JPH113095A - Speech synthesis device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To decrease a calculation amount at a time of speech synthesis and also to obtain a highly naturalized speech. SOLUTION: A speech unit information addition part 3 puts a speech unit vector on each frame of a natural speech pattern. An interframe distance calculation part 4 calculates a distance between each frame. An inter-pattern connecting frame authorization part 5 assumes a frame pair having a distance within a threshold to be the inter-pattern connecting frame pair. A transition weight calculation part 7 connects the inter-pattern connecting frame pair, adding them the transition weight according to the distance, forming a speech pattern connecting network, and storing it in a speech pattern connecting network storage part 8. At a speech synthesis, an optimal path on the above-mentioned speech pattern connecting network is searched to obtain a synthetic parameter. Thus, speech data are connected with each other in a frame unit smaller than a phoneme and a highly naturalized speech is obtained. Moreover, a calculation amount is decreased at the speech synthesis by setting a frame transition path only between acoustically connectable frames.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizer which performs a speech synthesis process by appropriately editing a characteristic spectrum obtained by analyzing a natural speech pattern, and in particular, realizes a more natural sound output. To reduce the amount of calculation and the amount of stored data during the speech synthesis processing operation.
The present invention relates to a speech synthesizer that gives a flexible adaptive ability to a generated vocabulary.

[0002]

2. Description of the Related Art Conventionally, as one of the text-to-speech synthesis methods, natural speech data obtained from a speaker is stored in a database for each speech unit, and a speech unit suitable for a target synthetic sentence is stored. There is an editing / synthesis method in which audio is generated by joining. In the conventional editing / synthesizing method, the speech data of speech units required for speech synthesis is selected from a database and used. It is necessary to create a database that is associated with segment data.

[0003] That is, for natural speech data recorded from a speaker, first, a phoneme, a short chain of phonemes (C (consonant) V
(Vowels), VC, VCV, etc.), speech unit segmentation and labeling using a priori speech unit segmentation information based on phoneme information such as syllables, words or phrases as speech symbols are performed. Then, a database in which the obtained speech unit segment information and label information are added to the speech data of the speech unit and registered is created in advance.

[0004] In this case, the connection of voice data that can complement each other in smaller voice segments determined by voice symbols used as label information of the voice segments is not considered. In addition, the connection point between the speech units is fixed to the end point of the segmented speech unit, or the search range and search direction (order) of the connection point are set depending on the type of the speech symbol. Is done. The selection of the speech unit and the search for the connection point are performed at the time of speech synthesis in consideration of the continuity of the phoneme environment and the spectrum.

[0005]

However, in the conventional text-to-speech synthesis based on the editing and synthesizing method,
There are the following problems. That is, as described above, the synthesized speech unit data is obtained by searching a database based on label information provided under some linguistic information or phoneme information, and thus the unit of the speech unit and the linguistic information or phoneme We need to be able to clearly express our relationship with information. Therefore, a speech unit based on a unit smaller than a speech unit such as a phoneme is not used.

In other words, in the conventional editing / synthesizing method,
Mutual complementation and connection of speech data in units smaller than the unit of the synthesized speech unit determined by the speech unit set by the speech symbol are not tested. Therefore, there is a problem that even if there is a chain of more effective feature vectors in the smallest frame unit, it is impossible to evaluate the effective feature vector during speech synthesis. This means that the editing and synthesizing method requires a relatively large storage capacity, which is an obstacle to reducing the scale of the speech synthesizing system.

[0007] For example, "Rule synthesis method for automatically generating synthesis units," Nakatsuta, Hamada, etc.
P57-64 "or" Ito, Nakatsuta, Hirokawa, etc., "Speech synthesis method using waveform COC unit" Proceedings of the Acoustical Society of Japan, October 1993, 2-8-17 P253-254 "
In the synthetic speech segment creation method using segment clustering using a distance scale based on a feature spectrum in a speech segment as shown in (1), cluster division is performed by labeling a unit of a phoneme or a concatenation of a plurality of phonemes. This is done based on the information. Therefore, the frame connection other than the speech unit connection point limited by the label information is not evaluated.

Further, in the above-mentioned editing and synthesizing method, it is necessary to select a necessary speech unit from a database on the basis of the label of the above-mentioned speech symbol by measuring compatibility each time.
Therefore, the amount of calculation required for searching the database and verifying the connection explosively increases according to the storage amount of the synthesized speech unit. This is because all speech unit candidates having the same speech symbol label are tested in the same column based on the label information at the time of speech unit selection. This is because even the test of the combination of is performed, an extra calculation process occurs.

"Abe, Takeda, Sakazaka, etc." Study on Connection Method of Speech Synthesis Units According to Phonological Environment "Speech Research Group Material SP89
In −66 P17 to 22 ″, the degree of freedom in the search range of the speech unit connection point is improved. However, the same problem as described above occurs in the selection of speech units having the same phoneme environment, and the synthesized speech When hypothesizing the unit sequence (subsequence), the candidate of the dividing point is determined by the speech symbol given to this synthesized speech unit segment and the dividing point candidate group used for forming the hypothesis. There is no degree of freedom in selection, and the search range of the connection point is limited by a limited combination of division points.

[0010] Furthermore, "Iwahashi, Kaigi, Sakasaka, etc."
ment Selection for ConcatinativeSynthesis Based on
Spectral Distortion Minimization ”, IEICE TRANS.VO
LE.76-A NO.11, NOV. 1993, PP1942-1948 ”, even if the hypothesis method of the candidate of the dividing point is used, the speech is measured based on the statistical phoneme continuity in the phonetic symbol. Since segment segmentation is performed, it is not possible to solve the problem that there is a speech segment connection in which a natural and smoother spectrum connection is performed than a segmented speech segment candidate pair even at an unselected division point.

[0011] This is an important problem for an editing and synthesizing method in which a speech unit is created from natural speech and speech synthesis is performed.

[0012] As described above, the search path of the route that can follow the spectrum transition in the feature spectrum space of the voice and the connection point between the path units is limited by the voice symbol used to describe the voice unit. . Therefore, in order to fully cover all spectral transitions, it is not only necessary that the samples of the feature spectrum fully fill the utterance space, but also that the interval defined by the speech unit (that is, Spectral transitions in other directions that can be taken within a path that can follow spectral transitions on the feature spectrum space of the voice need to be sufficient.

When there is a shortage of possible spectrum transitions in the section defined by the speech unit, it is necessary to further increase natural speech data to increase the number of possible spectrum transition paths. On the contrary, since redundant feature spectrum information is increased, the amount of unnecessary storage is increased, and the amount of calculation for searching for a connection point is greatly affected.

Further, since the correspondence between the speech unit and the speech symbol is fixed, another speech symbol is added to the speech unit obtained from the articulation state in which the acoustic features are very similar (that is, similar). In this case, the use of the feature spectrum of the speech unit as a synthesis parameter is not evaluated and examined, and there is a problem that the articulation transition information cannot be effectively used.

Further, when considering the use of a speech synthesizer in which the occurrence of sentences and phoneme chains (that is, the path distribution of spectrum transition) is limited depending on the use conditions, depending on the number of vocabularies, the speech synthesis by the analysis and synthesis method is considered. In some cases, using a text-to-speech synthesis system using a text-to-speech synthesis method is more advantageous on a system scale than using a device. However,
With the conventional text-to-speech synthesis method using edit synthesis, it is difficult to reduce the risk of connection between speech units that are not particularly preferable even in a limited sentence space, and the quality is improved when the use is limited. There is also a problem that it is difficult to achieve this.

Accordingly, an object of the present invention is to effectively utilize a feature spectrum obtained from natural speech for the generation of synthesis parameters, reduce the amount of calculation at the time of speech synthesis, suppress the enlargement of the system, and perform editing and synthesis. An object of the present invention is to provide a speech synthesizer capable of obtaining a synthesized speech having a high naturalness by utilizing features.

[0017]

According to a first aspect of the present invention, there is provided a speech synthesizer comprising: an input unit;
Frame analysis of natural speech obtained from a speaker is performed, and for a plurality of natural speech patterns in which each frame is labeled with a speech unit, a pair of frames belonging to different natural speech patterns and having similar acoustic features is identified. A voice pattern connection network storage unit that connects to each other and stores a voice pattern connection network formed by adding a transition weight to the connection between each frame, and a voice unit label of each frame constituting the voice pattern connection network. By reference, when tracing on the voice pattern connection network in the order of the voice unit label sequence based on the linguistic information input from the input unit, a search is made for an optimum route in which the cumulative total of the transition weights has the maximum value, and the search is performed. That outputs a sequence of feature spectra of frames on the optimal route as a composite parameter sequence A sequence search unit, a prosody generation unit for generating a prosody control parameter based on the input linguistic information, and a synthesized speech based on the synthesized parameter sequence from the parameter sequence search unit and the prosody control parameter from the prosody generation unit Is provided.

According to the above configuration, the synthesis parameters used for speech synthesis are connected to frames that can be acoustically connected between natural speech patterns formed by analyzing natural speech frames, and transition weights are applied to connections between frames. It is stored as an added voice pattern connection network. And
At the time of voice synthesis, an optimum route on the voice pattern connection network is searched to obtain synthesis parameters.
In this way, the synthesis parameters are smoothly connected in units of frames smaller than phonemes or the like, and a synthesized speech having high naturalness is generated.

According to a second aspect of the present invention, in the voice synthesizing apparatus according to the first aspect of the present invention, a natural voice pattern for storing a natural voice pattern obtained by frame analysis of a natural voice obtained from a speaker is stored. A storage unit, a speech unit information adding unit that attaches the speech unit label to each frame of the natural speech pattern, an inter-frame distance calculation unit that determines a distance between characteristic spectra of frames belonging to different natural speech patterns, By examining whether or not the distance is equal to or less than a predetermined value, it is possible to determine whether a pair of frames having the interframe distance can be a pair of inter-pattern connection frames connecting the different natural voice patterns. A connection frame verification unit, between a pair of frames that can be the above-mentioned pattern connection frame pair and between already connected frame pairs The transition weights at the time of transition are calculated, and the transition weights are added to the connections between the respective frames of the voice pattern connection network formed by connecting the inter-pattern connection frame pairs in a plurality of natural voice patterns to each other. It is characterized by including a transition weight calculator for storing in the pattern connection network storage.

According to the above configuration, a speech pattern connection network in which acoustically connectable frames are connected between natural speech patterns formed by frame analysis of natural speech and transition weights are added to connections between the frames. But,
It is automatically created based on the natural voice pattern stored in the natural voice pattern storage.

According to a third aspect of the present invention, in the voice synthesizing apparatus according to the second aspect of the present invention, the inter-frame distance calculating unit calculates a distance between characteristic spectra of two frames to be calculated. Then, a correction distance obtained by performing correction based on a power difference between the two feature spectra, a non-voice portion of the two feature spectra, a change direction of the two feature spectra, or a variance of the two feature spectra is obtained.

According to the above arrangement, the characteristic difference between the two feature spectra to be calculated is emphasized in the distance based on the feature spectrum obtained by the frame analysis for the natural speech obtained from the speaker. The distance between frames is calculated by performing such correction. Therefore, an inter-frame distance is calculated such that an acoustically similar frame pair can be accurately determined.

According to a fourth aspect of the present invention, in the voice synthesizing apparatus according to the second aspect of the present invention, when all the frames of the natural voice pattern are clustered, the voice unit label of the number of frames dropped into each cluster is used. A frame unit for storing a speech unit vector having a ratio of each element as an element in association with the cluster, wherein the speech unit information adding unit refers to the frame cluster storage unit and By attaching a speech unit vector associated with the cluster to which each frame belongs to the frame, a speech unit label is attached to each frame.

According to the above configuration, the correspondence between the frame as the node of the voice pattern connection network and the voice unit label is not one-to-one and not discrete,
It is made in the form of a vector having the passage probabilities for a plurality of voice unit labels as elements. Therefore, a plurality of articulation transition information represented by the nodes is effectively used, and a transition state to many speech unit labels is expressed.

According to a fifth aspect of the present invention, in the voice synthesizing apparatus according to the second aspect of the present invention, the inter-pattern connection frame testing unit determines that the inter-frame distance is equal to or less than another predetermined value smaller than the predetermined value. In the case of (1), there is provided a frame merging means for merging a pair of frames having the inter-frame distance into one frame.

According to the above configuration, a pair of frames that are acoustically very similar are merged into one frame, redundant feature spectrum information is deleted, and the storage capacity is reduced.

According to a sixth aspect of the present invention, in the voice synthesizing apparatus according to any one of the first to fifth aspects of the present invention, the voice unit label of each frame constituting the voice pattern connection network is referred to. Then, when the voice unit label follows the voice pattern connection network in the order of the voice unit label sequence of the known specific voice pattern, the optimum route in which the total of the transition weights exhibits the maximum value is searched, A network evaluation unit is provided which outputs a sequence of audio unit labels attached to frames on the route as a recognition result of the specific audio pattern.

According to the above configuration, the matching rate between the speech unit label string of the speech pattern whose speech unit label is known and the recognition result (speech unit label string), and the value of the cumulative transition weight presented by the above-mentioned optimal route are represented by: The suitability of the above voice pattern connection network for voice synthesis is evaluated.

According to a seventh aspect of the present invention, there is provided the voice synthesizing apparatus according to any one of the first to sixth aspects, wherein the network editing for changing a transition weight between frames in the voice pattern connection network is performed. It is characterized by having a part.

According to the above configuration, when performing speech synthesis with limited vocabulary, inappropriate interframe connections existing in the speech pattern connection network are easily corrected,
The quality of the synthesized speech based on the limited linguistic information is improved.

Here, the linguistic information is a concept including a text, a syntax structure, and a morpheme.

[0032]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, the present invention will be described in detail with reference to the illustrated embodiments. FIG. 1 is a block diagram of the speech synthesizer according to the present embodiment. This speech synthesizer is a text speech synthesizer having a speech pattern connection network creation device.

First, the voice pattern connection network creating device 1 will be described. The voice pattern connection network creation device 1 includes a natural voice pattern storage unit 2,
It has a voice unit information addition unit 3, an inter-frame distance calculation unit 4, an inter-pattern connection frame test unit 5, a transition weight calculation unit 7, a voice pattern connection network storage unit 8, and a frame cluster codebook storage unit 9.

The natural voice pattern storage unit 2 creates a frame by analyzing a natural voice recorded in advance and labeling each frame to indicate a voice unit (hereinafter, the attached label is called a voice unit label). Stored natural voice pattern groups. Further, a vector book is created by performing vector quantization on all the frames of each natural voice pattern, and the frames falling into the cluster to which each code is allocated are totalized for each voice unit label and normalized for each cluster. . Then, a vector (Equation (1)) having each normalized value as an element is stored in the frame cluster codebook storage unit 9 together with the codebook as a speech unit vector of the cluster.

(Equation 1)

Here, k-means clustering or the like is used to create the speech unit vector, and the relationship between the average distortion Dm (1 ≦ m ≦ K) in the cluster and the threshold Dr is Dr ≧ Dm (3) The number of divisions is increased by setting K ← K + 1 until the following condition is satisfied, and adjustment is made so that the difference in the articulation state is best classified.

The voice unit information adding unit 3 uses the codebook stored in the frame cluster codebook storage unit 9 and the voice unit vector of each code cluster to read out the natural voice pattern stored in the natural voice pattern storage unit 2. A voice unit vector is attached to each frame of the voice pattern. Thus, each frame is associated with the audio unit label.

The inter-frame distance calculator 4 calculates the distance between temporally consecutive frames in two natural voice patterns as described below, and uses the calculated distance d as the inter-pattern connection frame tester 5. To send to. In calculating the distance d between frames in this case, the feature spectrum (waveform, cepstrum, LPC (linear prediction analysis method) or LSP (line spectrum pair), etc.) obtained by the frame analysis is not used as it is, The distance in which the space where the above-mentioned characteristic spectrum exists is subjected to the correction inclination is used.

An example of the correction distance calculation formula in that case is shown in formula (4).

(Equation 2)

In the first term of the numerator in the correction term of equation (4), the distance between frames having different powers P is obtained by taking the absolute value of the logarithmic difference between the powers Pu and Pv of the respective characteristic spectra u and v. Enlarge. In the second term, the distance of a frame having a wide utterance spectrum space in a high frequency range without having a formant structure such as unvoiced fricatives and unvoiced affricates is corrected. Further, in the first term of the denominator, the characteristic spectra u, v
Is normalized by taking the inner product of the change in the norm and by the norm, thereby correcting the distance when the change directions of the characteristic spectra u and v are the same. A delta cepstrum or the like may be used as a characteristic spectrum for this correction. In the fourth term, the variance σ in the cluster c to which each of the characteristic spectra u and v belongs
_{Using c (u)} and σ _{c (v)} , the overall distance inclination is corrected by the distribution of the spectrum.

The inter-pattern connection frame verification unit 5 includes:
It is determined whether or not the value of the normalized distance d _norm (hereinafter simply referred to as distance d) between frames calculated by the inter-frame distance calculator 4 is within a predetermined value. Then, all the frame pairs exhibiting the distance value d within the predetermined value are extracted as a pattern connection frame pair. The processing from the calculation of the distance between frames performed by the inter-frame distance calculation unit 4 to the verification of the pair of inter-pattern connection frames performed by the inter-frame connection frame verification unit 5 includes “Ito, Kiyama, Kojima, Seki, Oka et al. “Refen for spotting in any section of the standard pattern
ce Interval-free continuous DP ”IEICE technical report,
SP95-34, P73-80 ", and spotting between arbitrary sections of a frame synchronization type can also be applied.

Further, the inter-pattern connection frame test section 5 has a frame merging means 6. Then, the frame merging unit 6 merges the characteristic spectra of the closer frame pairs among the inter-pattern connection frame pairs, thereby reducing the storage capacity and increasing the speed by reducing the number of paths during the speech synthesis operation. The transition weight calculation unit 7 connects the inter-pattern connection frame pairs extracted as described above. Then, a transition obtained by adding a transition weight b determined based on the distance between the two frames to the transition between the feature spectra corresponding to the connected frame pair is stored as a voice pattern connection network storage unit as voice pattern connection network information. 8 is stored. The processing up to this is the voice pattern connection network generation processing performed by the voice pattern connection network creation device 1.

FIG. 2 is a flowchart of a voice pattern connection network generation processing operation performed by the voice pattern connection network generation device 1. Hereinafter, the voice pattern connection network generation processing operation will be briefly described once again. Here, a natural voice pattern in which a natural voice is subjected to frame analysis and a voice unit label is attached to each frame is stored in the natural voice pattern storage unit 2, and the code book and the voice unit generated based on the natural voice pattern are stored. The vector is stored in the frame cluster codebook storage unit 9.

In step S1, the speech unit information adding unit 3 refers to the code book and the speech unit label stored in the frame cluster codebook storage unit 9 and stores the code book and the speech unit label in the natural speech pattern storage unit 2. The speech unit vector is attached to each frame of all the natural speech patterns. In step S2, the one-to-one inter-frame distance calculation unit 4 reads one natural sound pattern from the group of natural sound patterns to which the sound unit vector is attached.

In step S3, one of the natural voice patterns different from the natural voice pattern read in step S1 from the natural voice pattern storage unit 2 is read by the inter-frame distance calculation unit 4 as a reference pattern. In step S4, the interframe distance calculation unit 4 sets a time interval between temporally consecutive frames of the reference pattern frame read in step S3 and the natural voice pattern frame read in step S1. The distance d is given by the equation
It is calculated by (4). In step S5, the inter-pattern connection frame determining unit 5 determines in step S4
By examining whether or not the value of the inter-frame distance d calculated in is within a predetermined value, it is determined whether or not the two frames match a pair of inter-pattern connection frames connecting different natural voice patterns. . As a result, when it is determined that they match, the process proceeds to step S6, and when it is determined that they do not match, the process proceeds to step S8.

In step S6, the transition weight calculation section 7 calculates a transition weight b for transition between the inter-pattern connection frame pairs based on the distance d between the inter-pattern connection frame pairs. In step S7, the calculated transition weight b is associated with the inter-pattern connection frame pair by the transition weight calculation unit 7 and stored in the voice pattern connection network storage unit 8 as voice pattern connection network information. . In step S8, the inter-frame distance calculation unit 4 determines whether or not all the natural voice patterns stored in the natural voice pattern storage unit 2 have been read as reference patterns. As a result, if it has been read, the process proceeds to step S9. If it has not been read, the process returns to step S3 and shifts to the process with the next reference pattern.

In step S9, the inter-frame distance calculation unit 4 determines whether or not an unprocessed natural voice pattern exists in the natural voice pattern storage unit 2. As a result, if it exists, the process returns to the step S2 and shifts to the processing of the next natural voice pattern.
Proceed to. In step S10, the frame merging means 6 of the inter-pattern connection frame judging section 5 sets a frame interval smaller than another predetermined value smaller than the predetermined value used for determining the suitability of the inter-pattern connection frame pair in step S5. An inter-pattern connection frame pair exhibiting a distance d (hereinafter, referred to as a close connection frame pair) is selected. Then, the characteristic spectra corresponding to the close connection frame pair are merged. In step S11, the transition weight calculator 7
Thereby, the transition weight of the frame after the above-mentioned combination is calculated, and the inter-voice-pattern connection network storage unit 8 is updated with the combined frame and the transition weight. After that, the voice pattern connection network generation processing operation ends.

Using the speech pattern connection network created in this way, a language analysis unit 11, a speech unit label string conversion unit 12, a parameter string search unit 13, a speech synthesis unit 14,
The text-to-speech synthesis is performed by the prosody generation unit 15 and the speech output unit 16. Hereinafter, the text-to-speech synthesis will be described.

The language analysis unit 11 includes the text input unit 1
Language analysis is performed on the text given from 0, and the analysis result is sent to the voice unit label string conversion unit 12. The voice unit label sequence conversion unit 12 converts the input text into a voice unit label sequence using the same voice unit label group as the voice unit label group used in the voice unit information adding unit 3. In this case, in the conventional text-to-speech synthesis apparatus, a label string in units of linguistic information or phoneme information is uniquely determined with respect to the input text. By using it, it is possible to connect in units of frames smaller than phonemes, so it is possible to give multiple input speech unit label string candidates to the input text, including handling of unvoiced vowels and presence / absence of unusual connection Becomes In this case, a plurality of synthesized speech candidates are created based on the plurality of speech unit label string candidates obtained as described above, and a synthesized speech candidate that can produce the smoothest speech is selected from these synthesized speech candidates. Then, speech synthesis is performed.

Further, the voice unit label string conversion unit 12
As in the conventional text-to-speech synthesis apparatus, the continuation of the same label is compressed into one label to set a single speech unit length or to set the continuation length of a plurality of speech units. Instead, it is expanded to a frame-synchronous audio unit label sequence in which the number of repetitions of a frame having the same audio unit label and the variation range of the number of repetitions can be set. The speech unit label string converted by the speech unit label string conversion unit 12 is sent to the parameter string search unit 13.

The parameter sequence search unit 13 is the one having the highest likelihood in the voice pattern connection network stored in the voice pattern connection network storage unit 8.
Feature spectrum sequence (that is, connected naturally and smoothly)
By searching for a path giving the (synthesis parameter sequence), a synthesis parameter sequence for driving the speech synthesis unit 14 is generated.

FIG. 3 schematically shows a part of the above-described voice pattern connection network structure, and also schematically shows a path search process for one synthesis parameter sequence. Also, FIG.
FIG. 4 schematically shows paths of other synthesis parameter strings in the same voice pattern connection network structure as FIG. Hereinafter, an example of the path search operation of the composite parameter sequence performed by the parameter sequence search unit 13 will be described with reference to FIGS.

As described above, FIGS. 3 and 4 are cut-out portions of the voice pattern connection network. The voice pattern A (chain of circles) obtained from the utterance / abe / and the utterance / bbf / Voice pattern B (□
This is the part where the connection between the patterns occurs between the chain of marks) and the voice pattern C obtained from the utterance / dbc / (the chain of marks). In the figure, a solid line indicates an “invariant path (in the figure, simply described as“ invariant ”)” transitioning to the next frame in the same voice pattern, and a dashed line indicates a transition to the next frame of another voice pattern. A transition path (in the figure, simply described as “transition”) ”is shown. A transition weight is added to each “invariant path” and “transition path”.
(In FIG. 3 and FIG. 4, only a part of the transition weight is shown.) The speech unit vector is added to each node (feature spectrum of each frame) of the speech pattern connection network. For example, in the third frame of the voice pattern B, the voice unit vector (0.1 (/ a /),
0.6 (/ b /) and 0.3 (/ d /)) are added. This means that among the frames belonging to the cluster to which the third frame of the voice pattern B belongs, 10% are frames of the voice unit label / a /, 60% are frames of the voice unit label / b /,
It indicates that 30% of the frames are of the audio unit label / d /, which is not changed and each audio unit label / a /, / b /, /
The probability that the frame of d / transits to the third frame of the voice pattern B is shown.

From the above, the transition weight “0.6” of the transition path from the second frame of the voice pattern A to the third frame of the voice pattern B and the voice unit in the voice unit vector of the third frame of the voice pattern B The product of the label / b / with the element value "0.6" is the sound unit label of the third frame
In the case of b /, the probability of transition from the second frame of the audio pattern A to the third frame of the audio pattern B is obtained.
Hereinafter, the "probability" is used as the likelihood of the transition path. Therefore, the route in which the cumulative likelihood when the voice unit label string of the given text passes through each route on the voice pattern connection network is searched for, and each node is searched.
By connecting the characteristic spectrum (synthesis parameter) associated with the (frame), it is possible to generate a natural synthesized voice without any acoustic discomfort.

Hereinafter, a speech unit label string / aa obtained by developing the path of the speech patterns A, B, and C in frame synchronization.
When bbbbf / is a speech unit label string obtained from a text, a search for an optimal path through which the synthesis parameter string passes will be considered.

The partial likelihood calculation when the voice unit label sequence / aabbbbb / passes through the voice pattern connection network is performed as follows. As shown in FIG. 3, the initial frame is the first frame of the audio pattern A, and the audio unit vector has only an element of / a /. Then, the path that transitions to the second frame is only the invariant path to the same voice pattern A, and the transition weight of the invariant path is defined as 1.0, so the cumulative likelihood up to the second frame / a / is , The initial value “1.0”, and the sum of the transition weight of the invariant path of 1.0 and the element value of / a / in the audio unit vector of the second frame of the audio pattern A of 0.8, p (i | I = 2) = 1.0 + 1.0 × 0.8 = 1.8 (5) where i: frame number (maximum value N).

Next, from the second frame / a / to the third frame
There are two paths at the time of transition to / b /: an invariant path to the audio pattern A and a transition path to the audio pattern B. Therefore, first, an invariant path to the audio pattern A is considered. In this case, the accumulated likelihood p (2) up to the second frame
= 1.8, the product of the transition weight 1.0 of the invariant path to the third frame / b / and the element value 0.6 of / b / in the audio unit vector of the third frame of the audio pattern A is added. Thus, the cumulative likelihood p (3) up to the third frame is p (3) = 2.4. Next, a transition path to the voice pattern B is considered. In this case, the cumulative likelihood up to the second frame p (2) = 1.8
And the product of the transition weight 0.6 of the transition path to the third frame / b / and the element value 0.6 of / b / in the speech unit vector of the third frame of the speech pattern B are added to the third The accumulated likelihood p (3) up to the frame is p (3) = 2.16. Therefore, the transition from the second frame / a / to the third frame / b / is more likely to take the invariant path to the audio pattern A than to the transition path to the audio pattern B. And the cumulative likelihood P (3) up to the third frame is P
(3) = 2.4.

Similarly, the cumulative likelihood p (4) when transitioning from the third frame / b / to the fourth frame / b / is obtained.
However, in this case, the cumulative likelihood p (4) when the invariant path to the voice pattern A is taken is p (4) = 3.2. On the other hand, the cumulative likelihood p (4) when a transition path to the voice pattern B is taken is p (4) = 3.21. Therefore, the third
A transition path is taken as a transition from frame / b / to the fourth frame / b /, and the accumulated likelihood P (4) up to the fourth frame is P
(4) = 3.21.

Thus, the final term p (N), p (i + 1) ← p (i) + b _{n (i + 1)} n (i + 1)... (6) in the recurrence equation represented by equation (6) (1) = n (1) p (i): cumulative likelihood up to the i-th frame in the path of the synthesized voice pattern n (i): the evaluation target element of the voice unit vector of the i-th frame in the path of the synthesized voice pattern Value b _{n (i)} : weight of transition to the next frame i, that is, the cumulative likelihood value up to the last frame obtained by equation (7)

(Equation 3) Is selected as the optimal route.

Therefore, in the case of the example shown in FIG. 3, the cumulative likelihood p (7) up to the seventh frame which is the last frame is p (7) = 1.0 + (1.0 × 0.8) + (1.0 x 0.6) + (0.9 x 0.9) + (0.8 x 0.9) + (1.0 x 0.4) + (1.0 x 0.9) = 5.23… (8) Path (thick line in FIG. 3: chain line is a transition path)
Is the optimal path indicating the maximum likelihood.

FIG. 4 shows a path through which the text / speech unit label string / bbbbccc / obtained in the same manner as described above passes. In this case, the accumulated likelihood p (7) up to the seventh frame which is the last frame is p (7) = 0.5 + (1.0 × 0.5) + (1.0 × 0.7) + (1.0 × 0.9) + (1.0 × 0.2) + (1.0 × 0.7) + (1.0 × 1.0) = 4.5 (9) (Bold line in FIG. 4: dashed line is transition path)
Is the optimal path indicating the maximum likelihood.

A sequence of characteristic spectra corresponding to nodes on the voice pattern connection network through which the optimum route selected as described above passes is obtained as a synthesis parameter sequence.

As described above, when a plurality of speech unit label string candidates are given to the input text by the speech unit label string conversion unit 12, a plurality of synthesis parameters corresponding to each speech unit label string candidate Column candidates will be obtained. In that case, the maximum likelihood of each of the synthesis parameter sequence candidates may be compared, and the synthesis parameter sequence candidate exhibiting the largest maximum likelihood may be determined to be the target synthesis parameter sequence that can be uttered most smoothly.

The synthesis parameter sequence thus obtained is sent to the speech synthesis unit 14. Then, the speech synthesis unit 14 controls the prosody based on the prosody control parameters generated by the prosody generation unit 15 based on the result of the language analysis by the language analysis unit 11, and synthesizes the synthesized speech based on the synthesis parameter sequence. The generated synthesized voice signal is sent to the voice output unit 16. Then,
The sound output unit 16 drives the speaker 17 based on the synthesized sound signal to output a synthesized sound.

In the present embodiment, the speech synthesis method used in the speech synthesis unit 14 is not particularly limited.
The voice synthesis method relates to a frame analysis method used when generating a natural voice pattern to be stored in the natural voice pattern storage unit 2 of the voice pan connection network creating apparatus 1. In the case of the present embodiment, application of a zero-phase waveform superposition method, a linear sound synthesis method, a method using a linear LSP, a cepstrum synthesis method, or the like can be considered as the sound synthesis method. At this time, it does not relate to the language analysis by the language analysis unit 11 or the prosody generation method by the prosody generation unit 15.

Finally, the speech pattern connection network suitability evaluation processing by the network evaluation unit 18 and the speech pattern connection network editing by the network editing unit 19 will be described. Natural speech can be recognized using the speech pattern connection network stored in the speech pattern connection network storage unit 8. Therefore, the network evaluation unit 18 recognizes a voice with a known utterance using the voice pattern connection network, and evaluates the suitability of the voice pattern connection network based on the recognition result.

The suitability evaluation of the voice pattern connection network by the network evaluation section 18 is performed as follows.
That is, the equation used when calculating the cumulative likelihood p (N) up to the last frame by the parameter string search unit 13
The value n (i) of the evaluation target element of the speech unit vector in the i-th frame in (7) is calculated as the characteristic spectrum of the (i-1) -th frame of the input speech and the i-th frame in the speech pattern connection network. Consider the score defined by the distance between them. Then, using the above equation (7), the optimum route indicating the maximum score on the voice pattern connection network for the input voice whose voice unit label is known is obtained, and the node on the voice pattern connection network through which the optimum route passes is determined. The input speech is recognized based on the column of the assigned selected (passed) speech unit labels. In this case, as the recognized speech unit label string matches the input speech and the value of the maximum score is larger, the synthesized parameter string obtained based on the obtained optimal route is closer to the transition of natural speech. This indicates that the voice pattern connection network is sufficiently suitable for text-to-speech synthesis. It should be noted that the distance calculated by Expression (4) used by the interframe distance calculation unit 4 may be used instead of obtaining the score of the optimal path by Expression (7).

As a result of the evaluation of the suitability of the voice pattern connection network, if an inappropriate transition is found in the voice pattern connection network, an unnecessary transition path is deleted by the network editing unit 19. Thus, a better synthesized voice quality can be obtained. Here, the editing process by the network editing unit 19 may simply replace the inappropriate transition weight in the voice pattern connection network with “0”. Further, the structure of the voice pattern connection network may be converted.

As described above, in the present embodiment, the voice pattern connection network creating device 1 is provided. Then, the voice pattern connection network creating apparatus 1 refers to the code unit and the voice unit vector stored in the frame cluster codebook storage unit 9 by the voice unit information adding unit 3 and stores the code in the natural voice pattern storage unit 2. A speech unit vector is attached to the stored frame of the natural speech pattern. Then, a distance d between frames belonging to different natural voice patterns is calculated by an inter-frame distance calculation unit 4.
The inter-pattern connection frame testing unit 5 extracts all the frame pairs exhibiting the inter-frame distance d within a predetermined value, and sets them as inter-pattern connection frame pairs. The transition weight calculation unit 7 connects the above-mentioned inter-pattern connection frame pairs, and adds a transition weight b according to the distance d to transitions between feature spectra corresponding to the connected frame pairs to form a voice pattern connection network. Then, it is stored in the voice pattern connection network storage unit 8.

At the time of synthesized speech, the given text is subjected to language analysis by the language analysis unit 11, and the input text is converted to a speech unit label sequence by the speech unit label sequence conversion unit 12 based on the result of the language analysis. Based on the converted speech unit label string, the parameter string search unit 13 determines the optimal path on the speech pattern connection network that maximizes the cumulative likelihood p (N) up to the last frame according to Equation (7). select. Then, the voice synthesis unit 1
No. 4 generates a synthesized speech based on a synthesis parameter sequence that matches the optimum route selected as described above.

That is, in the present embodiment, the synthesis parameters are not stored as a database of speech units, but rather, natural sound patterns formed by frame analysis of natural speech in advance can be connected between acoustically connectable patterns. By examining the connection frame pair, in addition to the "invariant path" between frames in the natural voice pattern, a "transition path" with frames of other natural voice patterns is provided, and a transition weight is added to both paths. It is stored as the created voice pattern connection network. Then, at the time of voice synthesis, an optimum route on the voice pattern connection network is searched to obtain synthesis parameters. Therefore, at the time of speech synthesis by the above-mentioned edit synthesis method, it is possible to evaluate and evaluate the possibility of use as a synthesis parameter for a feature spectrum of another natural speech pattern obtained from an articulation state having a similar acoustic feature, Mutual complementation can be performed in smaller sections in the natural voice pattern. As a result, according to the present embodiment, speech data can be smoothly and naturally connected in units of frames smaller than the unit based on phoneme information, and a synthesis parameter sequence that is best suited to the input text can be obtained. You can.

In the voice pattern connection network stored in the voice pattern connection network storage section 8, a pattern connection frame pair that can be acoustically connected based on a natural voice pattern is checked in advance. The transition path is set only between the pattern connection frame pairs in the voice pattern. Therefore, at the time of speech synthesis, it is possible to avoid the execution of connection evaluation for a pair of frames that cannot be connected acoustically, and it is possible to greatly reduce the amount of calculation for searching for a connected frame.

The association between each node in the voice pattern connection network and the voice unit label is defined as a probability that each voice unit label passes through each node on the voice pattern connection network as shown in FIG. 3 or FIG. Has been given. Therefore, if a certain node can stably pass a plurality of speech unit labels, the feature spectrum corresponding to the node can be used in association with different speech unit labels on different synthesis parameter strings. Therefore, a large amount of spectrum transition information can be represented by a small amount of voice pattern connection network information, and a feature spectrum obtained from natural voice can be effectively used for generating synthesis parameters. This also eliminates redundant feature spectrum information storage and prevents an unnecessary increase in storage amount.

In the present embodiment, the evaluation of the voice pattern connection network stored in the voice pattern connection network storage section 8 is performed by the network evaluation section 18 on a limited text (known voice pattern). The evaluation is possible. Further, the network editing unit 19 can delete unnecessary transition paths that cause an error. Therefore, it is possible to overcome the drawback of the text-to-speech synthesis method in which it is difficult to reduce the risk that particularly undesired speech units are connected to each other even in the limited sentence space, and to improve the quality of the text-synthesis speech using the limited text. It can be enhanced.

[0074]

As is apparent from the above description, the speech synthesizing apparatus according to the first aspect of the present invention connects frames that can be acoustically connected between natural speech patterns formed by analyzing natural speech frames. A voice pattern connection network in which transition weights are added to connections between the voice pattern connection networks is stored in a voice pattern connection network storage unit. A search is made for the optimal path in which the cumulative total of transition weights when tracing upward has a maximum value to obtain a synthesized parameter sequence, and the speech synthesis unit performs, based on the obtained synthesized parameter sequence and the prosody parameter generated by the prosody generation unit. Since synthetic speech is generated, mutual complementation is performed in even smaller sections in the natural speech pattern, and synthesized speech is generated. It is possible to obtain a meter columns.
Therefore, according to the present invention, feature spectra can be smoothly connected in units of frames smaller than phonemes or the like, and synthetic speech having high naturalness can be generated.

Further, since the above-mentioned speech pattern connection network is formed by connecting only acoustically connectable frames between natural speech patterns, when searching for an optimum route at the time of synthetic speech, it is acoustically necessary. It is possible to avoid search processing for useless routes that cannot be connected. Therefore, the amount of calculation at the time of synthesized speech can be reduced.

In the speech synthesizer according to the second aspect of the present invention, a speech unit label is attached to each frame of the natural speech pattern by a speech unit information adding unit, and a distance between characteristic spectra of frames belonging to different natural speech patterns is provided. Is calculated by the inter-frame distance calculation unit, and the transition weight between the pair of frames in which the inter-frame distance is equal to or less than a predetermined value and which can be an inter-pattern connection frame pair is calculated by the transition weight calculation unit. The transition weights are added to the connections between the frames of the voice pattern connection network formed by connecting the inter-connection frame pairs to each other and stored in the voice pattern connection network storage unit. Acoustic connectable patterns are connected and transition between natural voice patterns A voice pattern connected network only is added, can be automatically created based on natural speech patterns stored in natural speech pattern storage unit.

The inter-frame distance calculator in the speech synthesizer according to the third aspect of the present invention calculates the distance between the characteristic spectra of two frames to be calculated.
The difference between the power of the two feature spectra, the non-speech location of the two feature spectra, the direction of change of the two feature spectra, or the corrected distance corrected by the variance of the two feature spectra is obtained. The distance between frames can be calculated by performing a correction that emphasizes the difference. Therefore, according to the present invention, it is possible to calculate an inter-frame distance by which a pair of acoustically similar frames can be accurately determined.

The voice unit information adding unit in the voice synthesizing apparatus according to the fourth aspect of the present invention refers to the frame cluster storage unit to store the voice unit associated with the cluster to which each frame of the natural voice pattern belongs. Since a speech unit label is attached to each frame by attaching a vector to the frame, a frame which is a node of the speech pattern connection network and a speech unit label in the form of a vector having a passage probability regarding a plurality of speech unit labels as an element. Can be associated with. Therefore, it is possible to associate a feature spectrum corresponding to one node of the voice pattern connection network with a different voice unit label, and a large number of articulation transition information can be represented by a small amount of voice pattern connection network information. Further, it is possible to eliminate the storage of redundant feature spectrum information and prevent an increase in storage capacity.

In the voice synthesizing apparatus according to the fifth aspect of the present invention, the inter-pattern connection frame testing unit may determine whether the inter-frame distance is equal to or less than another predetermined value smaller than the predetermined value. Since the pair of frames having the inter-frame distance is merged into one frame, the pair of acoustically similar frames is merged into one frame to delete redundant feature spectrum information.
The storage capacity can be reduced.

Further, in the speech synthesizer according to the present invention, when the network evaluation unit traces the voice pattern connection network in the order of the voice unit label sequence of the specific voice pattern having a known voice unit label, The optimal path in which the sum of the transition weights exhibits the maximum value is searched, and the sequence of the audio unit labels attached to the frames on the searched optimal path is output as the recognition result of the specific audio pattern. The suitability of the speech pattern connection network for speech synthesis can be evaluated by the matching rate between the speech unit label string of the known speech pattern and the recognition result (speech unit label string), and the value of the cumulative transition weight presented by the optimal path. .

The speech synthesizer according to the invention of claim 7 has a network editing unit and changes the transition weight between the frames in the speech pattern connection network, thereby performing speech synthesis with a limited vocabulary. In this case, inappropriate inter-frame connections existing in the voice pattern connection network can be easily corrected, and the quality of synthesized speech based on limited linguistic information can be improved.

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech synthesizer according to the present invention.

FIG. 2 is a flowchart of a voice pattern connection network generation processing operation performed by the voice pattern connection network creation device shown in FIG. 1;

FIG. 3 is a schematic diagram showing an optimal route of one synthesis parameter string passing through a part of a voice pattern connection network structure stored in a voice pattern connection network storage unit in FIG. 1;

FIG. 4 is a schematic diagram showing an optimum route of a synthesis parameter sequence different from FIG.

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Voice pattern connection network creation apparatus 2 ... Natural voice pattern storage part 3 ... Voice unit information addition part 4 ... Inter-frame distance calculation part 5 ... Inter-pattern connection frame verification part 6 ... Frame merging means
7: transition weight calculation unit, 8: voice pattern connection network storage unit, 9: frame cluster codebook storage unit, 10: text input unit, 11 ...
Language analysis unit, 12: voice unit label string conversion unit, 13
... parameter string search unit, 14 ... speech synthesis unit,
15: Prosody generation unit, 16: Voice output unit,
17 speaker, 18 network evaluation unit,
19 Network editing department.

Claims (7)

[Claims] An input unit configured to analyze a natural voice obtained from a speaker and perform a frame analysis on each of the plurality of natural voice patterns with a voice unit label attached to each frame; A pair of frames having similar characteristic spectra are connected to each other, and a voice pattern connection network storing section storing a voice pattern connection network formed by adding a transition weight to the connection between the frames; and When referring to the audio unit label of each frame to be configured and tracing on the audio pattern connection network in the order of the audio unit label sequence based on the linguistic information input from the input unit, the total of the transition weights has a maximum value. The optimal route to be presented is searched, and the sequence of the feature spectrum of the frame on the searched optimal route is synthesized. A parameter sequence search unit that outputs a data sequence, a prosody generation unit that generates a prosody control parameter based on the input linguistic information, a synthesized parameter sequence from the parameter sequence search unit, and a prosody from the prosody generation unit A speech synthesis device comprising a speech synthesis unit that generates a synthesized speech based on a control parameter. 2. The natural voice pattern storage unit according to claim 1, wherein a natural voice pattern storage unit stores a natural voice pattern obtained by frame analysis of a natural voice obtained from a speaker. A speech unit information adding unit for attaching the speech unit label to each frame; an inter-frame distance calculating unit for calculating a distance between characteristic spectra of frames belonging to different natural speech patterns; and whether the inter-frame distance is equal to or less than a predetermined value. An inter-pattern connection frame test unit that tests whether a pair of frames having the inter-frame distance can be a pair of inter-pattern connection frames that connects the different natural voice patterns. The transition weight when transitioning between frame pairs that can be connected frame pairs and between already connected frame pairs For this reason, the transition weight is added to the connection between each frame of the voice pattern connection network formed by connecting the above-mentioned inter-pattern connection frame pairs in a plurality of natural voice patterns to each other and stored in the voice pattern connection network storage unit. A speech synthesizer comprising a transition weight calculator. 3. The speech synthesizer according to claim 2, wherein said inter-frame distance calculation unit calculates a distance to be calculated.
For the distance between the feature spectra of two frames, it is necessary to find the correction distance obtained by correcting the power difference between the two feature spectra, the silent part of the two feature spectra, the change direction of the two feature spectra, or the variance of the two feature spectra. Characteristic speech synthesizer. 4. The voice synthesizing apparatus according to claim 2, wherein when all frames of the natural voice pattern are clustered, voice having a ratio of the number of frames dropped into each cluster for each voice unit label as an element. A frame cluster storage unit that stores a unit vector in association with the cluster; wherein the audio unit information adding unit refers to the frame cluster storage unit and associates the unit vector with a cluster to which each frame of the natural audio pattern belongs. A speech unit label attached to each frame by attaching a given speech unit vector to the frame. 5. The voice synthesizing device according to claim 2, wherein the inter-pattern connection frame test unit determines that the inter-frame connection frame is smaller than the other predetermined value if the inter-frame distance is smaller than the predetermined value. A speech synthesizer comprising a frame merging means for merging a pair of frames having a distance into one frame. 6. The voice synthesizing apparatus according to claim 1, wherein the voice unit label is known by referring to the voice unit label of each frame constituting the voice pattern connection network. When tracing on the voice pattern connection network in the order of the voice unit label sequence of the specific voice pattern, a search is made for an optimum route in which the total of the transition weights has the maximum value, and the route is added to the frame on the searched optimum route. A network evaluation unit that outputs a sequence of the voice unit labels as the recognition result of the specific voice pattern, a matching rate between the voice unit label sequence of the specific voice pattern and the recognition result, and a cumulative transition represented by the optimal path. A speech synthesizer characterized in that the suitability of the speech pattern connection network can be evaluated by the value of the weight. 7. The speech synthesizer according to claim 1, further comprising a network editing unit for changing a transition weight between the frames in the speech pattern connection network. Speech synthesizer.