JPH0115079B2 - - Google Patents

Description

[Detailed Description of the Invention] A. Technical Field of the Invention The present invention relates to a word speech recognition method, in particular, to a word speech recognition method, which is based on input feature parameter time series for unknown input word speech.
In order to efficiently recognize various word lengths, multiple input reduced parameter time series are generated by dividing into multiple intervals, and a registered reduced parameter time series with a time series length corresponding to the registered word length and the above-mentioned reduced parameter time series are generated. The present invention relates to a word speech recognition method in which a candidate word to be recognized is selected by sequentially comparing a plurality of input reduction parameter time series with those having the same time series length.

B. Prior Art and Problems In a speech recognition system, feature parameters representing the features of each phoneme are extracted using the frequency analysis results of a speech signal, and pre-registered words corresponding to the extracted feature parameters and registered words are extracted. The unknown input speech is recognized by comparing it with the feature parameters that are available. That is, for example, the first formant frequency and the second formant frequency are sampled and used as the characteristic parameters. However, the amount of data to be processed for the above matching increases, and as the number of recognized categories increases, the time required for the above matching process increases.

Therefore, Japanese Patent Application No. 52-43972 and Patent Application No. 53966
As seen in Japanese Patent Application No. 53-53967, Japanese Patent Application No. 53965-1983, various methods have been proposed for increasing the recognition rate with a smaller number of features to be compared.

In particular, the present inventors have proposed, in Japanese Patent Application No. 55-62059, a system for efficiently determining recognition target word candidates using a relatively simple algorithm. This method divides the input feature parameter time series for input speech into decimal intervals, extracts a reduced feature parameter time series consisting of the average value obtained by averaging the parameter values for each interval, and extracts the reduced feature parameter time series. In this method, a candidate word to be recognized is selected by , and a comparison is made against the candidate word, thereby greatly improving the processing speed.
According to this method, when performing calculations to determine matching candidates, it is sufficient to perform mutual calculations on a predetermined small number of average value parameters, such as five, and the calculation time is significantly reduced.

In general, words to be recognized include a mixture of long words and short words. The conventional method described above, in which the number of parameters to be matched is determined in advance regardless of the length of the word, is considered to have one advantage in terms of processing simplicity, but it is difficult to recognize the accuracy of any word. Considering that the number of parameters to be compared, that is, the registered parameter time series length, can be changed depending on the word length, more efficient recognition will be possible.

C. Object and Structure of the Invention The present invention focuses on the above-mentioned points, improves and launches the conventional method described above, and makes it possible to improve processing speed and reduce the memory amount of registered dictionaries without reducing the recognition rate. It is intended to be. Therefore, the present invention assumes a plurality of word lengths for the input speech, obtains input reduction parameter time series of several lengths, and performs verification using the parameter time series according to the registered word length. It was designed to do so. That is, the word speech recognition method of the present invention analyzes the speech signal of unknown input word speech, and based on the time series of input feature parameters extracted from the speech signal, parameters in each section divided into predetermined sections are determined. Based on the value, the input reduction parameter time series is configured to generate an input reduction parameter time series whose time series length is smaller than the input feature parameter time series, and the input reduction parameter time series is compared with a registered reduction parameter time series registered in advance. In a word speech recognition method for recognizing unknown input word speech, a plurality of number of sections are predetermined for the section division from the beginning to the end of the unknown input word speech, and each section division with a different number of sections is A registered reduced parameter time series that is configured to generate multiple types of input reduced parameter time series with different corresponding parameter time series lengths from the input feature parameter time series, and has a parameter time series length determined in advance for each registered word. and an input reduction parameter time series that has the same parameter time series length as the registered word to be compared among the plurality of types of input reduction parameter time series, to determine recognition target candidate words and identify unknown input words. It is characterized by recognition of word sounds. This will be explained below with reference to the drawings.

D. Embodiments of the Invention Fig. 1 is an explanatory diagram for explaining the relationship between registered words and parameter time series lengths, and Fig. 2 is an explanatory diagram for explaining a processing example for generating a reduced parameter time series in the present invention. , FIG. 3 is an explanatory diagram for explaining another processing example for generating a reduced parameter time series in the present invention, FIG. 4 is an example configuration of the present invention, and FIG. 5 corresponds to the above-mentioned FIG. 3. An explanatory diagram in the form of a flowchart of a process for determining a section in a process example is shown.

For example, if the registered word is a place name, the first
As shown in the figure, registered words include relatively long words such as "Hokkaido", short words such as "tsu", and average words such as "Tokyo".
There are various types. Generally, longer words are considered to contain more feature changes than shorter words. In the present invention, a smaller number of input reduced parameter time series are generated from the input feature parameter time series for input speech, and they can still be compared with the registered reduced parameter time series that have been extracted and registered using the same method in advance. However, it is better to match long words that include many feature changes with a long parameter time series length, and match short words with a small parameter time series length.
This can be said to be desirable in terms of the balance between recognition accuracy and processing speed. Therefore, as shown in Figure 1, depending on the registered word length, for example, for "Hokkaido", the registered parameter time series length is "5", and for "Tsu".
For example, the parameter time series length is determined in advance, such as "3", and the registered reduced parameter time series corresponding to the time series length is registered.

However, it is unclear what time series length should be used for the input reduction parameter time series obtained from unknown input word speech, and it is necessary to use the same time series length for matching. In the present invention, as will be described in detail later, input reduction parameter time series of all predetermined time series lengths are prepared.

Generation of a reduced parameter time series when a parameter time series length is given is performed, for example, as follows.

As shown in Figure 2, between time T ₀ and T _E ,
Assuming that there is a sampled feature parameter P, for example, the parameter time series length Nf
If "5" is required as
Divide the time from T ₀ to T _E into five equal parts and set the time T _E /
5. Determine 2T _E /5, 3T _E /5, 4T _E /5, T _E. Then, the feature parameter values from time T ₀ to T _E /5 are averaged, and the values from time T E /5 to T _E /5 are averaged.
Average each feature parameter value up to 2T _E /5, ... average each feature parameter value between 4T _E /5 and T _E , and create a reduced parameter time series consisting of the five average value parameters. Make it extract.

In addition, instead of dividing the time equally, as shown in Figure 3, by dividing the cumulative variation of the characteristic parameters equally, a reduced parameter time series with a predetermined parameter time series length is generated. It's okay.

In the case of the example shown in Fig. 3, the sections are divided at equal intervals on the time axis in the example shown in Fig. 2, whereas the intervals at points where the rate of change of the characteristic parameter is relatively large are selected to be small. I have to. That is, when the characteristic parameter P is assumed to be as shown in FIG. 2, the cumulative amount of variation of this parameter P, that is, the cumulative amount of variation, is plotted with time as the horizontal axis as shown in FIG. For the figure drawn in this way, the value 1/5TAV is obtained by dividing the maximum value of cumulative variation TAV into 5 equal parts,
Select 2/5TAV, 3/5TAV, 4/5TAV, TAV.
And the cumulative fluctuation amount is the above value 1/5TAV, 2/5TAV,
Extract the time points T ₁ , T ₂ , ...T _E , and average the values of each feature parameter shown in Figure ₂ from _time T ₀ to T ₁ . The values of each characteristic parameter shown in Figure 2 are averaged, and..., at time T ₄
The values of each characteristic parameter shown in FIG. 2 between T _E and T E are averaged, and a reduced parameter time series consisting of five average value parameters is extracted. When obtaining a reduced parameter time series with a parameter time series length of, for example, "3", it is of course divided into three sections, and a reduced parameter time series consisting of the average value parameter of each section is extracted.

The average value parameter of the reduced parameter time series is calculated, for example, as shown in Japanese Patent Application No. 55-62059, but since it is well known, detailed explanation will be omitted.

Note that instead of averaging the feature parameter values, it may be simplified to extract a reduced parameter time series consisting of interval boundary values.

FIG. 4 shows the configuration of an embodiment of the present invention. In the figure, 1 is a group of band filters, 2 is a parameter extraction circuit, 3 is a parameter average interval determination circuit, 4 is a parameter average circuit, 5 is a switching circuit, 6 is a registered word reduction parameter time series registration unit, 7-1 to 7-n
denotes an input reduced parameter time series buffer for each parameter time series length N ₁ -N _o , 8 represents a reduced parameter time series collation unit, and 9 represents a candidate word determination unit.

The input audio signal is input to band filter group 1,
The parameter extraction circuit 2 extracts input feature parameters corresponding to the input audio signal. For example, if n types of parameter time series lengths are determined from N ₁ to _No , the parameter average interval determination circuit 3 determines the time point for each parameter time series length Nf in the example shown in FIG. After extracting T _E , the period from T ₀ to T _E is divided into Nf equal time points T _E /
Determine Nf, 2T _E /Nf, 3T _E /Nf, ..., T _E, respectively. The time points T ₁ , T _{2 .} . . shown in FIG. 3 will be described later with reference to FIG. 5. Once the sections are determined based on the above points, the parameter averaging circuit 4 calculates the average value of the parameter values for each section based on the input feature parameters.

n parameter averaging interval determining circuits 3 and n parameter averaging circuits 4 may be provided each corresponding to the parameter time series lengths N ₁ to _No , and processing may be performed in parallel, or one circuit may be configured to perform processing in parallel. By using it repeatedly, it may be possible to process it serially.

The switching circuit 5 converts each input reduction parameter time series buffer 7-1 in the registration mode and the recognition mode.
This circuit switches between 7-n and 7-n. In the registration mode, since the word length of the word to be registered is known, an appropriate parameter time series length Nf can be determined in advance for each word. Therefore, the parameter averaging circuit 4 extracts a reduced parameter time series with an appropriate parameter time series length Nf,
The registered word reduction parameter is registered in the time series registration section 6 together with the character code of the registered word.

In recognition mode, the word length of the input speech is unknown. Therefore, as described above, the reduced parameter time series for all parameter time series lengths N ₁ to _{N o} are extracted by the parameter averaging circuit 4, and the input reduced parameter time series are extracted via the switching circuit 5. The consequences are led to and stored in series buffers 7-1 to 7-n.

In the recognition mode, the reduction parameter time series collation unit 8 sequentially reads out the registered reduction parameter time series of registered words from the registered word reduction parameter time series registration unit 6, and stores n input reduction parameter time series buffers 7-1 to 7. -n, select one that matches the time series length of the registered reduction parameter time series,
Check against the input reduction parameter time series stored there.

Distance D between unknown input word and registered word ⁱ
(〓, ^〓i ) is D(〓, ^〓i )=1/Nd( ^〓N , ^〓i ) or D(〓, ^〓i )=C(N)・d( ^〓N , ^〓i )
is given by

Here, N is the parameter time series length of ⁱ , and ^N is a parameter whose time series length is N among the parameters extracted from the unknown input word speech. Also, in the above equation, d(〓 ^N , 〓 ⁱ ) is d(〓 ^N , 〓 ⁱ )= _N 〓 ^N=1 _J 〓 ^j=1 |x ^N _j (n)−S ⁱ _j (n)| Given. x ^N _j (n) and s ⁱ _j (n) correspond to the input reduction parameter time series and the registered reduction parameter time series, respectively.

Based on the distance D calculated by the reduced parameter time series matching unit 8, the candidate word determination unit 9 calculates
It is determined whether the registered word is suitable as a candidate word to be recognized. The candidate word names determined in this way are output from the candidate word determination section 9.

When determining the time points T ₁ , T ₂ , . . . shown in FIG. 3 above, the parameter average interval determination circuit 3 shown in FIG. . That is, when extracting a time series whose parameter time series length is Nf, the following processing is performed.

(1) Based on the parameters extracted by the parameter extraction circuit 2, the cumulative variation amount TAV as shown in FIG. 3 is extracted.

(2) And the value obtained by dividing the value of the cumulative fluctuation amount TAV into Nf equal parts
Determine DTAV.

(3) Then, in order to find the first time point _T1 , set J = 1, set the above value DTAV in register AVH, and set value T in timing start register TS(J).
Set (I).

(4) Thereafter, feature parameter values are accumulated sequentially until the cumulative value AV(I) of the feature parameters becomes equal to or larger than the contents of the register AVH.

(5) When the cumulative value AV(I) is equal to or greater than the contents of the register AVH, the timing value T(I) at that time
is set in the register TE(1 ₎ for time T1, the value T(I+1) is set in the register TS(J+1), and the value (AVH+DTAV) is set in the register AVH.
and set J=2 to find the next time _T2 .
shall be.

(6) Similarly, feature parameter values are accumulated until the cumulative value AV(I) becomes equal to or greater than the contents of the register AVH. That is, time points T ₂ , T ₃ , and T ₄ are found.

(7) Then, when the accumulation circuit I reaches the value N, that is, when the accumulation process reaches the accumulation of the characteristic parameters corresponding to the time T _E shown in FIG.
T _E is determined.

E. Effects of the Invention As explained above, according to the present invention, efficient word speech recognition becomes possible.

Generally, if you try to improve the recognition rate, you sacrifice the processing speed, and if you try to increase the processing speed, the recognition rate deteriorates, but according to the present invention, the most appropriate parameters for each registered word Since the time series length can be selected, processing speed can be improved compared to conventional methods without reducing the recognition rate. Furthermore, the memory capacity of the registered dictionary can also be reduced.

[Brief explanation of drawings]

FIG. 1 is an explanatory diagram for explaining the relationship between registered words and parameter time series lengths, FIG. 2 is an explanatory diagram for explaining a processing example for generating a reduced parameter time series in the present invention, and FIG. An explanatory diagram for explaining another processing example for generating a reduced parameter time series in the present invention, FIG. 4 shows the configuration of an embodiment of the present invention, and FIG. 5 shows interval determination in a processing example corresponding to the above-mentioned FIG. 3. An explanatory diagram in the form of a flowchart is shown for the process of performing. In the figure, 3 is a parameter average interval determination circuit, 4 is a parameter average circuit, 6 is a registered word reduction parameter time series registration unit, 7-1 to 7-n are input reduction parameter time series buffers, and 8 is a reduction parameter time series comparison. represents the department.

Claims (1)

[Claims] 1. Analyze the audio signal of unknown input word audio, and based on the input feature parameter time series extracted from the audio signal, based on the parameter values in each section divided into predetermined sections. , is configured to generate an input reduction parameter time series whose time series length is smaller than the input feature parameter time series, and compares the input reduction parameter time series with a registered reduction parameter time series registered in advance,
In a word speech recognition method for recognizing unknown input word speech, a plurality of division numbers for the section division from the start to the end of the unknown input word speech are predetermined, and parameters corresponding to each section division with a different number of sections are provided. The registered reduced parameter time series is configured to generate multiple types of input reduced parameter time series with different time series lengths from the input feature parameter time series, and has a parameter time series length determined in advance for each registered word; Candidate words to be recognized are determined by comparing registered words to be matched with input reduction parameter time series having the same parameter time series length among multiple types of input reduction parameter time series, and recognition target words are determined. A word speech recognition method characterized by performing recognition.