JPS5879297A - Voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition device, and more particularly to a monosyllabic speech recognition device using a balatgelaf as an auxiliary V-fiber device.

In this invention, artificial r: JIIly et al.'s tongue movement information is dynamically analyzed to classify input speech into four consonant groups, and the classification results are converted into speech il! by four types of =-do. !
Transfer to fiber processing equipment. And the voice recognition device is a proprietary kll.
Among the identified single syllables having a predetermined degree of similarity or higher, the one that matches -T with the consonant group and has the highest degree of similarity φ is output as the yu code of the input speech. As a result, among the consonants that were previously difficult to identify, consonants with different tongue movements can now be clearly identified! &! ! ! As a result, we were able to improve the AM rate by 2 to 5 points compared to the conventional method, although it differs from person to person.

That is, this first invention consists of a first parameter extracting device 2 for extracting feature parameters of monosyllabic speech (VS) inputted through a microphone P-phone 1, and a paratogelatus from an artificial palate 5 as shown in FIG. A bat graph analysis device 4 that analyzes the signal PG and outputs the corresponding consonant group code GC, and a feature parameter FP extracted by the parameter extraction device 2 and a monosyllabic feature parameter stored in advance in an internal storage device. Among the monosyllabic codes with a predetermined degree of similarity or higher obtained from and k, the consonant group code GCk from the batgraph analysis device 4 matches the consonant part and has the highest degree of similarity, as the monosyllabic code Me. -1 [
As shown in FIG. 2, the parameter extraction device 2 includes a preprocessing circuit 4 that amplifies and preprocesses the audio signal
, a band pass filter group n that divides the preprocessed audio signal VSA into bands having different center frequencies, a channel selection circuit that sequentially selects each divided band signal by a control signal 081, and a selected Channel signal CH
The control signal 082 includes a sampling circuit for sampling the signal at a predetermined timing according to the control signal 082. In addition, the balatogelaf analysis device [4 includes an artificial palate control device 41 that sends out a paratogelaf signal PG from the artificial palate 5 at regular intervals, a data conversion circuit 42 that converts this output information into required data, and a voice Ial111! The device consists of a consonant group classification device that outputs a consonant group code GC in response to a vowel part start signal C83 and a transfer request signal C84. Note that the voice recognition device 3 is v4.
A keyboard for verification and correction is connected, and display #6
1ala easily while watching! Corrections can be made.

In such a configuration (f'), the audio signal S from the microphone 1 is amplified and preprocessed by the preprocessing circuit 21 in the parameter extraction device 2, and is applied to a group of band pass filters B having mutually different center frequencies. Here, the audio signal divided into each band is sequentially selected by a channel selection circuit B, and sent to a subsequent sampling circuit U for sampling.

The timing of the channel selection circuit C and the sampling circuit C is controlled by the control signals C8l and C82 from the voice recognition device 3, and the voice signals divided into each band are arranged in series in time, and the digital information is It is output to the audio ignition device 3 as a feature parameter.

On the other hand, the paratogek analysis device 4 inputs only the consonant part of the utterance out of the paratogelaph signal PG from the artificial palate 5 at the time of utterance, and analyzes the movement of the tongue based on temporal information changes. In other words, the information from the artificial palate control device 41, which transmits bit information (for example, 63 bits in total) of the terminals distributed over the entire surface of the artificial mouth II8 in time series at regular intervals, is transmitted to the subsequent stage. Consonant group classification device 430 converts and organizes the data in accordance with the input format in data conversion circuit 42.

In this case, for example, the serial PIF column information is backed up in 8-bit units and sent out. Therefore, in the consonant group classification inscription, the input consonant group classification is performed in the following manner.

In the first place, $sotographs are usually used to understand the position and movement of the tongue, which is one of the important elements that determine consonants, and are widely used and studied from the standpoint of phonetics.
In this invention, information on the position and movement of the tongue is pattern-analyzed to classify consonants into four types of consonant groups. Here, when using the Balatgelaf, the uttered sounds can be broadly classified into those in which the tongue contacts the upper surface of the artificial palate 5.
#.

1・n・y′・*rI,%,1. %, 11 and
Non-contact 'w', 'to', 'h', 'm', %b
l.

If it is divided into %pI and %gI, it can be classified into sores, and even things that come into contact with the tongue can be classified more finely than the tongue position M. However, because there are individual differences in the movement of the tongue, and because this invention uses the Balatgelaf as a supplement for sound %I recognition, it is possible to classify it with certainty, and it does not play a sufficient role as an auxiliary device. It is classified into 4 m'l consonant groups as shown in Table 1. Table 1 also shows the grooving fiber opening ratio.

Therefore, the relationship between the movement of the utterance field y and the voice power is like that of the fourth position, but in this invention, only the Balatogelaph pattern corresponding to the consonant part is used as a Radograph pattern,
It is determined which of the four types shown in FIG. 4(c) this pattern corresponds to. In order to classify into these four types of groups, the one-consonant group classification device 6 continues to input baratogeraf information at regular intervals until it receives the vowel part start signal C83 from the speech recognition device 3. The above-described group classification operation is started from the time when C83 is input, and the classification result is prepared as a consonant group code GC unique to each consonant group and is put on standby. He is a voice recognition system 'fR' that has obtained unique recognition results by comparing %fine parameters.
When it receives the transfer request signal C84 from s, it outputs the prepared consonant group code GC to the speech recognition device f3.

The speech recognition device [13] stores the characteristic parameters of each predetermined monosyllable along with its monosyllable code in its built-in storage device, and determines which monosyllable the % fine parameter FP of the input speech from the parameter extraction device 2 is similar to. The monosyllabic code having a predetermined degree of 1II4EJ is selected. To be more specific, audio V! The identification storage 3 synchronizes the parameter extraction device 2 with control signals C81 and C82, and inputs the band ψ-divided audio signal as the characteristic parameter PP. Then, the input feature parameters FP are sequentially stored in the internal storage device, and at the same time, the vowel start pitch of the input speech is determined from the value of the feature parameters FP. Voice V identification device f11
When it is confirmed that the incoming voice has shifted from early utterance to vowel utterance, step 3 outputs a vowel part start signal C83 to the balat geraf analysis device 4 which has already been inputting balat geraf data at regular intervals. Thereafter, the voice II fiber device 3 successively inputs the characteristic parameters of the vowel part for 6 fixed times, and inputs the 4111 parameters of the pronunciation-4 food. Pre-processing up to similarity calculation, such as normalization of the deprived W% feature-parameter and partial extraction, is performed as required, and the similarity with a predetermined monosyllabic feature parameter is calculated by a so-called pattern matching method. As a result, one or more monosyllabic codes whose degree of similarity is equal to or higher than a predetermined value are selected, and at this point, the consonant group classification is finished, and a transfer request signal C84 is sent to the standby Balatgeraf analysis device 4. and enter the consonant group code OC. Then, the consonant part of the monosyllable =-do selected in the previous stage (for example, 1 to 1# → %kI, %, Cl-4%, #) is compared with the consonant group code GC, Among the selected monosyllabic codes, select one that matches the consonant group code GC, and if there is one, it is the recognition result, and if multiple codes match or temperament, the one with the highest Ijj elevation among them is selected. Let σ be the recognition result. The above procedure corresponds to input audio. After determining the monosyllabic code Me, the Shishisho fI
6 displays the character corresponding to that monosyllabic yycy.

Figure 5 shows how the single syllable is called in its final form, % da1
This is a schematic diagram illustrating the case where the user inputs the following utterance as an example. The center of the circle is % da I, which is the % signature parameter of the input voice, and the circle indicates a predetermined degree of similarity. However, if you do not use baratogerahu, '21', which has the highest degree of similarity, becomes a single syllable, so it becomes a komama, but as a consonant group code, 't, d,
Given n', % da # and % na in the circle
The one closest to the input voice is selected, and 'da' is correctly displayed.
9. In this way, by supplementarily using the information on the position and movement of the tongue from the paratogelaph, the recognition rate is improved by 2 to 5qII.

As described above, according to the present invention, paratogelaf is used as an auxiliary device for speech recognition, and consonants are grouped into 4 types, resulting in 9! Because it is designed to have a unique character, it is a monosyllabic IP with a clear voice! A rut can be achieved.

[Brief explanation of drawings]

FIG. 1 is a block configuration diagram showing an embodiment of the present invention, FIG. 2 is a block diagram showing details of an example of a parameter extraction device used in this invention, and FIG. 3 is a block diagram showing a detailed example of a parameter extraction device used in this invention. A block diagram showing an example in detail, FIGS. 4(a) and 4(b) are diagrams for explaining its operation, and FIG. 5 is a diagram showing how a single syllable is determined according to the present invention. 1...Microphone, 2...Parameter extraction device,
3... Speech recognition device, 4... Paratogelaf analysis device, 5... Artificial palate, 6V-display device.

Claims (1)

[Scope of Claims] a. A parameter extraction device for extracting characteristic parameters of an input single-tone cylinder sound and voice; The Balatgeraf analysis device to output, C0, a monosyllabic code with a predetermined degree of similarity or more obtained by comparing the feature parameters extracted by the parameter extraction device and the monosyllabic special parameters stored in advance in the internal storage device. a speech recognition device that outputs a consonant part that matches the consonant group code from the Balatgelaf analysis device and has the highest degree of similarity as the monosyllabic code; d. a monosyllabic code from the speech recognition device; An audio cabinet meeting device characterized by comprising: a display device that displays; and a display device that displays.