JP2697995B2 - Utterance training device - Google Patents

Abstract

PURPOSE:To improve the ability to discriminate a consonant with relatively simple constitution by calculating the coefficient of a decision making function from the LPC cepstrum of a trainee's vocalized sound, and comparing it with the coefficient of specific standard data and making a binary decision on the consonant. CONSTITUTION:The trainee's vocalized sound is converted by a microphone 2 into a speech signal, an LPF 3 extracts a reduced component among the frequency components of the speech signal, and an AD converter 4 digitizes the component. An arithmetic processing part DSP 5 analyzes the coefficient of the linear decision making function from the LPC cepstrum of the digitized speech signal. Then the contents of a ROM 6 stored with the coefficient of the linear decision making function calculated from the LPC cepstrum of the standard speech data is compared with the contents of the DSP 5 and a CPU 7 makes the binary decision on the consonant. After an image processing part 8 processes the comparison result, an image processing signal is displayed on a display 9. Then two consonants such as 'r' and 'l' or 'm' and 'n', three consonants such as 'b', 'd', and 'g' or 'p', 't', and 'k' or five vowels such as 'a', 'i', 'u', 'e', and 'o' can be recognized respectively by using the linear or nonlinear decision making function.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an utterance speech trainer for performing pronunciation training.

[0002]

2. Description of the Related Art For example, a pronunciation training device disclosed in JP-A-64-49081 is known. According to the present invention, the frequency of the first and second formants of the voice data of the trainee is compared with the frequencies of the first and second formants of the standard data to determine the voice, and the determination result is determined by the video processing device. Was displayed.

[0003]

However, according to this method, speech can be determined mainly by vowels.
Therefore, as an algorithm for recognizing consonants, for example, s, h, and z with high accuracy, neural networks and DP (D
It is conceivable to use dynamic programming (dynamic programming) or the like. However, the processing capability and storage capacity of the arithmetic processing unit built in the training machine body are relatively small, and it has been difficult to use these algorithms. Accordingly, it is an object of the present invention to provide an utterance speech trainer capable of determining consonants with a relatively simple configuration.

[0004]

In order to solve the above-mentioned problems, the present invention provides a microphone (hereinafter, referred to as a microphone) for converting a trainee's utterance into a voice signal, and an LPC cepstrum of the voice signal. Analysis means for calculating the coefficient of the discriminant function from the characteristic parameter of the audio signal; storage means for preliminarily storing the coefficient of the discrimination function calculated from the characteristic parameter of the standard audio data; and the coefficient calculated by the analysis means and the storage means And a display means for displaying the result of the determination by comparing the coefficients stored in.

[0005] The analysis means may be configured to calculate a coefficient of the discriminant function after performing a binary selection based on an average zero-cross value of a consonant section of the audio signal in advance.

The determining means includes a determining function G for the standard data.
In ab (x), a binary decision may be made between the discriminant function of the standard data and the discriminant function of the trainee except for the range of | Gab (x) | <k.

[0007]

The coefficient of the discriminant function is calculated from the LPC cepstrum of the trainee's utterance, and this coefficient is compared with a predetermined standard data coefficient to make a binary judgment of the consonant.

If the binary selection based on the average zero-cross value of the consonant section of the audio signal is performed in advance, the number of sets of discriminant functions required for the binary decision can be reduced. Note that the preliminary selection differs depending on the combination of consonants to be determined, and an autocorrelation, a function pitch frequency audio power, or the like may be used other than the average zero cross value.

For the binary determination, | Gab (x) | <k
If a certain range is excluded from the comparison target (that is, if the voice of the trainee falls in this range, it may be determined to any value), the determination can be made in consideration of the ambiguity.

[0010]

Embodiments of the present invention will be described below with reference to the accompanying drawings. FIG. 1 is a configuration diagram of an example of the utterance speech trainer according to the present invention, FIG. 2 is a graph of the frequency of each consonant versus the average zero-cross value, FIG. 3 is a mapping diagram of a discriminant function for each consonant, and FIG. 6 is a graph of rate versus ambiguity.

First, the invention according to claim 1 will be described. In FIG. 1, an utterance speech trainer 1 according to the present invention includes a microphone 2 that converts a trainee's utterance into an audio signal, and a low-pass filter (hereinafter, LPF) that extracts low-frequency components from the frequency components of the audio signal. 3) and this LPF3
A / D converter (hereinafter, referred to as A / D) 4 for converting an output signal of the digital signal into a digital signal, and an operation as analysis means for calculating a coefficient of a linear discriminant function from an LPC cepstrum of the digitally converted audio signal A processing unit (hereinafter referred to as DSP) 5 and a ROM (Random Access Mem) as a storage unit in which coefficients of a linear discriminant function calculated from the LPC cepstrum of the standard audio data are stored in advance.
ory) 6 to control the entire apparatus and
A central processing unit (hereinafter, referred to as a CPU) 7 as a determination unit for performing a binary determination of a consonant by comparing the contents of the SP 5 and the ROM 6, and an image processing unit 8 for performing image processing on the comparison result.
And a display 9 as display means for displaying the image-processed signal. In this embodiment, a case where three consonants s, h, and z are recognized will be described as an example. However, the present invention is not limited to this. For example, b, d, g, p, t, k, n, and It may be configured to recognize each of three consonants such as m, r, y, w, n, s, c, and sh. Further, the present invention can also be used when recognizing two consonants r and l or m and n, and recognizing vowels a, i, u, e, and o. As the discriminant function, not only a linear discriminant function but also a nonlinear discriminant function may be used.

Next, the DSP 5 and the CPU 7 will be described. The DSP 5 performs the following operation. That is, the LPC cepstrum of the speech signal recognition target frame is represented by x (x1, x
2,..., Xn), the linear discriminant function is represented by the following equation, and this operation is performed. Gab (x) = (x− (Ma + Mb) / 2) tΣ−1 (Ma−Mb) (1) Then, it is determined whether x belongs to class a or b.
The determination is made by PU7, and the determination is performed as follows.

That is, if Gab (x) ≧ 0, x belongs to class a, and if Gab (x) <0, x belongs to class b. In this embodiment, three discriminant functions for discriminating two classes, namely, Gs for discriminating between s and h,
h, Ghz for determining h and z, G for determining z and s
The consonant classes s, h, and z are determined using zs. The final consonant determination of the input voice is determined to be the consonant most frequently counted in the frame of the consonant section. The determination may be made based on the sum of the discriminant functions of the consonant sections.

Next, the invention according to claim 2 will be described. The following processing can be further performed by the DSP 5. That is, in this process, in order to improve the average identification rate in the consonant section, the preliminary selection is performed using the automatically cut average zero cross. FIG. 2 shows an example of the distribution of the average zero-cross value in the consonant section used for the preliminary selection. That is, when the average zero-cross value is from 0 to about 15 (V1), the frequency of z is relatively large in FIG.
(A) and (B) in which the frequency of z and h are relatively large in (V2), and the frequency of s is comparatively large in (V3) from about 60 to 100 (V3). Is determined to be s. By performing such preliminary selection in advance, the number of sets of linear discriminant functions required for class determination can be reduced, and the average identification rate can be increased by about 3% as compared with the conventional example. The trainee can train while comparing both the zero-cross value obtained in real time and the main distribution on a screen on the display 9 described later.

Next, the third aspect of the present invention will be described. Since the average discriminating power of s, h, and z was relatively low at 83% even after the preliminary selection, the present invention is intended to improve the average discriminating power. One of the reasons why the average discriminating power is low may be that the determination condition (determining whether Gab is greater than 0) is too strict.

FIG. 3 is a diagram in which the three sets of discriminant functions obtained for each consonant are mapped in a triangular shape, and are converted so that the distribution near zero can be easily understood. The shape is not limited to a triangular shape, but may be a polygonal shape. For example, if there are two sets of discriminant functions, the discriminant function may have a square shape, and if there are five discriminant functions, a pentagonal shape may be used. The figure mapped into the triangle is displayed on the display 9. As shown in FIGS. 7A and 7C, s and z are relatively easy to identify because the distribution is relatively concentrated on the vertices of each triangle, but h in FIG. Relatively many features close to s were recognized, and identification was difficult. Therefore, "ambiguity" is adopted as the determination condition near zero.

The determination incorporating the ambiguity is mainly performed by the CPU 7. In the present invention, | Gab (x) |
The range in which <k is set is a threshold (ambiguity) k, and the threshold has a certain width. When the value of Gab (x) is in the range of −k to + k, it is determined to be either class a or b. Also decided to be good. As described above, in the present invention, the binary determination based on 0 is conventionally excluded from the determination target in the range from -k to + k (that is, the range in which the binary determination is difficult). By doing so, the discrimination power can be improved. FIG. 4 shows the identification result when k is changed from 0 to 7. As can be seen from this figure, when k was set to about 3 or more, the average discriminating power for s, h, and z was about 94.5 on average. That is, the average discriminating power can be improved by introducing ambiguity into the determination condition near zero. On the other hand, by changing the value of k, the difficulty of the training content can be changed, and the range of application can be expanded.

According to the above configuration, since the processing can be performed sufficiently even by using a circuit element for a personal computer having a relatively small processing capacity and a small storage capacity of the arithmetic processing section, the apparatus can be reduced in size and cost. It is also suitable for.

[0019]

According to the present invention, consonants can be identified with a relatively simple configuration, and the discriminating power can be improved. Therefore, it is also suitable for reducing the size and cost of the device.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of an example of an utterance speech training device according to the present invention.

FIG. 2 is a graph of frequency versus average zero-cross value for each consonant.

FIG. 3 is a mapping diagram of a discriminant function for each consonant.

FIG. 4 is a graph of the discrimination rate versus the degree of ambiguity.

[Explanation of symbols]

 REFERENCE SIGNS LIST 1 utterance speech training device 2 microphone 3 low-pass filter 4 A / D converter 5 arithmetic processing unit 6 ROM 7 central processing unit 8 image processing unit 9 display

Claims (3)

(57) [Claims] 1. A microphone for converting a trainee's uttered sound into a voice signal, an analysis means for calculating a discriminant function coefficient from a characteristic parameter of the voice signal such as an LPC cepstrum, and a characteristic parameter of standard voice data. Storage means for preliminarily storing the coefficient of the discriminant function, judgment means for performing a binary judgment of a consonant by comparing the coefficient calculated by the analysis means with the coefficient stored in the storage means, and display means for displaying the judgment result An utterance speech training device characterized by comprising: 2. The utterance utterance according to claim 1, wherein said analysis means calculates a coefficient of said discriminant function after performing a binary selection based on an average zero-cross value of a consonant section of said audio signal in advance. Training device. 3. When the discriminant function of the standard data is Gab (x), x is an LPC cepstrum, k is a constant, and a and b are binary, | Gab (x) | <k The utterance utterance training device according to claim 1 or 2, wherein the binary determination is performed except for (1).