JPS6118200B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to improvements in speech recognition devices. Speech recognition devices, which are devices that recognize voices such as numbers and command words, are useful as a means of inputting data into computer systems or as a means of controlling various industrial machines, and the labor savings achieved by using them. The effects of increasing efficiency and efficiency are extremely high.

On the other hand, research on speech recognition has made significant progress in recent years, and although there are some moves towards commercialization, the recognition performance obtained is still not sufficient. As the operating principle of speech recognition devices, a pattern matching method is often adopted. in this way,
A standard pattern (hereinafter referred to as standard pattern) is set in advance for each word in the vocabulary to be recognized, and when unknown speech (hereinafter referred to as input pattern) is input, it is compared with each of the above patterns. Recognition is performed by determining the most similar standard pattern. This pattern matching method has a simple principle and can easily follow changes in vocabulary and changes in speaker by changing the standard pattern. It has the disadvantage of being unstable. One of the major causes of variation in speech patterns is variation in speaking speed.

This is true even when the same person pronounces the same word.
This is a phenomenon in which the speed of vocalization changes each time, resulting in a significant expansion or contraction on the time axis of the voice pattern. In normal speech, the variation in speech rate is thought to be about 30% or less. At this level, the
The pattern matching method based on dynamic programming (hereinafter referred to as DP matching method) described in ``Pattern Comparison Device'' and the sequential method described in ``Japanese Patent Application No. 4542/1983''. Therefore, it can be normalized. That is, as long as the speaker tries to make normal speech, accurate recognition can be performed using conventional techniques. However, in reality, high-speed utterances are often performed to improve the data input speed, in some cases up to 30% faster than the standard pattern.
It is also possible for the words to be uttered at speeds higher than that. When generated at such a high speed, even with the above-mentioned DP matching method, it is impossible to perform complete normalization, resulting in erroneous recognition.

The present invention improves the drawback that conventional speech recognition devices are prone to malfunctions due to wide variations in generation speed, and provides speech recognition that enables accurate data input at the speed desired by the speaker. The purpose is to realize and provide equipment.

The device according to the present invention includes a speech rate setting section for setting a speech rate desired by a speaker, an analysis section configured such that the analysis frame period changes based on a signal from the speech rate setting section, and a recognition section for executing a determination operation based on the analysis results obtained by the section.

The detailed configuration of the present invention will be explained below based on the drawings. FIG. 1 is a block diagram showing the basic configuration of the present invention. The input voice inputted from the microphone 10 is converted by the analysis unit 11 into a time series of characteristic vectors a _i =(a _1i , a _2i , . . . a _Ki ) (1).

A=a ₁ , a ₂ , ...a _i ...a _I (2) Here, a _i is time sampled at time interval Δt.

Specific analysis processing includes well-known frequency analysis,
Various methods such as autocorrelation analysis and linear prediction analysis are possible, but below we will consider frequency analysis as a typical example. In this case, the spectrum a _i in equation (1) is constructed with elements of values obtained by time-sampling the analysis outputs of the bandpass filter group of the K channel at the analysis frame period Δt. The speech rate setting section 12 has a mechanism such as a dial with a scale that allows the user to specify the speech rate by manual operation.The analysis frame period Δt is determined by the designation of the dial with a scale, and the analysis frame period Δt is determined by the signal d. The analyzer is configured to instruct the analysis unit. The recognition unit 13 performs a recognition operation by processing the utterance pattern A given by the analysis unit 11 as shown in equation (2). Various methods can be considered as the principle of recognition operation, but here we will consider a pattern matching method as the simplest example. That is, when the vocabulary to be recognized consists of N words where n = 1, 2, ..., N, the standard pattern B ⁿ = b ⁿ ₁ , b ⁿ _J2 , ... b ⁿ is used for each word n. Prepare and memorize _i ,...b ⁿ _J (3) in advance. It is assumed that the time axis j of this standard pattern is different from the time axis i of the previous input pattern A and is time-sampled at intervals of a fixed analysis frame period Δτ. When input pattern A is given, a comparison operation is performed between these standard patterns and similarity is calculated. Let us now denote the degree of similarity between the standard pattern B ⁿ and the input pattern A as S(A, B ⁿ ) (4). When this degree of similarity is calculated for all standard patterns from n=1 to N, the maximum value n=n is determined. Input pattern A is recognized as word n, and the recognition result is signal r
is output as

When operating according to the above principle, it is possible to correct the difference between the speed at which the standard pattern is uttered and the speed at which the input pattern is uttered by adjusting the analysis frame period Δt of the input pattern. can.

FIG. 2 is a diagram for explaining how the influence of a change in speaking speed is corrected. Reference numeral 21 indicates a standard pattern B ⁿ which is time-sampled at a constant frame period Δτ (for example, 20 ms). Reference numeral 22 is an example of an input pattern in which the same word as the standard pattern is uttered at twice the speed (1/2 time length). If time sampling is performed with the same analysis frame period as the analysis frame period Δr of the standard pattern, the pattern will be half the length of the standard pattern. If the analysis frame period of the input pattern is now halved, the resulting input pattern A will have the same length as the standard pattern B ⁿ , and the influence of the change in speech rate will be cancelled. Of course, it is impossible to artificially control the speech rate to accurately double it. For this reason, even if the analysis frame period of the input pattern is doubled, the length of input pattern A may differ to some extent from the length of the corresponding standard pattern. By adopting the method, this degree of difference in time length can be normalized and does not cause any problems. Even if high-precision time normalization such as the Yoshiba DP matching method is not adopted, after aligning the input pattern with the length of the shorter of the standard patterns and discarding the surplus parts of the other patterns, the simple Even if pattern matching is performed, if the analysis frame periodic control of this method is used,
Approximate time normalization effect can be obtained.

According to the configuration described above, the speaker can set his or her preferred (or required) speaking speed in advance, and by doing so, even if the voice is uttered at a significantly high speed (or slowly), it can be correctly uttered. Become recognized.

FIG. 3 is a block diagram showing an example of the structure of a mechanism for controlling the analysis frame period in the analysis section 11.

The input audio signal x is analyzed into N-channel analysis bands by a frequency analyzer 110 consisting of a group of bandpass filters and a level detector, and the signals a ₁ ,
It is output as a ₂ , ..., a _N. The speech rate setting section 12 is composed of a dial mechanism with a scale, and when the dial is set at scale 1, the number 10 is output as a signal d, and the dial scale is set to 1.
The signal d increases or decreases by 1 every time it is increased or decreased by 0.1.
increased or decreased. A reference clock generating circuit 116 generates a clock pulse Cl having a period of 200 μs. The adder circuit 113 calculates the arithmetic sum of the numerical value given by the signal d and the numerical value given by the signal d indicating the contents of the register 114. This arithmetic sum is stored in register 114 in synchronization with clock pulse Cl.
written to. Therefore, the contents of register 114 are equal to the period of clock pulse Cl, i.e. 200 μs.
It continues to be incremented by the value specified by the signal d.

The decoding circuit 115 detects the contents of the register 114.
Frame clock pulse fcl at the moment it exceeds 1000
occurs. The contents of register 114 are reset to 0 by this frame clock pulse fcl. Thereafter, addition is repeated every 200 μs. Therefore, when the value specified by the signal d is 10, a frame clock pulse fcl is generated with a period of (1000/10) x 200 μs = 20 ms.

When the value specified by the signal d is α times 10, the period of the frame pulse fcl is approximately 1/
It is clear that it will be α times as large.

The multiplexer 111 time-samples the outputs a ₁ -a _N of the frequency analyzer 110 in a scanning manner every time the frame clock pulse fcl is applied, and sends the sampled signals to the A/D converter 112. As the output signal a of the A/D converter, a digital signal group of analysis results a ₁ to a _N is obtained in synchronization with the frame clock pulse fcl.

As described above, a circuit has been realized in which the analysis frame period is set shorter as the speech rate is set faster.

In the above example, when the standard pattern B ⁿ is fixedly built into the recognition unit, the standard pattern may be time-sampled at a fixed analysis frame period Δτ (for example, 20 ms). However, when a speaker is replaced, a standard pattern is usually registered in the new speaker's own voice in order to follow the individuality of the new speaker.

The embodiment shown in FIG. 4 is used when registering a standard pattern,
FIG. 2 is a block diagram of a speech recognition device configured to easily and reliably control the analysis frame period in each of two modes when actually inputting data. In this figure, the analysis section is shown divided into an analysis processing section 40 and an analysis frame pulse generation section 41. Adder circuit 1 in the block diagram of Figure 3
13, a register 114, a decoding circuit 115, and a reference clock generation circuit 116 correspond to the analysis frame pulse generation section 41, and a frequency analysis section 110, a multiplexer 111, and an A/D converter 112 correspond to the analysis processing section 40. Further, the recognition section includes a standard pattern storage section 43, an input pattern buffer 44, a similarity calculation section 45, and a maximum value detection section 46. The speech rate setting section 12 is the same as that explained in FIG. The reference analysis frame period generator generates a pulse with a fixed period Δτ (for example, 20 ms). Changeover switches 47 and 48 operate in conjunction. That is, in the standard pattern registration mode, the changeover switches 47 and 48 are both connected to the contact point 1. Since the changeover switch 47 is connected to contact 1, the period of the analysis frame clock pulse fcl supplied to the analysis processing section 40 is Δτ (for example,
20ms).

Therefore, the standard pattern uttered at this time is time sampled at the analysis frame period Δτ. Standard pattern B ⁿ (n=1 to N) thus obtained
is sent to the standard pattern storage section 43 and stored therein by the action of the changeover switch 48. When in the mode for actually inputting data, selector switches 47 and 4
8 are both connected to contact 2. Therefore, the analysis frame pulse fcl supplied to the analysis processing section 40
is sent from the analysis frame pulse generator 41. Therefore, in the data input mode, the period of the analysis frame pulse fcl is controlled by the speech rate setting section 12. Therefore, the speaker can set the analysis frame period Δτ according to his or her desired speaking speed. The input pattern A time-sampled at an appropriate analysis frame period is sent to the input pattern buffer 44 by the action of the changeover switch 48 and subjected to recognition processing. Although the specific processing content for recognition is not directly related to the present invention, an example is as follows. Input pattern A stored in input pattern buffer 44 and standard pattern B ⁿ stored in standard pattern storage section 43
A comparison operation is performed by the similarity calculation unit 45 between (n=1 to N), and the similarity S (A, B ⁿ ), n=1
~N is calculated. The maximum value detection unit 46 determines the word n=n for which the degree of similarity S(A, B ⁿ ) is maximum, and
This is output as the recognition result γ.

According to the configuration of this embodiment, in the standard pattern registration mode, the analysis frame period is automatically set to a fixed reference analysis frame period Δτ (for example, 20 ms). If changeover switch 47
If the pulse from the analysis frame pulse generator 41 is sent directly to the analysis processing unit without the
In addition to operating the speech rate setting section 12, it is necessary to adjust the analysis frame period to the standard analysis frame period Δτ (for example, 20 ms). If you perform two types of operations in this way, you may forget to perform one operation, etc.
However, malfunctions are likely to occur. With a configuration such as the embodiment shown in FIG. 4, the above problems can be completely avoided.

Although the configuration of the present invention has been described above based on examples, these descriptions do not limit the scope of the present invention. In particular, a circuit that makes the analysis frame period variable can be realized using various circuits, and is not limited to the example shown in FIG. 3. The specific configuration of the recognition unit is not particularly limited in the present invention. In addition, although Fig. 3 shows an example in which voice analysis processing is performed by frequency analysis,
The principles of the present invention can also be applied to other analysis methods, such as autocorrelation analysis and formant analysis.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing the basic configuration of the present invention, in which 10 is a microphone, 11 is an analysis section, 12 is a speech rate setting section, and 13 is a recognition section. FIG. 2 is a diagram showing the principle of the present invention. FIG. 3 is a diagram showing an embodiment of the present invention, in which 110 is a frequency analysis section, 111 is a multiplexer, and 112 is a diagram showing an embodiment of the present invention.
113 is an A/D converter, 113 is an adder circuit, 114 is a register, 115 is a decoding circuit, and 116 is a reference clock generation circuit. FIG. 4 is a diagram showing a second embodiment of the present invention, in which 40 is an analysis processing section, 41 is an analysis frame clock generation section, 42 is a reference analysis frame clock generation section, 43 is a standard pattern storage section, and 44 is a diagram showing a second embodiment of the present invention. 45 is an input pattern buffer, 45 is a similarity calculation section, 46 is a maximum value detection section, 12 is a speech rate setting section, and 47 and 48 are mutually interlocked changeover switches.

Claims (1)

[Claims] 1. In a speech recognition device consisting of an analysis section that analyzes an input speech signal and a recognition section that makes a judgment by processing the analysis result signal, there is a method for setting the speech rate desired by the speaker. 1. A speech recognition device comprising a speech rate setting section, the analysis frame period of the analysis section being controllable by a signal from the speech rate setting section.