JPS6342278B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a triggering method for a speech recognition device, and in particular, it is an object of the present invention to provide a triggering method for a speech recognition device that can reduce malfunctions caused by noise or the like. Generally, a speech recognition device passes an audio signal picked up by a microphone through a microphone amplifier, and then passes it through a feature extraction section configured using a bandpass filter and a full-wave rectification circuit. AD by switching sequentially using a multiplexer.
The audio data is supplied to a converter, converted into a digital signal by this AD converter, stored in the storage unit as audio data, and again, the audio data obtained through the above-mentioned circuit is compared with the audio data stored in the storage unit. It was configured to recognize speech through processing. In this case, the digital conversion of the audio signal from the bandpass filter in storage mode is performed at high speed, and the time interval (hereinafter referred to as frame interval) until the digital conversion of the audio signal from the bandpass filter in storage mode is performed at high speed. The captured data (called frame data) ranges from several milliseconds to several tens of milliseconds, and speech is recognized by processing this data. Although there are various problems with the above-mentioned speech recognition device, the following are cited as factors that cause malfunctions during speech recognition. (1) Changes in the speed of the speaker's speech production. (2) Activation by words other than the recognition target word. (3) Activation due to environmental noise (especially pulsed noise). Among the above items, (1) and (2) can reduce misrecognition by performing advanced analysis processing on the collected data, but (3) is when the voice is input during a malfunction due to noise. Because misrecognition or rejection sometimes occurs, it has been difficult to reduce misrecognition. The current trigger method is considered to be the cause of malfunctions due to noise. The current trigger method that is commonly used is to capture data when the energy of the input signal exceeds a set threshold level, and when the energy of the input signal becomes below the threshold level, the data is captured for a certain period of time. It is a method that stops data acquisition after a certain period of time has elapsed, and there are differences in whether the energy of the input signal is processed over the entire band or only in a specific band, but basically the energy of the input signal is processed. We deal with the size of. Therefore, from the above content, when noise such as noise is input at a certain level or higher, data capture is started. For this reason, when noise is superimposed on speech, or when speech recognition is performed in a place where a lot of shocking noise is generated, there is a drawback that malfunctions inevitably increase. The present invention solves these conventional drawbacks, and uses not only a comparison of the energy levels of input signals but also a difference in frequency characteristics as a feature of voice as a criterion for determination. Generally, when operating a recognition device in a noisy environment, if the background noise input signal is set to below the trigger level as an initial setting, malfunctions due to stationary noise are less likely to occur, but non-stationary noise such as impact sounds and periodic noises is difficult to predict, and the sound is likely to cause a frame data capture operation. Even if this frame data capture operation is performed, frame data capture can be made difficult by determining whether the signal is voice or noise based on the duration of the signal. However, when noise occurs immediately before voice generation, the noise first triggers the noise frame data, and then the voice frame data is collected, which tends to cause erroneous recognition due to processing. Since this erroneous recognition operation continues until the processing is completed, the acquisition of new audio signals is rejected. The spectrum of noise, especially impact sound, has flat characteristics over a wide frequency band. The spectrum of speech, on the other hand, differs depending on the words produced because the length of the vocal tract is constant in the tube, but it produces multiple peaks (hereinafter referred to as formants) on the frequency axis. This formant has little frequency variation in the vowel part of speech,
It also has characteristics that are constant over time. In contrast,
In the present invention, the initial stationary part frame data of the input signal is extracted and the city value distance from this extracted frame data is determined, so it is possible to instantly determine whether the input signal is noise or speech to be recognized. However, it has the advantage of being able to enter a re-input system when there is noise. Further, even if the input signal is a voice, the city value distance becomes large for words with different initial stationary parts, which has the advantage that malfunctions caused when different voices are input can be reduced. Here, there are the following two types of speech recognition to which the present invention is applied. One is the case of recognition using the registered limited word method for specific speakers, and the other is the case of recognition using the limited word method for non-specific speakers. In the former case, at the time of audio registration, the city value distance between the extracted frame data of the regular frames at the beginning of the registered words is determined. In the latter case, the city value distance from the phonological data at the beginning of the word to be used is determined. Hereinafter, the present invention will be explained with reference to drawings of embodiments. 1 and 2 show an embodiment of the present invention. In FIG. 1, an audio signal picked up by a microphone 1 passes through a microphone amplifier 2 and then passes through band filters 3-1 and 3-2. ......3-n, the audio signal is converted to DC by the full-wave rectifier circuit 4-1, 4-2...4-n, and the multiplexer 5 converts the audio signal into direct current through each band filter 4-1, 4-2...4- After the outputs of n are sequentially switched, they are digitally converted by the A/D converter 6 and stored in a storage section (not shown) as audio data. In this type of device, the band filter group is digitally converted at high speed, but the time interval (hereinafter referred to as frame interval, and the captured data group is called frame data) is several times until the band filter group is digitally converted again. It takes from m seconds to several tens of m seconds, and speech is recognized by processing this data. Here, in a registration mode in which data from the A/D converter 6 is taken into the storage section, or in a recognition mode in which data from the A/D converter 6 is collated with data registered in the storage section, as shown in FIG. An audio trigger based on the flowchart is executed.
In FIG. 2, numeral 11 is a step for taking in data of one frame, numeral 12 is a comparison step between the level of the input signal and the threshold level, and numeral 13 is a determination step using the level trigger. As a result of the judgment in this judgment step 13, if the threshold level is exceeded, the process proceeds to the next frequency trigger step 14.
If the level is below, return to step 11. In the frequency trigger step 14, the steady portion of the input signal is determined and used as a trigger frame. Also,
The street value distance between the limited word and the trigger frame, which has been separately analyzed and found, is determined, and the magnitude of the distance is used to test whether the limited word is phonological or not, and a determination is made in the next determination step 15. Furthermore, if the steady part of the input signal cannot be extracted even after several tens of milliseconds, it is regarded as noise and the process returns to step 11. If the test determination result in determination step 15 is that the city value distance is small, the process proceeds to step 16 of the next known data collection and recognition process, and the processed result is output. Next, a method for detecting a steady state in the frequency trigger step 14 will be described. The data frame captured through the determination step 13 using the level trigger is expressed as the fifth frame _XJ (when _frequency analysis is performed on the I channel, and the maximum value in X _J is L _J ). Furthermore, the distance between the J-th frame and the K-th frame is D _JK =

[Formula] Alternatively, define D _JK =〓 i|X _J −X _K |・1/L _J. Next, we will describe how to find the actual potential. For example, focusing on the J-th frame, and assuming that the potential of the J-th frame is P _J , the average value of the distance between the N frames before and after it is expressed as P _J = 1/N _J 〓 ^K=JN D _JK・d/ L _J . However, N<J. d is the gain coefficient. When the potential P _J calculated from the above formula becomes less than the city value distance setting threshold level, that J
Set the frame as the trigger frame. If the distance does not fall below the city value distance setting threshold level even after capturing up to 50ms, stop the process and proceed to step 1.
Return to Next, a simplified method will be described. This method allows high-speed processing because it does not require complicated exercises. However, since the accuracy of stationary part extraction by processing is not achieved, it is desirable to use the city value distance setting threshold level raised. The contents of this simplification method are as follows in level trigger determination step 13.
After receiving OK, one frame of data is captured several tens of milliseconds later, and this is used as a trigger frame. As is clear from the above explanation, according to the trigger method of the speech recognition device of the present invention, in order to judge the effectiveness of data capture as pre-processing for speech recognition, energy comparison between registered speech data and input speech data is performed. At the same time, the frame data of the stationary part at the beginning of a word is extracted and the difference in frequency characteristics of that part is incorporated into the judgment criteria. It is possible to accurately determine whether it is noise or speech to be recognized.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the speech recognition device of the present invention, and FIG. 2 is a flowchart for explaining a triggering method of the same device. 1...Microphone, 2...Mic amplifier, 3
...bandwidth filter, 4...full wave rectifier circuit, 5...
Multiplexer, 6...A/D converter.

Claims (1)

[Scope of Claims] 1. When performing voice recognition by collating input voice data obtained from a microphone with registered voice data registered in advance in a storage unit, as pre-processing of the voice recognition, the input voice data is In addition to determining the energy of the input audio data and comparing it with a predetermined level, the constant part frame data of the beginning of the word is determined by determining the city value distance between the preceding and following frames of the input audio data, and the constant part frame data of the beginning of the word of the input audio data is determined. A triggering method for a speech recognition device that determines the effectiveness of data capture by determining the street value distance between the prefix frame data and the steady frame data at the beginning of a word of the registered speech data. 2. The method for triggering a speech recognition device according to claim 1, wherein the stationary part frame data is formant spectrum or phoneme data.