JPH03125200A - Voice synthesizing method - Google Patents

Abstract

PURPOSE:To synthesize a voice which has fluctuations close to a natural voice by converting a phoneme symbol sequence into a parameter matrix and then converting the parameter matrix into a parameter matrix of the voice which has fluctuations close to the natural voice. CONSTITUTION:The conversion into the parameter matrix is performed by a rule synthesizing method which uses the parameter matrix averaged in each phoneme environment as a synthesis unit and this converted parameter matrix 1 has its time base matched with that of the parameter matrix 2 of the natural voice for learning corresponding to a predetermined phoneme sequence, phoneme by phoneme, by linear expansion or contraction in the time-base direction to obtain an input signal 3 to a neural network. Then, outputs of respective units in an input layer 4 are multiplied by respective coupling coefficients 5 and the total of all the units in the input layer 4 is calculated to obtain an input to each unit of an intermediate layer 6. Outputs of respective units in the intermediate layer 6 are multiplied by respective coupling coefficients 7 and the total is calculated to obtain an input to an output layer 8. Consequently, the voice having fluctuations close to the natural voice can be synthesized.

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application Field] The present invention relates to a speech synthesis method for making synthesized speech obtained by a rule synthesis method into highly natural synthesized speech.

[Prior art] In the rule synthesis of speech, it is important to reflect the articulatory effect due to the phonetic environment, etc., and there are methods for extracting synthesis units from a relatively wide sound environment such as CVCl3 continuous phonemes, transformation rules, and spectrum activation. Methods such as a control method that transforms synthesis units when they are connected, and a method that uses clustering based on the phonological environment to generate synthesis units that reflect the effects of articulatory combination have been proposed.

In the method of extracting synthetic units from a relatively wide phonological environment such as the above-mentioned CvC53 continuous phoneme, or the method of transforming the synthetic units when they are connected, the range of articulatory combinations and the phonological environment are determined based on a priori knowledge, so the synthesis units are It took a lot of effort to create, and there were problems with the flexibility of the system, such as changing the speaker and the number of composite units.

On the other hand, in the clustering method based on the phonological environment, synthesis units are automatically generated using statistical methods only from phonologically labeled learning data.
The effort of creating synthetic units is reduced without using a priori knowledge. However, since a parameter matrix averaged for each phoneme environment is used as a synthesis unit, fluctuations in speech (subtle changes in strength and pitch within each phoneme) are not observed, and the naturalness is not sufficient. In other words, in this method based on clustering of the phonological environment, first, a set of parameter matrices (parameter time series) given the same phonological label in the learning data is used as an initial cluster, and the preceding and following of each variable matrix in the cluster is set as an initial cluster. Cluster division is performed based on phoneme, and this is continued until a certain evaluation value (such as the average value of intra-cluster variance) falls below a certain threshold. The centroid matrix of each cluster obtained in this way was registered as a synthesis unit. When calculating the center of gravity and dispersion of a cluster consisting of parameter matrices of different lengths, each parameter matrix was linearly expanded and contracted in the time direction using a fixed dimension method, and the center of gravity was calculated by converting it into a fixed dimension. Therefore, each centroid matrix was a parameter matrix averaged for each phonetic environment, and had the average duration length of the cluster. For this reason, in vowel stationary parts, etc., the sound quality becomes vocoder-like without fluctuations, which is characteristic of synthesized speech, and lacks naturalness. It was also necessary to improve the naturalness of phoneme duration and prosody.

[Means for Solving the Problems] This invention uses a neural network to learn mapping relationships from parameter matrices averaged for each phonetic environment to natural speech, and converts phonetic symbol strings manually created to synthesize speech into phonemes. The method is characterized by converting into a parameter matrix averaged for each environment, inputting the parameter matrix to the learned neural network, and obtaining a speech parameter matrix having fluctuations close to natural speech as an output of the neural network. . This method differs from conventional methods in that the synthesis units are not uniformly averaged, but have fluctuations, resulting in improved naturalness.

[Example] FIG. 1 is an explanatory diagram of the present invention. In this figure, the predetermined phonetic symbol strings include /b, a, u, o, n.
A case is shown in which a phoneme symbol string / is input.

First, this predetermined phonetic symbol string is converted into an LSP parameter matrix by a rule synthesis method that uses parameter matrices averaged for each phonetic environment as synthesis units, and this converted SP parameter matrix l
is aligned with the LSP parameter matrix 2 of natural speech for learning corresponding to the predetermined phoneme symbol string for each phoneme by linear expansion/contraction in the time axis direction, and is used as an input signal 3 to the neural network. . The analysis order of LSP parameters is 16th order, and the analysis window length is 20n+s.

The frame shift is 5Ills. Input N4 of the neural network has 480 units of 30 frames of 16th order parameters. Input the parameters by sliding the appropriate number of frames here. The teacher signal for the output layer 8 of the neural network is the parameter matrix 2 of the natural speech for learning, which is temporally synchronized with the input parameter matrix (that is, corresponds to the same phoneme), and is slid in the same way as the input signal. . Therefore, the number of units in the output layer 8 is also the same as that in the input layer 4, which is 480.
It is.

The units in the input layer 4 are simply input terminals, and input parameter values are used as output values without any special conversion. The output of each unit in the input layer 4 is multiplied by a coupling coefficient 5, and the sum of all the units in the input layer 4 is taken as an input to each unit in the intermediate layer 6. That is, the input value neJ of unit j of the intermediate layer 6 is expressed by the first equation.

However, ga 4. is the coupling coefficient from input layer unit i to hidden layer unit j, U! is the output value of input layer unit i.

In each unit of the intermediate layer 6, human power is converted by a nonlinear function such as a sigmoid function and becomes an output. That is, the output value uj of unit j in the intermediate layer 6 is expressed by the second equation.

u j = 1/(1+exp-"" θ, )) (2nd formula) where θ is the threshold value of unit j.

The output of each unit of the intermediate layer 6 is multiplied by the coupling coefficient 7 as in the case of the human-powered layer 4, and the sum is taken to become the input of the output layer 8. That is, the input value ne to the unit of the output layer 8
5 becomes the third equation.

net, -Σ匈7Xu, (3rd formula) where N is the number of units in the middle layer 6.

The output u to the unit of the output layer 8 is given by the fourth equation.

u w =1/(1+exp-(″″′″θ)) (4th formula) However, θ3 is the threshold value of the unit.

Here, the LSP parameter value of the natural voice, which is the teacher signal for unit k of the output layer 8, is set to 1. Then, the network error evaluation value E is given by the following equation.

The coupling coefficient Wlljs Wji is changed by the steepest descent method so as to minimize the value of E. Repeat this step every time you slide the input signal and teacher signal a few frames.
Continue learning until the value of E becomes sufficiently small. As a result,
The output signal becomes extremely close to the teacher signal (natural speech parameter matrix for learning), which means that fluctuations close to natural speech are given to input signal 1. The number of frames for sliding the input signal and the teacher signal is determined as appropriate based on preliminary studies. Further, the number of units in the intermediate layer 6 is also determined as appropriate based on preliminary studies and the like.

When synthesizing speech using the neural network trained in the above steps, first, a given phonetic symbol string is synthesized into an LSP parameter matrix using a rule synthesis method that uses parameter matrices averaged for each phonetic environment as synthesis units. Convert to

Next, the LSP parameter matrix is input to the learned neural network every 30 frames, and
By calculating according to the formula, output is obtained for each 30 frames. By combining these outputs, an LSP variation matrix of speech having a length corresponding to a given phonetic symbol string and having fluctuations close to that of natural speech is obtained. Although the phonetic symbol string is converted into an LSP parameter matrix in the above description, it may be converted into other parameter matrices such as a Percall parameter matrix.

"Effects of the Invention" As explained above, according to the present invention, a phonetic symbol string is converted into a parameter matrix by a rule synthesis method that uses a parameter matrix averaged for each phonetic environment as a synthesis unit, and the parameter matrix is converted into a parameter matrix. Using a trained neural network, it is possible to automatically convert speech into a parameter matrix with a ti-ratio close to that of natural speech, and to create conversion rules (functions) using a priori knowledge. Since there is no need to describe it, it has the advantage that it is possible to easily synthesize speech with fluctuations similar to natural speech.

[Brief explanation of the drawing]

Figure 1 is an explanatory diagram of the present invention, where ■ is a parameter matrix obtained by the conventional rule synthesis method, 2 is a parameter matrix of natural speech for learning, and 3 is the duration length of each phoneme of 1 by linear expansion and contraction. 2 is the parameter matrix combined with 2, 4 is the input layer of the neural network, 5 and 7 are the connections whose coupling coefficients are changed by learning, 6 is the middle layer, 8 is the output layer, and 9 is the natural speech obtained as the output of the neural network. This is a speech parameter matrix with a tI error close to .

Claims (1)

[Claims] (1) Convert a predetermined phoneme symbol string into a speech feature parameter matrix using a rule synthesis method that generates synthesis units that reflect the effects of articulatory combination through clustering based on the phonetic environment, and correspond to the phoneme symbol string. The duration length of the feature parameter matrix obtained by the rule synthesis method is adjusted for each phoneme to the feature parameter matrix of the natural speech for learning, and the feature parameter matrix obtained by combining the duration lengths is input, and the input and temporal A neural network is trained using the feature parameter matrix of the natural speech for learning synchronized with the training signal as a teacher signal, the input phonetic symbol string is converted into a speech feature parameter matrix by the rule synthesis method, and the converted feature parameter matrix is A speech synthesis method characterized by inputting the above-mentioned learned neural network into the learned neural network, and using the output of the neural network as synthesized speech.