JPH0677198B2 - Speech recognition method - Google Patents

Description

The present invention relates to a voice recognition method for recognizing input voice information.

[Prior Art] The voice recognition device first A / D-converts the input voice,
The output is sent to the feature extraction unit. The feature extraction unit calculates the power information of the sound and calculates the spectrum information of the sound by using a method such as fast Fourier transform.

The type of information that the standard pattern stored in the reference pattern storage unit has is the same as the type of information calculated by the feature extraction unit. The calculation is performed for each same type of information amount of the standard pattern, and the final similarity is calculated by multiplying the obtained individual information amount by a predetermined value and adding them together.

[Problems to be Solved by the Invention] By the way, in a conventional voice recognition device, discrimination between voiced sound and unvoiced sound, further discrimination between unvoiced sound and unvoiced consonant, discrimination between vowel sound and nasal consonant in voiced sound. Etc. use the power information of voice, divide the frequency band into low band, middle band, high band,
This has been done by utilizing the ratio of frequency components included in the band.

However, if the input voice contains a lot of noise,
The power information of the consonant at the beginning of a word is often lost in the power of noise, and even in the consonant in a word, its spectrum is dragged by the spectrum of the vowels before and after, and qualitative information is lost, and consonants are lost. It was not easy to identify.

Moreover, the vowel / u / spectrum is very similar to the nasal consonant / m /, / n / spectrum, so the misidentification rate of these was also high.

[Purpose] The present invention has been made in view of the above-mentioned problems of the prior art, and a voice recognition is performed by preselecting a recognition result of voice information according to a relative relationship of a plurality of peak levels included in input voice information. It is an object of the present invention to provide a voice recognition method that increases the recognition rate of recognition processing and shortens the processing time.

[Means for Solving the Problems] In order to solve the above problems, the speech recognition method of the present invention includes the following steps. That is, in a voice recognition method for recognizing input voice information, a plurality of peak levels included in the input voice information are detected, a relative relationship between the plurality of detected peak levels is calculated, and the calculated A recognition candidate of the input voice information is preselected from the relative relationship of the peak levels.

[Operation] In the process of the present invention, a plurality of peak levels included in the input voice information are detected, and a relative relationship between the detected plurality of peak levels is calculated. Then, a recognition candidate of the input voice information is preselected from the calculated relative relationship of the peak levels.

Embodiments Embodiments according to the present invention will be described in detail below with reference to the accompanying drawings.

FIG. 1 is a block diagram of the voice recognition apparatus of this embodiment.

In the figure, 1 is a microphone that converts a voice into an electric signal in a voice input unit, and 2 is a sample that samples the voice converted into an electric signal every 5 to 10 msec and digitizes an analog to be quantized A It is a / D converter. Reference numeral 3 is a buffer memory for storing a part of the output of the A / D converter, and reference numeral 4 is a peak value detecting circuit for sequentially reading data from the buffer memory 3 to obtain a peak value. 4a is a central processing unit CPU in the crest value detection circuit 4, 4b is a RAM in which the program of the flow chart of FIG. 3 to be described later is stored, and 4c is a work area or a crest value to be described later. Used for d
RAM used as a buffer for (1), d (2), d (3)
Is. Reference numeral 5 denotes a peak value change calculation circuit for calculating a temporal change in the peak value. Reference numeral 6 identifies whether the input voice is a voiced sound or an unvoiced sound based on the value of the peak value obtained by the peak value change calculation circuit 5, and further identifies a silent voice and an unvoiced consonant, and a vowel nose consonant. This is a discriminating circuit. Reference numeral 7b is a standard pattern storage unit that stores standard patterns classified by peak value for each group, and 7a is a memory having a standard pattern storage unit 7b. 8 is a frequency range of 200-6000 hertz 8-3
It is a feature extraction unit that consists of a band-pass filter group divided into 0 channels and extracts features such as power signals and spectrum information. Reference numeral 9 is a buffer memory for storing the input voice feature amount until the standard pattern storage section classified by the peak value is selected, and 10 selects the standard pattern storage section 7b judged by the judgment circuit 6. It is a switch. Reference numeral 11 denotes a pattern matching unit which compares the standard patterns read by switching the input voice feature amount switch 10 and calculates the similarity between the two. Reference numeral 12 is a recognition result output unit that outputs the standard pattern having the maximum similarity calculated by the pattern matching unit 11 as a recognition result.

The operation of this embodiment will be described in detail below.

First, the input voice is converted into a digital amount by the A / D converter 2, and its output is sent to the peak value detection circuit 4 and the feature extraction unit 8 via the buffer memory 3. Although the sampling frequency of A / D conversion and the number of quantization bits per sample are variable, in this embodiment, sampling is performed at 12 kHz and 1 sample is 12 bits (1 bit out of 12 bits is a code bit). Quantized. In this case, 1 second of voice is represented by 12000 points of data.

FIG. 2 is a graphic representation of the output of the A / D converter 2.

Since the A / D conversion is performed in real time, the buffer memory 3 is placed in front of the peak value detecting circuit 4, and the peak value detecting circuit 4 sequentially reads the sample data from the buffer memory 3.

FIG. 3 is a flow chart for processing by the CPU 4a inside the peak value detection circuit 4.

Here, d (1), d (2), d (3) are buffer memories 3
This is an array for storing the data read from, and 1, 2, 3 represent the order of sample data. Also, d + is a positive peak value, d
− Represents a negative peak value.

First, in step S1, the first two data stored in the buffer memory 3 are read into d (1) and d (2), respectively. In step S2, it is judged whether all the data in the buffer memory 3 have been read, and when the reading is completed, the process ends in step S9. However, since it has not ended at this point, the process proceeds to step S3, and the next data is read from the buffer memory 3. Read in (3).

In step S4,6, the magnitude relationship of d (1), d (2), d (3) is examined. For example, d (1) <d (2) and d (2)>
When d (2)> 0 in d (3), d (2) is a positive peak value, and d (1)> d (2) and d (2) <d (3).
Then, when d (2) <0, d (2) has a negative peak value, and therefore, when the above conditions are satisfied, step S5,7
Then, d (2) is stored in d + and d-, and the order of the data is stored, and the process proceeds to step S8.

If the above conditions are not satisfied, proceed directly to step S8. At step S8, the current d (2) is stored in d (1), similarly d (3) is stored in d (2), and
The process moves to S2, it is judged whether or not the process is finished, and when it is not finished, the data is newly read from the buffer memory 3, the data is stored in d (3), the same process is repeated, and all the data are processed. Is done.

The number of data is the value obtained by multiplying the measurement time by 12000.

As described above, when obtaining the peak value, the following is done.

[Explanation for Obtaining Positive Crest Value] That is, when d (1) ≦ d (2), further d
(2) and d (3) are compared. d (3) <d (2)
Then, the value of 2 is at the peak, so the value of d (2) is the peak value. Furthermore, the sign of d (2) may be checked to see if this is a positive peak value. If d (2)> 0, d (2) is substituted for d +, the values of d + and n are stored, and the processing from step S8 onward is performed.

In other cases, for example, d (3) ≧ d (2), d
(2) If ≤0, etc., the process directly following step S8 is performed.

[Explanation for Obtaining Negative Crest Value] That is, when d (1) ≧ d (2), further d
(2) and d (3) are compared.

If d (3)> d (2), the value of 2 is at the peak, so the value of d (2) is the peak value. Furthermore, the sign of d (2) may be checked to see if this is a negative peak value. If d (2) <0, substitute d (2) for d-
The values of − and n are saved, and the processing from step S8 onward is performed.

In other cases, for example, d (3) ≦ d (2),
If d (2) ≧ 0 or the like, the processing directly after step S8 is performed.

In FIG. 2, examples of the positions of the positive crest value d + and the negative crest value d− obtained by the crest value detection circuit 4 are represented by symbols ▽ and Δ, respectively.

The crest value change calculation circuit 5 of this embodiment calculates the following characteristic amount from the output of the crest value detection circuit 4 described above.
Further, in the following formulas, d + (n) and d- (n) represent a set of time information n and peak value information d + and d-.

Ratio of sum of positive peak value and negative peak value within a fixed time; p1: Σ {d + (n); n ≦ T} / Σ {d− (n); n ≦
T} Ratio of adjacent crest values of the same sign and the distance between them; p2: d + (n-1) / d + (n) p2 (n, t): {time of n to be d + (n)}-{d +
N-1 time of (n-1)} p3: | d- (n-1) | / | d- (n) | p3 (n, t): {n time of d + (n)} -{D-
(N-1 time as (n-1)} The ratio of the crest values of the adjacent different codes and the distance between them; p4 (n, +): d + (n-1) / | d- (n) | p4 ( n, t): {time of n that becomes d + (n)}-{d-
N-1 time of (n-1)} p5 (n,-): | d- (n-1) | / d + (n) p5 (n, t): {d + (n) of n Time}-{d-
(N-1 time of (n-1)} In the discrimination circuit 6, the magnitude of the crest value and the feature amount obtained by the crest value change calculation circuit 5 are combined to convert the input voice into a voiced sound. Identification of unvoiced sounds, further identification of unvoiced consonants and unvoiced consonants, unvoiced recognition of vowels and consonants.

Further, although the sampling is performed at 12 kHz in this embodiment, the standard pattern stored in the standard pattern storage unit 7b is selected based on the following criteria.

1) Discrimination between voiced sound and unvoiced sound The difference between the values of d + (n) and d- (n) is 100 or more, and p
If 4 (n, +)> 1.3 or p5 (n,-)> 0.76 is satisfied, it is judged as voiced sound. If not, it is determined as unvoiced sound.

2) Discrimination between silent and unvoiced consonants p2 (n, t) <
When 3 or p3 (n, t) <3 is satisfied, it is determined that there is no sound. If not, it is determined to be unvoiced consonant.

3) Discrimination between vowels and consonants In contrast to the voiced sound determined in 1), if p1> 1.5, it is determined as a vowel. If not, it is determined to be a consonant.

Standard pattern candidates selected based on such a criterion are stored in the standard pattern storage unit 7b. 1 out of this standard pattern
Select one to select. This selection process is selected by the switch 10 in FIG. 1, and the standard patterns are sequentially read from the selected standard pattern storage unit 7b, output from the pattern and the feature extraction unit, and temporarily stored in the buffer memory 9. The characteristic pattern of the input voice that has been input is sent to the pattern matching unit 11. The pattern matching unit 11 calculates the similarity between the two, and the recognition result output unit 12 outputs the standard pattern with the maximum similarity as the recognition result from the recognition result output unit 12.

As described above, according to the present embodiment, a more accurate recognition result is output by selecting some input voice information candidates.

In the present embodiment, the group of standard patterns is selected based on the information on the time change of the crest value.However, the standard pattern storage unit is created by newly incorporating the time change information of the crest value in the feature amount of each standard pattern. Even if not divided into groups, the same effect as that of the embodiment can be obtained. As the time change information, there are other spectrum outlines, zero crossings, and the like. Further, by applying this embodiment, it is possible to configure a high-speed and highly-reliable voice identification device, for example, a typewriter equipped with voice recognition.

Further, in the present embodiment, the sampling is performed at 12 kHz, but it is not specified to this. One sampling is 12 bits, but this is not specified to 12 bits either.

[Effect] As described above, according to the present invention, the recognition rate of the voice recognition process is improved by preselecting the recognition result of the voice information according to the relative relationship of the plurality of peak levels included in the input voice information. It is possible to provide a voice recognition method that enhances the processing time and shortens the processing time.

[Brief description of drawings]

FIG. 1 is a block diagram of the voice recognition apparatus of this embodiment, FIG. 2 is a graphic representation of the output data of the input voice after A / D conversion, and FIG. 3 is the peak value detection of the embodiment. It is a flow chart showing processing. In the figure, 1 ... Microphone, 2 ... A / D converter, 3,9 ...
Buffer memory, 4 ... Crest value detection circuit, 4a ... CPU, 4
b …… ROM, 4c …… RAM, 5 …… Crest value change calculation circuit, 6
...... Discrimination circuit, 7a ...... Memory, 7b ...... Standard pattern storage section, 8 ...... Feature extraction section, 10 ...... Switch, 11 ...... Pattern matching section, 12 ...... Recognition result output section.

Claims (1)

[Claims] 1. A voice recognition method for recognizing input voice information, wherein a plurality of peak levels included in the input voice information are detected, and a relative relationship between the plurality of detected peak levels is calculated, A voice recognition method, characterized in that a recognition candidate of the inputted voice information is preselected from the calculated relative relation of the peak levels.