JP2001092481A - Method for rule speech synthesis - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method for rule speech synthesis whereby speech synthesis is performed by utilizing CVs and speech elements in a VC unit prepared beforehand and whereby a speech dictionary volume and distortion in a connection plane are reduced. SOLUTION: The method for rule speech synthesis in this invention is characterized by dividing CVs and speech elements in a VC unit into synthesis units smaller than phoneme units by using HMnet, creating an intra-phoneme network over all the synthesis units, and selecting speech elements corresponding to an input phoneme train on the basis of the obtained intra-phoneme network.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a rule speech synthesis method.

[0002]

2. Description of the Related Art In rule-based speech synthesis, speech synthesis is generally performed by selecting a connection unit (speech unit) extracted from speech data in accordance with a sentence to be synthesized, correcting the prosody, and connecting.

The connection units are (C, V), (CV, V
C), VCV, etc. are used. Here, C indicates a consonant, and V indicates a vowel.

The number of connection units is the smallest for phonemes (C,
V) is a connection unit, but at the boundary between phonemes, since the acoustic properties differ depending on the types of the preceding and succeeding phonemes, if each phoneme has only one segment, the connection surface Distortion increases. For this reason, it is generally necessary to have a plurality of segments for each phoneme. Having multiple segments for each connection unit improves sound quality,
This will increase the size of the voice dictionary.

By the way, when (CV, VC) is used as a connection unit, about 500 units are required to synthesize Japanese, even if only one unit is provided for each connection unit. . Further, from the viewpoint of sound quality, if each connection unit has only one segment, a preceding or succeeding phoneme may connect different connection units, which may cause significant connection distortion.

When (CV, VC) is used as a connection unit, similar waveforms are included in different connection units. In order to eliminate such a portion, the connection unit may be made finer, and a portion having similar acoustic properties may be shared as one connection unit (synthesis unit). In addition, the influence of different preceding and succeeding phonemes does not appear equally for all phonemes. Therefore, if the connection unit is fine, the connection unit can be efficiently increased only in a portion having a large influence.

[0007] Such a concept holds even when a phoneme (C, V) is used as a connection unit, and has already been described by Ito et al.
-222898). In other words, multiple arcs are connected via nodes for each phoneme,
A sound parameter network to which sound characteristic parameters corresponding to each arc are attached is stored, a series of arcs is searched according to an input phoneme sequence, a sound characteristic parameter sequence is formed, and speech is synthesized. is there.

In this method, a network is formed for each phoneme. On the other hand, it is not easy to apply this method to a speech database made up of CV and VC units prepared in advance, since it is necessary to construct a network that spans all synthesis units.

[0009]

SUMMARY OF THE INVENTION The present invention relates to a voice synthesizing method for synthesizing voice using a unit of CV or VC prepared in advance. It is an object of the present invention to provide a rule-based speech synthesis method capable of reducing surface distortion.

[0010]

The rule speech synthesis method according to the present invention divides a CV / VC unit segment into smaller synthesis units than phoneme units using HMnet, and forms a network within a phoneme spanning all synthesis units. Is generated, and a segment corresponding to the input phoneme sequence is selected based on the obtained intra-phoneme network.

The intra-phoneme network is obtained, for example, from a plurality of first states obtained from the starting end portions of the individual CV and VC units prepared in advance and the central portion of all the CV and VC units. One second state and individual CVs,
A step of dividing a state by an iterative state division method from an initial state of HMNet composed of a plurality of third states obtained from a terminal part of a VC unit fragment, and HMNet divided to an arbitrary number of states; Individual CV, V
It is generated by the step of determining a representative segment of each state of HMNet based on the segment in C units.

The step of determining a representative segment of each state of HMNet is performed, for example, by dividing individual segments in units of CV and VC and an arbitrary number of states, and dividing each learning segment into each state from HMNet. Then, a step of creating a candidate of a segment of each state, and a connection point between all the states of HMNet plus all combinations of the terminal and the starting end of the same phoneme of HMNet are defined as a total connection point. The method comprises the steps of selecting a combination of segments that minimizes the sum of connection distortions at the connection points from the segment candidates in each state and setting the combination as a representative segment.

[0013]

Embodiments of the present invention will be described below with reference to the drawings.

[1] Description of Features of the Invention

According to the present invention, there is provided a speech synthesizing method using CV and VC as connection units, wherein the method comprises a hidden Markov network (HMne).
A speech dictionary is compressed by forming a network for all connection units using t).

HMNet is a technique used in speech recognition. During learning of the HMM, by sharing the state between different phonemes and allophones, the parameters of the entire model are reduced, and statistically stable. ("Takami, Sagayama:" Automatic Generation of Hidden Markov Network by Sequential State Division Method (SSS) "," Proc.
−5-1- (1991-1-9) ”). In the present invention,
One connection unit of synthesis (synthesis unit) was assigned to one state of HMNet.

FIG. 1 shows an example of an intra-phoneme network obtained by dividing a segment in CV and VC units into smaller synthesis units than phoneme units using HMNet.

Each ellipse indicates each state, and each state has a composite 1
Connection units (units smaller than phoneme units) are assigned. The symbols a-s, z, and n in the ellipse indicate that a segment starting with [a] and ending with [s], [z], [n] passes.

[2] Description of method for generating intra-phoneme network

Hereinafter, a method of generating a network within a phoneme by dividing a CV / VC segment into smaller pieces by using HMNet will be described.

FIG. 2 shows the initial state of HMNet.

The left column shows all CVs prepared for learning,
Of the VC segments, a segment beginning with [a] is extracted, and the state consisting of the average value and the variance value of the acoustic parameters calculated from the beginning portion is similarly set from the segments beginning with [a *] and [u]. Represents the plurality of first states generated in such a manner as [u *]. The central state represents one second state including the average value and the variance of the acoustic parameters calculated from the central portions of all the CV and VC segments. The right column shows a plurality of third states including the average value and the variance value of the acoustic parameters calculated from the terminal portions of the individual CV and VC segments.

For the state division, a sequential state division method was used. In the sequential state division method used for speech recognition, division in the context direction is performed based on the phoneme type or phoneme environment. In this embodiment, however, the number of states is increased regardless of the phoneme type. .

The reason for this is to take into account that the distortion of the acoustic parameters at the time of speech synthesis is affected not only by the phoneme environment but also by the phoneme environment, and to assign a state to those factors.

By the successive state division method, for example, FIG.
As shown in, when the state is divided into an arbitrary number of states,
The representative segment of each state is determined as follows.

(1) Each learning unit (CV, VC unit)
And the number of states is divided into
Using the erbi algorithm, each learning segment is divided into each state, and candidate segments for each state are created.

That is, the learning unit is divided into the states based on the ratio of the time length in each state on the path corresponding to an arbitrary learning unit in HMNet. Assuming that the time length of a state that does not involve a transition from the same state to the same state is 1, for a state that involves a transition from the same state to the same state, the number of transitions from the same state to the same state is n. Then, the time length in this state is (1 + n).

(2) The connection relation between all the states is obtained based on the state transition of HMNet.

(3) The total connection point is obtained by adding all the combinations of the terminal and the start of the same phoneme of HMNet to the above connection relation.

(4) A combination of segments that minimizes the sum of the connection distortions at the total connection points is selected from the segment candidates in each state and is used as a representative segment.

For the selection of the representative element in the above (4), simulated annealing (SA) was used. The connection distortion was calculated by the following equation (1).

[0032]

(Equation 1)

D _F0 , D _pow and D _cep represent a difference in fundamental frequency, a difference in power, and a cepstrum distance at a connection point of the two combining units connected respectively. Also,
w _FO , w _pow and w _cep represent weighting factors by which D _F0 , D _pow and D _cep are multiplied, respectively.

[3] Description of speech synthesis method

The intra-phoneme network (HMn) shown in FIG.
(et) will be described.

Here, a description will be given of a case where "asa" is synthesized. "Asa" is a fragment of "as" and "s-
a ".

(1) First, as shown in FIG.
The path through which s "passes is extracted.

(2) Next, as shown in FIG.
The path through which a "passes is extracted.

(3) Then, as shown in FIG. 5, these routes are connected.

In this case, there is a path passing through the state A and a path passing through the state B. Therefore, it is checked which path has a small connection distortion, and a modal path with a small connection distortion is selected. After the representative segments in each state on the selected route are corrected to appropriate height, length, and size, they are connected to generate synthesized speech.

[0041]

According to the present invention, a CV prepared in advance,
In a speech synthesis method for performing speech synthesis using segments in VC units, it is possible to reduce the size of a speech dictionary and reduce distortion on a connection surface.

[Brief description of the drawings]

FIG. 1 shows a CV, VC unit fragment using HMNet.
It is a schematic diagram which shows an example of the intra-phoneme network obtained by dividing | segmenting into a synthesis unit finer than a phoneme unit.

FIG. 2 is a schematic diagram showing an initial state of HMNet.

FIG. 3 is a schematic diagram showing a path through which “as” of HMNet of FIG. 1 passes.

FIG. 4 is a schematic diagram illustrating a path through which “sa” passes among HMNets of FIG. 1;

FIG. 5 is a schematic diagram showing a route obtained by connecting the route in FIG. 3 and the route in FIG. 4;

Continuation of the front page (72) Inventor Hideharu Nishida 2-5-5 Keihanhondori, Moriguchi-shi, Osaka Sanyo Electric Co., Ltd. (72) Inventor Mitsumi Okura 2-5-5-1 Keihanhondori, Moriguchi-shi, Osaka No. Sanyo Electric Co., Ltd. (72) Inventor Hiroki Onishi 2-5-5 Keihanhondori, Moriguchi-shi, Osaka F-term in Sanyo Electric Co., Ltd. (reference) 5D045 AA07

Claims (3)

[Claims] 1. A unit of CV and VC units is divided into synthesis units smaller than phoneme units using HMNet, a network within the phoneme over all the synthesis units is generated, and based on the obtained network within the phoneme. A rule speech synthesis method in which a segment corresponding to an input phoneme sequence is selected. 2. The intra-phoneme network is obtained from a plurality of first states obtained from a starting end portion of each CV / VC unit piece prepared in advance and a central portion of all CV / VC unit pieces. From the initial state of HMNet, which is composed of one second state and a plurality of third states obtained from the terminal portions of the individual CV and VC units,
Splitting the states by the sequential state splitting method, and HMNet split into an arbitrary number of states, and individual C
2. The rule speech synthesis method according to claim 1, wherein: a step of determining a representative segment of each state of HMNet based on the segments in V and VC units. 3. The step of deciding a representative segment of each state of HMNet is performed by dividing individual segments in units of CV and VC and an arbitrary number of states and dividing each learning segment into each state from HMNet. And
Creating a candidate for each state segment; and HMn
The sum of the connection points between all the states of et and all the combinations of the end and start ends of the same phoneme of HMNet is defined as the total connection point, and the element that minimizes the sum of the connection distortion at the total connection point 3. The rule speech synthesis method according to claim 2, further comprising the step of: selecting a combination of segments from the segment candidates in each state to be a representative segment.