JPS6336679B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a word speech recognition device based on speech zero-cross information.

Many of the speech recognition devices are large-scale devices that use a large number of bandpass filters, spectrum analysis means, etc., and are complex in structure and expensive. On the other hand, there is a demand for a speech recognition device that is simple in configuration and inexpensive, even though the number of words to be recognized is small and the recognition rate is low.

One of the simple speech recognition methods uses zero cross information of speech waveforms. A comparator detects zero crosses in the audio signal, and a computer or the like is used to process it, thereby extracting and recognizing the characteristics of words. In reality, if you simply take the zero-crossings of the input signal, as shown in Part 1, zero-crossings will also be detected in areas other than audio due to noise contained in the input signal. Therefore, in addition to the zero-cross information, the envelope of the input signal is detected, the section where the envelope exceeds a predetermined value is regarded as a speech section, and the zero-cross information in that section is recognized. use However, according to this method, the circuit for envelope detection becomes complicated.

Other methods of separating voice sections include adding hysteresis to the comparator or shifting the reference voltage of the comparator from the average value of the input signal so that the comparator does not respond to noise whose amplitude is small compared to the voice. It is possible to do so. FIGS. 2 and 3 respectively show the input waveform and the output waveform of the comparator when the comparator has hysteresis and when the reference voltage of the comparator is shifted. By appropriately setting hysteresis or appropriately shifting the reference voltage for such a noise-containing input waveform, it is possible to prevent a change in the output state of the comparator during a section without audio. Therefore, recognition is performed using information on sections where the output state changes. Strictly speaking, the information obtained at this time is not zero-cross information, but in the case of audio signals whose amplitude is larger than that of noise, the hysteresis of the comparator and the deviation of the reference voltage do not have much of an effect, and almost the same information as zero-cross information can be obtained. It will be done. However, zero-cross information may be lost for consonant parts with small amplitudes. In a recognition method using zero-cross information, it is desirable that there is a large difference in the number of zero-crosses between a portion such as a vowel where there are relatively few zero-crosses and a portion where there are many zero-crosses such as a fricative. However, since the amplitude of the fricative part is small, the above method has the problem that zero cross information is lost, the characteristics are difficult to appear, and sufficient recognition accuracy cannot be obtained.

Therefore, an object of the present invention is to improve recognition accuracy in a device that uses a comparator to obtain zero-cross information and performs word speech recognition using the information.

The present invention provides a first comparison means for comparing the average level of the audio signal with a reference value, and a second comparison means for comparing the audio signal with a reference value that is deviated from the average level of the audio signal by a predetermined amount. means for detecting a voice section from the output of the second comparing means;
It consists of recognition means that performs voice recognition from the zero cross information output from the first comparison means in the detected voice section, and it is possible to improve recognition accuracy with a simple configuration. Examples of the present invention will be described below.

FIG. 4 shows the configuration of an embodiment of a speech recognition device according to the present invention. In the figure, an audio signal input from a microphone 1 is amplified by a preamplifier 2, a high-pass filter 3 for removing pitch components, and an offset adjustment section 4 for adjusting the average value of the input audio signal to zero level. After passing through, it is input to comparators 5 and 6. Comparator 5
This method essentially compares the audio signal with a reference value that deviates by a predetermined amount from the average level of the audio signal. It consists of not reacting when the situation arises. Here, instead of providing hysteresis, the reference voltage of the comparator may be shifted from the average value (zero level) of the input signal, as shown in FIG. 2, so that it does not react to noise. On the other hand, the comparator 6 is not provided with hysteresis, and the reference voltage is made to match the average value of the input signal.

The outputs of comparators 5 and 6 are input to the input terminal and interrupt input terminal of microcomputer 7, respectively. Microcomputers are usually about 4 bits in order to keep the overall price as low as possible. Note that the RAM 8 is externally attached when the RAM capacity inside the microcomputer is small.

The microcomputer 7 sequentially reads the input from the comparator 5 and increments the contents of the memory allocated for counting by one each time the state of the input changes. During this time, an interrupt is generated each time the input from the comparator 6 changes from high level to low level, and the contents of the memory allocated for counting by the comparator 6 are incremented by one in the interrupt processing routine. Furthermore, by generating a timer interrupt every 10msec, 10msec
(This is considered one frame.) Comparator 5 inside
and 6, the number of times the output state changes are obtained. The number of times the output state of the comparator 6 changes during one frame is called the number of zero crosses, and the number of times the output state of the comparator 5 changes is called the number of quasi-zero crosses. The flow of these processes is as shown in FIG. 6, 61, 63, 62, and 65.

First, a speech section is detected using the quasi-zero cross number. As mentioned above, since the comparator 5 is designed not to react to noise, the quasi-zero cross number is not zero in frames of voice sections, but is considered to be zero in frames without voice. Therefore, it is possible to detect speech sections using the quasi-zero cross number. However, even within a speech section, there are silent sections before plosives, so it is insufficient to simply consider a portion where the number of quasi-zero crosses is not zero to be a speech section. Therefore, the voice section is determined for the sequence of quasi-zero crosses using the method shown in FIG. When a predetermined number (for example, 5) of frames in which the number of quasi-zero crosses is not zero are consecutive, the first frame in which the number of quasi-zero crosses is not zero is taken as the beginning of the voice section. After a voice section starts, if a predetermined number (for example, 30) of frames in which the number of quasi-zero crosses is zero continue, the last frame in which the number of quasi-zero crosses is zero is determined as the end of the voice section. If the speech section detected in this way does not reach a certain length (for example, 20 frames), it is too short to be word speech and is not considered as a speech section. Furthermore, if the detected voice section exceeds a predetermined length (for example, 120 frames), it is too long to be a word voice, so it is not considered as a voice section.

In parallel with the voice section detection 64, the number of zero crosses is detected.
Writing 67 to RAM 8 is performed. As shown in FIG. 6, it is determined whether or not to write the number of zero crosses based on the value of the number of quasi-zero crosses. The portion surrounded by broken lines in FIG. 6 is the processing within the microcomputer. This situation is specifically shown in FIG. 7. Here, N is the number of frames written to the RAM 8, and M is the number of frames in which the number of quasi-zero crosses is zero. First, when a frame in which the number of quasi-zero crosses is not zero appears, the number of zero crosses for each frame is sequentially written from the beginning of the area allocated for storing the number of zero crosses in RAM 8 by the write control 66 for writing the number of zero crosses to the RAM. Write. After starting writing, a predetermined frame (here 5 frames)
If a frame with a quasi-zero cross number of zero appears before reaching , it is determined that it is not a speech section, and a new starting point of the speech section is searched. After the audio section starts,
When the number of quasi-zero crosses reaches 0 for a predetermined frame (30 frames in this case) consecutively, writing ends as the voice section ends, and it is determined whether the written voice section is longer than a predetermined frame (20 frames in this case), and if so. If not, it is not regarded as a voice section and a new start point of the voice section is searched. If it is determined that it is a voice section, counting and writing of the number of zero crosses and quasi-zero crosses is finished, and the process moves to the next process. Also, the predetermined frame (here, the sum of the maximum value of the voice section, 120 frames, and 30 frames for detecting the end of the voice section)
150 frames) If the end of the voice section is not detected even after writing, it is not regarded as a voice section and a new start point of the voice section is searched.

When it is determined that it is a voice section and the writing is completed, the end point of the voice section is a frame that goes back a predetermined frame (in this case, 30 frames) from the end of the written frame. When the voice section is determined in this way, 16 frames are sampled at equal intervals starting from the number of zero crosses within the voice section. If the audio interval is n frames, the sampled frames are 1, n/16+1, 2・n/16+1, 3・n/
16+1,..., 15·n/16+1st frame. Since the microcomputer used here is a 4-bit machine, dividing by 16 is achieved by ignoring the lower 1 word of the memory allocated for frame calculation, and the above calculation can be performed only by addition. . Furthermore, since frame numbers and RAM addresses correspond in good order, sample frame calculations are actually calculated based on RAM addresses rather than frame numbers.

Recognition is performed based on the number of 16 zero crosses for each word obtained in this way. First, the number of 16 zero crosses mentioned above for each word to be recognized.
Register in RAM8. This registered pattern is called a standard pattern. After inputting the registration mode into the microcomputer 7 by inputting a switch, when a word is uttered from the microphone, the RAM
16 zero crosses were sampled from the number of zero crosses written in 8, and the RAM was allocated for standard pattern registration.
Data is written sequentially from the beginning of the area.

When registration is completed, the mode is set to recognition by inputting a switch, and then voice is input from the microphone. The i-th value of the 16 zero-cross numbers sampled from the input voice is Si (i = 1, ..., 16), and the i-th value of the j-th registered word of the standard pattern is Tji.
When (i = 1, ..., 16, j = 1, ..., W (W is the number of registered words)), the distance Dj between the input voice and the j-th word is Dj = ₁₆ 2 ^{i = 1} 1Tji −Si1 shall be given. Dj is calculated for each standard pattern, and the word corresponding to the standard pattern that gives the minimum value of Dj is taken as a recognition result and displayed on a display element (not shown) such as an LED.

As described above, this embodiment uses two comparators for detecting speech intervals and calculating the distance between words, and a microcomputer to detect speech intervals with a simple circuit, and also detects consonant parts with small amplitude. It is also possible to detect zero crosses with relatively high accuracy. Also, since the processing for recognition is easy,
It can be realized with a microcomputer of about 4 bits.

If the frequency of zero crosses is very high due to noise, etc., and there is a risk that the microcomputer's interrupt processing may not be able to keep up with the time, it may be necessary to provide a slight hysteresis to the comparator 6, or set the reference voltage to the audio signal. may be biased from the average value.

[Brief explanation of the drawing]

Figure 1 is a waveform diagram to explain the operation of a comparator in a conventional word speech recognition device, Figure 2 is a waveform diagram to explain the operation of a comparator with hysteresis, and Figure 3 is a waveform diagram to explain the operation of a comparator with hysteresis. A waveform diagram for explaining the operation of a comparator with
Fig. 4 is a block diagram of a word speech recognition device according to an embodiment of the present invention, Fig. 5 is a flowchart of speech section detection in the same embodiment, and Figs. 6 and 7 show how the number of zero crosses is written to RAM. This is a flowchart. 1...Microphone, 2...Preamplifier, 3...
...High pass filter, 4...Offset adjustment section,
5, 6...Comparator, 7...Microcomputer, 8...RAM.

Claims (1)

[Claims] 1. A first comparing means that compares the audio signal with a reference value corresponding to the average level of the audio signal, and a second comparing means that compares the audio signal with a reference value that deviates from the average level of the audio signal by a predetermined amount. and a means for detecting a speech interval from the output of the second comparison means, and a word by comparing the zero cross information of the output of the first comparison means in the detected speech interval with standard zero cross information. A word speech recognition device comprising a recognition means for performing speech recognition.