattern stored in the pattern storage means, and when reproducing the unit pattern based on the expression data, if the step value is a positive value, increases the pitch period by a unit length, A pitch pattern generation device, wherein when the step value is a negative value, processing for reducing the pitch period by a unit length is performed. 20. The pitch pattern generating device according to claim 19, wherein the pattern deforming means applies a constant bias to a step value representing the inclination of each polygonal line to smoothly deform the unit pattern in the pitch direction. A pitch pattern generating device characterized by being applied. 21. The pitch pattern generation device according to claim 20, wherein the pattern transformation means treats the pitch update period by a complement notation of "2", and the absolute value of the step value as a result of applying the bias has a maximum value. A pitch pattern generation device characterized in that when it exceeds, it is set to the maximum absolute value of the opposite sign. 22. In the pitch pattern generation device according to claim 20 or 21, when the correspondence between the pitch change amount and the bias is obtained in advance, the pattern deforming means uses the correspondence. A pitch pattern generation device characterized in that a unit pattern is deformed in the pitch direction. 23. The pitch pattern generation device according to claim 1, further comprising rule storage means for storing a rule including a condition in association with a unit pattern stored in the pattern storage means. Cage,
The pattern search means searches the pattern storage means for a unit pattern satisfying the rule by checking the condition of the rule stored in the rule storage means for each input language information unit such as a mora. A pitch pattern generation device characterized in that. 24. The pitch pattern generation device according to claim 23, wherein the condition of the rule includes a range of values of a plurality of variables that represent characteristics of a linguistic information unit, and the pattern search means is a rule. A pitch pattern generation device characterized in that a unit pattern is searched for by adopting the rule when the value of the input linguistic information unit is within the range for all variables included in the condition . 25. The pitch pattern generation device according to claim 24, wherein a position from the beginning or end of the linguistic information unit is used as a variable included in the rule condition, or the linguistic information unit is included. A pitch pattern generation device characterized by using a boundary type of a phrase with an adjacent phrase, using accent information of a phrase including the linguistic information unit, or using phonological information of the linguistic information unit. 26. In the pitch pattern generation device according to claim 23, when one rule is associated with a plurality of unit patterns and a plurality of unit patterns satisfying the rule are retrieved, a fluctuation is given. A pitch pattern generation device characterized by selecting one unit pattern.