JPS6165297A - Voice recognition system - 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]

[Field of Application of the Invention] The present invention relates to a speech recognition device, and particularly to a speech recognition method suitable for preventing erroneous recognition due to non-stationary noise. BACKGROUND OF THE INVENTION Speech recognition devices typically monitor the power of input speech and
A speech section to be recognized is cut out and recognition processing is performed within that section. Therefore, when a voice or noise with a certain level of power is input from a microphone or the like, a recognition result is output for that voice section. For this reason, the voices of other people in the vicinity, one's own tweets, etc. are output as tlili-incapable. Furthermore, since the matching process is performed on the entire input voice section, if you say ``Eh~Tokyo'' when you want the word ``Tokyo'' to be recognized, it will not be recognized. In this manner, conventional recognition devices may misrecognize or fail to recognize due to unsteady noise or human speech errors. As a countermeasure against this problem, for example, there is a method of registering noise in advance and not outputting a result for that sound, as described in Japanese Patent Application Laid-Open No. 58-23097 ``Speech Recognition Device''. Although it has been proposed, there are limitations in terms of registration and number of recognized words, and it cannot be said to be sufficient. [Object of the Invention] The object of the present invention is to prevent misrecognition and g
The object of the present invention is to provide a highly practical voice recognition method with a high crystallinity rate while avoiding confusion. [Summary of the Invention] The present invention provides a speech recognition method that uses continuous DP processing (continuous matching processing) and takes advantage of the advantage of continuous DP in that it does not require speech section extraction. In other words, matching with the registered pattern is not performed only in the input speech section, and if there are no suitable candidates within the speech section, it is assumed to be noise or erroneous pronunciation and recognition continues, thereby eliminating unnecessary speech or non-speech. To provide a speech recognition method (hereinafter referred to as word boxing method) that recognizes only target input speech without being affected by stationary noise. [Embodiments of the Invention] FIG. 1 shows the configuration of a speech recognition device to which the present invention is applied. In the figure, 1 is A that digitizes the input audio.
/D conversion unit, 2 is a voice analysis unit that obtains voice power for each frequency band and extracts a characteristic pattern; 3 is a voice analysis unit that extracts a characteristic pattern;

[
- A distance calculation unit that calculates the distance between the standard pattern registered in advance in the pattern memory 5 and the input feature pattern; 4 is a continuous DP process based on the distance calculation result;
A verification unit 6 that performs verification by H is a control unit that controls each part of the 2.5° 4 and also selects and recognizes the verification results obtained from the verification unit 4. In addition, continuous DP
Regarding the verification section by processing, please refer to Japanese Patent Application Laid-Open No. 57-83880.
Please refer to Japanese Patent Publication ``Matching Method Switching Control Method'' and Japanese Patent Application Laid-Open No. 55-2205 ``Real-time Continuous Audio Data Planning Apparatus.'' Now, in the continuous DP method applied to the present invention, when continuous audio is input, each input frame is matched with each pre-registered standard pattern, and the pattern is similar to a certain standard pattern. There is a voice and pine along with its standard pattern number. The matched audio section and matching score (hereinafter referred to as DP value) are output. That is, matching results can be obtained for each input frame. Therefore, when matching an input cover pattern with a registered pattern, it is not necessarily necessary to accurately cut out the input speech section and perform matching on the entire section. Therefore, when a voice is input as shown in chart 70 in the f82 diagram, continuous DP processing is performed continuously until the voice section ends, and recognition results are given priority to those with a small DP value regardless of the input voice section. . At this time, as a validity check of the recognition result, short speech section candidates that are likely to occur in %絖DP are deleted to check for erroneous recognition due to profanity. Alternatively, the maximum and minimum matching sections of the 1' leaf to be m3 sweetfish may be determined in advance, and recognition may be made only by those words. For example, when the standard pattern of "Douki tsu" is registered, if the voice is input as "Nie... Touki 1", the voice section will be as shown in (C) in Figure 3 ◇ This section The matching value with the standard pattern "Touki Tatsu" changes as shown by α in FIG. 6, and at the minimum point A, the voice section and DP value that are matched with the pattern number of "Touki Sotsu" are output. Here, although the matched speech section is (α), recognition results can be obtained regardless of the input speech section. On the other hand, the noise "knee" is not registered, so D
It is rare that a minimum point of P value appears. Even if a minimum point like point B of β in FIG. 3 appears in relation to the registered word, it may result in a low DP value and will be removed. Also, if there is an input such as [Kyoha......] as noise, continuous DP processing is continuously performed during recognition, so the "key eyes" of "Tokyo" and "
Similar to ``Kyou'' in ``Today'', a P minimum point may appear, such as γ of 0 in FIG. At this time, the matched section is also output, and a matching length like <b> is obtained. In this case, when checking the validity of the recognition result,
Because the matching interval is shorter than specified, it is excluded from the candidates. In this way, by removing candidates that are matched in a speech interval shorter than the reference value compared to the registered words,
In continuous DB processing. It is possible to prevent misrecognition due to partial matching, which is also a drawback. In this case, the matching value and appropriate matching interval are determined as follows. The matching value must be determined depending on the constraints on the operation bits of the matching unit, that is, the dynamic range, but it has been found from previous studies that 10 degrees (8 degrees) is sufficient for the normalized output. Considering margin for determination, the ratio of the correct matching value to the maximum directness needs to be 4 to 5 times, so based on this constraint, the threshold value for determining the correct answer is determined to be about 200. Furthermore, the matching interval must be determined depending on the word length being handled, and is 0.8 to 1.0 for each word length stored in the standard pattern table for each word.
The correct answer is approximately twice the range. This means that the expansion and contraction in the time axis direction that can occur in human vocalization is ±20
This is because previous experiments have shown that it is about %. It goes without saying that these matching values and matching intervals are subject to change depending on the system. [Effects of the Invention] According to the present invention, words generated by mistake during speech recognition, surrounding non-stationary noise, the breath of the speaker, etc. will not cause unrecognizability or misrecognition, and the level of recognition will be reduced to a practical level. It has a remarkable effect on improving the recognition rate of speech recognition devices.

[Brief explanation of drawings]

Fig. 1 is a block diagram of a speech recognition device to which the present invention is applied, Fig. 2 is a flowchart for explaining the recognition method, and Fig. 3 is a flowchart for explaining the recognition method.
The figure is a waveform diagram showing the relationship between voice sections, voice power, and DP values. L: A/D Separation 2: Person Voice Editing Section 3: Old M Edit $IS 4: Part Number: Keiko Hater Memorandum: Wholesaler (4Φ Right Dopa) Agent Patent Attorney Akira Takahashi Husband's Day (3 Figure 2 Figure 3

Claims (1)

[Claims] In a speech recognition device consisting of a speech input section, an analysis section, a distance calculation section, a matching section, a standard pattern memory, and a control section, the matching section is equipped with a continuous matching means, and the speech digitized by the speech input section is matched. The data is input to the analysis section and features included in the speech are extracted, and the control section uses the data to detect speech sections.On the other hand, the registered pattern stored in advance in the standard pattern memory and the input speech pattern are is continuously verified in the verification section, and in the control section,
A matching value and a matching interval that are output from the matching section are input, and only when the matching value is smaller than a preset threshold and the input speech interval and the matching interval have a predetermined relationship, the word is recognized as a recognition result. A voice recognition method featuring: