KR100241345B1 - Simplified intonation stylization for ktobi db construction - Google Patents

Abstract

The present invention relates to a method for extracting only the main change parts of intonation from the intonation signal which affects the naturalness of the synthesizer as a signal processing for simplifying the complex and various intonation patterns in speech, and a method of simplifying them into a stepped intonation model. From the result of analysis of accent marginalization, defining accents for each symbol into stepped accent model and accent style according to accent pattern notation system, and signal processing step of simplifying accent pattern to stepped type according to this model. By doing this, it is possible to investigate the consistency of the accent information that is manually labeled by a person or to classify intonation patterns by hand, greatly reducing the time required to construct the database for intonation modeling, and to classify intonation patterns essential for automating database construction. Has the effect of easily extracting features.

Description

How Kthiobia Simplifies Accent Curves for Database Construction

The present invention relates to a method of simplifying an intonation curve for constructing a KToBI database.

Accent patterns and utterance information of spoken sentences are important factors for the naturalness of the Korean unlimited speech synthesizer.

In particular, many studies have been attempted to regularize such information in a large rhyme database, and the necessity of expressing the intonation pattern information as a symbol has been raised.

The Korean Tone and Break Indices (hereinafter referred to as KToBI) accent pattern notation has been proposed for this purpose, and the accent modeling, duration modeling, break reading, etc., are performed using the rhyme database described under this accent system. Statistical extraction of rules

Conventionally, an expert who can add rhyme information to a voice database is needed to build a large rhyme database, and even if it is possible to add rhyme information by an expert, it takes much time to construct.

In particular, manual work frequently causes errors, and since the consistency of the database may vary depending on the expert, there is a problem that the automation of information addition is urgently required.

In order to solve the above problem, the present invention is a method for facilitating classification of accent patterns by simplifying complex and various accent patterns in speech, and inflicting sentences to investigate the consistency between manually labeled symbols and intonation patterns. Simplification, and the simplified accent aims to fully reflect accents and accents, which are marginalizations of accents that are important to nature.

1 is a flow chart of a simplified process with a stepped model according to the present invention,

2A to 2E are waveform diagrams of each step intonation curve applied to the present invention and the jaw curve modeled in the final stepped form;

3 is an original accent, stepped accent, labeled accent symbol waveform in accordance with the present invention.

In order to achieve the above object, the present invention, step of defining the accent according to each symbol from the result of the marginal analysis of intonation into a stepped accent model and intonation style according to the accent pattern notation system, and the accent pattern according to the defined model It characterized in that it comprises a signal processing step to simplify the step.

Hereinafter, the present invention will be described in detail with reference to the accompanying drawings.

Simplified intonation should fully reflect accents and accents, which are the periphery of accents to be examined, so as to minimize the complexity of patterns due to excessive intonation changes due to phonological environment and detection errors of intonations. Go through the process.

1 is a flowchart of a simplification process with a stepped model according to the present invention, showing a signal processing process for simplifying an intonation pattern.

Input the original accent curve (S1) to replace zero for the unvoiced section of the original accent curve and the section with a pitch value of less than 70 Hz (S2), and 3-point median filtering Perform (S3).

The filtering process (S3) eliminates the generation of pitch pulses due to phonological environment, for example, rupture, friction, rupture sound environment, which can cause a large error when simplifying the conventional intonation.

After performing the above, a linear linear equation is created using the pitch value of the voiced sound at both ends of the unvoiced sound as a boundary value, and the linear equation is linearly allocated to the unvoiced sound interval.

This is a process for continuing the disconnection of the pitch due to unvoiced sound (S4).

After allocating the pitch, the low-pass filtering is performed so that only the accent margin to be examined, that is, the accent accent or the accent accent (S5).

At this time, the cutoff frequency of the filter is changed to 200㎐, 500㎐, 1,000㎐, 1,200㎐, 1,500㎐, 2,000㎐ by step, and experimentally determine the cutoff frequency of 1,200㎐, which shows the accent marginalization when filtering. .

In the low-filtered intonation curve, the peak to valley of the intonation is found, and the midpoint between the peak and the valley is simplified stepwise (S6).

Then, the final stepped accent curve is made by substituting again for zero in the bundle section (S7).

2A to 2E are waveform diagrams of each step accent curve applied to the present invention and a jaw curve modeled in the final stepped step, and are output values for each step and finally simplified to stepped accents.

2A is an accent curve, 2b is a median filtering output, 2c is a linearization process, 2d is a low pass filtering result, and 2e is a final stepped intonation.

3 is a diagram of the original intonation, stepped intonation, and labeled intonation symbol waveforms in accordance with the present invention, showing stepped intonation and intonation symbols.

Comparing the intonation symbol described above and the intonation pattern represented by the stepped shapes in the following [Table 1] and the stepped intonation can be seen whether the representation of each symbol is consistently labeled.

For example, in the case of LH%, it can be found from the stepped accent whether the intonation pattern at the intonation symbol position is High and the previous intonation pattern is Low.

Stepped accents were randomly selected test data that corresponded to about 10% of the total number of KToBI symbols to evaluate how well modeling the accent patterns in the original accent curve.

As a criterion for evaluating whether the accent pattern in the circular accent curve is consistent with the stepped accent pattern, the evaluation target is the accent of frequently occurring symbols, that is, H-, LHa, HL%, LH%, H%, L%, etc. The pattern was 98.9% for females and 97.2% for males.

As shown in Table 2 below, the overall performance of the stepped model is 93.2% and 94.1% for men and women, respectively, and this result is a reliable figure for analyzing the consistency of manually labeled KToBI symbols from a large amount of data. have.

In addition, stepped model accents can automatically label KToBI or clearly display accent patterns in analysis, making it easy for non-experts to label KToBI.

Stepped accent modeling performance evaluation (correct number / total number (%)) Symbol H- LHa HL% LH% H% (+ lev) L% (+ lev) sccore Female 162/173 (93.6) 114/123 (92.7) 51/55 (92.7) 20/22 (90.9) 33/33 (100.0) 37/33 (100.0) 94.1 Male 166/176 (94.3) 100/116 (86.2) ) 77/81 (95.1) 16/16 (100.0) 35/35 (100.0) 93.2

As described above, according to the present invention, the method of extracting only the main change portion of the intonation which affects the naturalness of the synthesizer from the intonation signal, and the method of simplifying the step into the step function, the method of manually labeling intonation of the human It is easy to classify intonation patterns when checking the consistency or by hand, which can greatly reduce the time required to construct the database for intonation modeling, and it is easy to extract the intonation pattern classification features essential for automating database construction.

Claims (3)

Signal processing to simplify the complex and various intonation patterns of speech, A first process of defining intonation according to each symbol from the result of the marginalization analysis of intonation into a stepped intonation model and intonation style according to the intonation pattern notation system; A method of simplifying an intonation curve for constructing a KToBI database, comprising a second process of signal processing to simplify the intonation pattern according to a defined model. The method of claim 1, wherein the first process is A first step of inputting a circular intonation curve; A second step of substituting zero for an unvoiced sound section of the input circular intonation curve and a pitch value section below a predetermined frequency and performing three-point median filtering; A third step of creating a linear equation having the boundary value of the pitch value of the voiced sound at both ends of the unvoiced sound after performing the filtering; And a fourth step of linearly allocating a pitch to the unvoiced sound interval using the linear equation. The method of claim 1, wherein the second process A first step of performing low-pass filtering so that only an accent accent or an accent intonation to be examined after the pitch is assigned (S5); Finding a peak and a valley of the intonation in the low-filtered intonation curve, and then simplifying the intermediate point of the peak and the valley into a step (S6); Simplifying the intonation curve for the construction of the KToBI database, characterized in that it comprises a third step of simplifying and then replacing the bundle interval to zero again to create a final stepped accent curve (S7).

Images