JPH02278295A - Voice pattern registering system and voice recognizing device - Google Patents

Abstract

PURPOSE:To execute a correct recognition even in a noise by eliminating a noise part from a stationary part of a noise to which a sudden noise is applied and replace it with a correct pattern. CONSTITUTION:A voice is collected by a microphone 1 and converted to an electric signal, the converted electric signal is brought to frequency analysis by a filter bank 2, and from a result of analysis, a position of a large frequency component is fetched by a peak detecting part 4, and whether the position of a large frequency component is continued for a prescribed time or not is decided by a time counter 5 and a comparing part 6. In the case plural spectrum stable parts continued exceeding the prescribed time exist, whether the adjacent spectrum stable parts are the same kind or not is decided by a pattern comparing part 8, and at the time of the same kind, they are decided that one spectrum stable part is continued and corrected by a pattern replacing part 9 and brought to pattern registration. In such a way, even in a noise, a correct recognition can be executed.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pattern registration method for speech recognition, and a speech recognition device using 28 patterns created by the pattern registration method.

As moxibustion-based speech recognition devices approach real eyebrows, research has been conducted to consider their use in real environments. In particular, in the case of voice recognition, ambient noise is a problem, and recognition cannot be achieved unless it is possible to distinguish between voice and noise! Several methods have been proposed for recognition in I sounds. For example, the spectrum of noise when no audio is input is stored in memory, and when audio is input, the noise component is subtracted from the spectral component and then tL& is performed (Proceedings of the Acoustical Society of Japan, 1977) March P141-1
42 3-4-6. Experiments on noise processing for word recognition)
is well known. However, this method may have the opposite effect if the spectral components of the noise stored in memory change. In addition, you can measure the loudness of the noise before the audio is input, and when a louder sound is input or the noise is lower than the measured value, you can measure the noise level at the beginning and end of the audio, respectively. There are things that can be extracted from the noise. However, with this method, if the noise increases during voice production, the voice level after detecting the start point will not fall below a pre-measured threshold. This causes a phenomenon in which the end of the audio cannot be found. Also, 2
It has two audio input means, one for audio and one for ambient noise.
On the other hand, the main input is ambient noise, and the filter characteristics are changed to minimize the error between the two.
There is a so-called adaptive filter (Voice Research Group Materials 581-81 p651.
noise removal using a processor). This method is highly accurate and has the advantage that even if the input noise is not constant, a signal with reduced noise can be extracted.

However, this method requires a large amount of calculation to converge to the minimum error.1 If data is acquired every 10 m5 as in normal speech recognition, it is difficult to converge within the acquisition time. No matter what method is used, it is impossible to deal with sudden noises that occur while uttering voice.

The present invention has been made in view of the above-mentioned shortcomings of the prior art, and provides a pattern creation method for modifying and recognizing characteristic patterns even when noise occurs during voice input. It was done for that purpose.

In order to achieve the above object, the present invention includes a part that collects audio and converts it into an electrical signal, a part that analyzes the frequency of the converted electrical signal, and a part that analyzes the frequency components from the analyzed results. If there are multiple parts where the position where the frequency component is large continues for a predetermined time or not (hereinafter referred to as spectral stable part), the adjacent spectrum This method is characterized by determining whether the stable parts are of the same type, and if they are of the same type, correcting them as one continuous spectral stable part and registering the pattern. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention, in which 1 is a microphone, 2 is a filter bank section,
3 is a sampling section, 4 is a peak detection section, 5 is a time counter, 6 is a comparison section, 7 is a memory section, 8 is a pattern comparison section, 9 is a pattern replacement section, and 10 is a registration memory section. This was done based on the fact that recognition results can be narrowed down with fairly high accuracy based on the constant arrangement of phonemes contained within and the order of appearance.
Specifically, there is a part that collects audio and converts it into an electrical signal, a part that analyzes the frequency of the converted electrical signal, and a part that extracts the major positions of frequency components from the analyzed results and calculates the major positions of the frequency components. If there is a part that determines whether or not the spectral stability continues for a predetermined time and a part that continues for a predetermined time or more (hereinafter referred to as spectral stable parts), it is determined whether the adjacent spectral stable parts are of the same type. judge,
When they are of the same type, they are corrected and registered as a continuous spectral stable region. To explain this using Figure 1, first,
Voice is input through the microphone 1 and converted into an electrical signal. As a frequency analysis section that performs frequency analysis of the converted signal, for example, band pass filter group 2 may be used. Alternatively, sample the waveform 3 and then perform FFT
The frequency may be converted by A peak detection unit 4 detects a frequency with a large component among the results of frequency conversion.

Here, we compare the magnitudes of the components on the frequency axis and extract the maximum value, but other methods, such as taking the difference between adjacent data on the frequency axis and determining the peak where the sign is reversed, are used. It goes without saying that this method is also good. Next, a time counter 5 measures the length of time in which the extracted local maximum values continue, and a comparator 6 compares this with a predetermined time length to check whether it is longer than that. The recommended length of time is about 30m5.

A mark is placed on a portion longer than this in the memory section 7.

What is marked here is the vowel and /S/.

These are phonemes such as /f/ and /l f/. Moreover, the stationary stable parts (spectral stable parts) of these phonemes are marked. For example, when two different vowels are consecutive, the pattern loses stability at the transition between vowels, so it is divided into two. By comparing adjacent portions of the marked portions in the pattern comparing section 8, it is determined whether two adjacent portions have the same phoneme or not. This can be determined by determining the degree of similarity between adjacent parts and determining whether the value is greater than a certain value. If they have the same phoneme, we connect the two parts and combine them into one. When grouping the marks, you can make them consecutive, or you can insert the same vowel pattern to rearrange the marks. It is also effective to register such a pattern in this way, first limit the recognition target by the pattern before normal recognition, and then perform normal Ly22. If the number is not very large, or if there are no vowels or constant consonants in the same position and combination, Ly& identification results can be obtained as is.

If a sudden noise occurs during vowel pronunciation, the noise will enter the vowel, but the vowel can be restored by the above operation. Furthermore, when noise is mixed into a non-stationary consonant, or when it is attached to the end of a vowel, it is possible to return to the same state as when noise was not mixed in by correcting the mark as described above.

Note that the example shown in FIG. 1 is an example in which this method is used for preliminary recognition, and there is no restriction on what kind of method the recognition unit uses. Even in the preliminary Ly3 knowledge, the same pattern in the dictionary is retrieved from the arrangement of marks, so there is no particular need for pattern matching. Of course, pattern matching may be used.

Of course - IchikumaAs is clear from the above explanation, according to the present invention, it becomes possible to remove the noise part from the steady part of speech to which sudden noise has been added and replace it with the correct pattern. , I was able to make correct recognition even in noise.

[Brief explanation of drawings]

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention. 1...Microphone, 2...Filter bank section,
3 sampling section, 4... peak detection section, 5... time counter, 6... comparison section, 7... memory section, 8.
...Pattern comparison section, 9...Pattern replacement section, 1
0...Registered memory section. Figure 1 Patent applicant Ricoh Co., Ltd.

Claims (1)

[Claims] 1. A part that collects audio and converts it into an electrical signal, a part that analyzes the frequency of the converted electrical signal, and extracts the major positions of frequency components from the analyzed results. If there are multiple parts that determine whether or not a large position continues for a predetermined time and parts that continue for more than a predetermined time (hereinafter referred to as spectral stable parts), adjacent spectral stable parts are of the same type. A voice pattern registration method characterized by determining whether or not they are the same, and registering the patterns by correcting them as if they are a continuous spectral stable part if they are the same. 2. A speech recognition device using a pattern registered by the pattern registration method according to claim 1.