JPH04295895A - voice recognition device - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition device that derives a recognition result by pattern matching between a characteristic pattern of an input voice and a characteristic pattern of a voice to be recognized that has been registered in advance.

0002

2. Description of the Related Art In conventional word speech recognition devices, the degree of similarity between two speech patterns is generally determined by time-normalized matching of spectral parameter time series. However, since this method does not take energy information into consideration, the degree of similarity may be high even between patterns with different energy shapes. In order to solve this problem, a method has been proposed that uses a distance measure that takes into account not only the spectral distance but also the energy distance, and aims to improve the recognition rate by reflecting the difference in energy shape in the distance (for example, 2-41760).

0003

[Problem to be Solved by the Invention] However, in the above speech recognition device, although it is possible to control the range of the time-normalized matching path to some extent by using the energy distance, it is difficult to accurately follow the energy change state. There is no guarantee that pattern matching will be performed. Therefore, the degree of similarity between patterns with different energy shapes may become high, resulting in erroneous recognition.

[0004] Furthermore, in recognizing words with similar phonemes, it is important to accurately compare the differences between the steady and transient parts of speech. It can be considered that the steady part and transient part of voice almost correspond to the steady part and transient part of energy change. Since the stationary part occupies a large proportion of the voice, it is possible to distinguish the difference using conventional methods, but the transient part occupies a small proportion of the voice, so it is necessary to compare the two patterns through accurate matching. For example, the person's name "SATO""KATO"
In the case of , the difference between the two lies mostly in the initial consonants "S" and "K" that correspond to the transitional part of the voice, and unless these can be matched accurately, no improvement in recognition performance can be expected. However, with conventional speech recognition devices, it is difficult to accurately match the steady and transient parts, and differences in the transient part are difficult to reflect in the final similarity, which can cause erroneous recognition, especially between similar words. It had become.

The present invention solves the above-mentioned conventional problems, and provides a speech recognition device that can significantly reduce misrecognition between words with different energy shapes and between words with similar phonemes through simple control. The purpose is to provide the following.

[0006]

[Means for Solving the Problems] In order to solve the above problems, the speech recognition device of the present invention includes a speech analysis section that extracts a time series of speech spectral parameters and normalized logarithmic energy parameters, and an energy adjacent frame. An energy change degree determination unit that determines the energy change degree according to the transition pattern of energy change between, i.e., rise, fall, and no change, and a matching rule that is arbitrarily determined based on the energy change degree of the input pattern and the standard pattern. The system also includes a pattern matching section that limits routes.

[0007] Furthermore, the speech recognition device of the present invention causes the pattern matching section to determine a matching range according to a rule for association with an arbitrarily determined standard pattern according to the degree of energy change between adjacent frames of the input speech pattern. This configuration is configured to perform time-normalized matching between two patterns while increasing the weight of the matching path in a portion where the degree of energy change between adjacent frames is large.

[0008]

[Operation] With the above-described configuration, the present invention can perform matching according to the state of energy change in the speech pattern, so that the steady and transient parts of the speech can be accurately matched, and words with different energy shapes can be matched. It is possible to provide a speech recognition device that can significantly reduce misrecognition.

[0009] Furthermore, the pattern matching section is configured to perform
The matching range is limited according to rules for association with an arbitrarily defined standard pattern, and time-normalized matching between two patterns is performed while increasing the weight of the matching path in areas where the degree of energy change between adjacent frames is large. By configuring this to A speech recognition device that can suppress recognition can be provided.

0010

DESCRIPTION OF THE PREFERRED EMBODIMENTS A speech recognition apparatus according to an embodiment of the present invention will be described below with reference to the drawings. (FIG. 1) is a block diagram of a speech recognition device according to an embodiment of the present invention. In FIG. 1, reference numeral 1 denotes a speech analysis section that extracts feature parameters of speech; a spectral parameter extraction section 11 that extracts a time series of feature parameters of a speech spectrum at regular intervals from an input speech signal; and an energy parameter extraction unit 12 that extracts a normalized value of the logarithmic energy value of . 2 is an energy change degree determination unit that calculates the energy change degree; an energy difference sign calculation unit 21 that determines the state of energy change from the energy difference value; It is composed of an energy change degree calculation section 22 that calculates the degree of change. Reference numeral 3 denotes an input pattern memory that stores a time series of characteristic parameters of input speech, and 4 denotes a standard pattern memory that stores a time series of characteristic parameters of a voice to be recognized. Reference numeral 5 denotes a pattern matching unit that calculates the degree of similarity between the input pattern and the standard pattern, and performs matching according to the restrictions in the table 51 and the table 31 that describes the restrictions on the correspondence between the input pattern and the standard pattern according to the energy change state. The control unit 52 includes a control unit 52 that executes pattern matching while limiting the range of , and a weighting unit 53 that increases the weight of the matching path in areas where energy changes are large. Reference numeral 6 denotes a similarity comparison unit that selects a standard pattern having the maximum similarity from the similarity between the input pattern and each standard pattern.

Next, the operation of the speech recognition apparatus in this embodiment will be explained in detail using FIG. 1. The audio signal a inputted through a microphone or the like is converted into feature parameters such as linear prediction coefficients at regular time intervals by a spectral parameter extraction unit 11, for example, by linear prediction analysis. The audio signal a is also sent to the energy parameter extractor 12, where logarithmic energy values are calculated at fixed time intervals. This logarithmic energy value is further converted into a value normalized between the maximum and minimum energy values in the voice section. This normalization of energy is performed in order to be able to handle values such as energy changes in a unified manner, since the magnitude of energy varies depending on each utterance or word. The time series of the spectrum parameters and energy parameters calculated by the spectrum parameter extraction unit 11 and the energy parameter extraction unit 12 are stored in the input pattern memory 3.

Using the time series of normalized logarithmic energy values calculated by the energy parameter extraction unit 12, the energy difference sign calculation unit 21 determines the energy change pattern, that is, rising, falling, or no change, for each frame. . Here, let i be an arbitrary frame of the input audio pattern,
Let Ei be the normalized logarithmic energy value of frame i.

First, frame i and the previous frame (i-1
) is calculated using the following equation.

[0014]

[Math 1]

[0015] The sign of the difference value ΔEi obtained by (Equation 1),
In other words, find + (plus) and - (minus), and set + to the value 1.
, give the value -1 to -. Note that if the absolute value of the difference value is smaller than an arbitrarily determined constant a, that is, if the condition of the following equation is satisfied, the change in energy is considered to be small and a value of 0 is given.

[0016]

[Math 2]

Let Fi be the change pattern (-1, 0, 1) of the energy difference of any frame i obtained here. Using this time series of Fi, the energy change degree calculation unit 22 calculates the energy change degree Si in each frame using the following equation.

[0018]

[Math 3]

This value is defined as a quantity representing the tendency of energy change between an arbitrary frame i and one frame before and after it. The time series of the degree of change Si calculated in this way is stored in the input pattern memory 3.

It is assumed that the spectral parameters of each standard pattern to be recognized and the time series of the energy change degree S are stored in the standard pattern memory 4 in advance.

Next, the operation of the pattern matching section 5 will be explained. Here, it is assumed that an arbitrary frame of the standard pattern to be processed is j, and the degree of energy change in frame j is Sj. Pattern matching between the input pattern and the standard pattern is performed by time-normalized matching (DP) using dynamic programming according to slope constraints, as shown in (Figure 2), for example
It's called matching. ) is executed. In FIG. 2, the numerical values on the arrows represent the weights when the matching path moves in each arrow direction. The recurrence formula is as follows.

[0022]

[Math 4]

Here, dij is the vector distance of the spectral feature patterns of the input pattern i frame and the standard pattern j frame, and gij is the cumulative distance thereof.

It is now assumed that matching calculations are being performed using i-frames of the input pattern and j-frames of the standard pattern. First, the control unit 52 calculates, for example, (Table 1) the value of the energy change degree Si of frame i of the input pattern.
) table 51 that defines the restrictions on mapping as shown in
According to (Equation 4), it is determined whether to calculate the recurrence formula (Equation 4).

[0025]

[Table 1]

For example, when Si is 2, according to Table 1, matching calculation (Equation 4) is executed when Sj is 1 or 2, and is not executed in other cases. In this way, pattern matching calculations are performed for each standard pattern according to the restrictions in the table 51, and the degree of similarity is calculated. An example of the results of a matching calculation performed by the pattern matching section 5 between two patterns having similar energy shapes is shown in FIG. 3 (FIG. 3). The shaded area in FIG. 3 represents the range in which matching calculations can be performed. (Figure 3) As is clearer,
It can be seen that the corresponding transient parts and steady parts of the energy changes of the input pattern and the standard pattern are optimally matched. As a result, optimal matching between the transient and steady parts of the audio is performed.

Furthermore, when the pattern matching section 5 executes pattern matching between two patterns with different energy shapes, the matching path cannot reach the end point,
At this stage, this standard pattern can be excluded from recognition candidate speech.

By the way, between words with similar phonemes, the spectra often differ only in the transient part of the speech. If this difference in the transition portion is reflected in the degree of similarity obtained as a result of pattern matching, it becomes easier to distinguish between the two.

Therefore, the weighting unit 53 calculates (
The weight of the slope restriction shown in FIG. 2) is changed so that it is weighted more heavily in the transition part. As an example of weighting, a method as shown in FIG. 4 can be considered. This is an example in which twice the weight is applied when the energy change degree S is 2, that is, when the energy change is definitely on an upward trend (FIG. 2). In (Fig. 4), the numerical values on the arrows represent the weights when the matching path advances in each arrow direction, (a) when Si = 2 and Sj = 2, (b) when S
When i=2 and Sj≠2, (c) is Si≠2 and Sj
(d) shows the case when Si≠2 and Sj≠2. In this way, while changing the weight of the slope restriction according to the value of S, the controller 53 controls D according to the restrictions in the table 51.
Perform P matching calculation. As a result, the difference in the energy transition portion is more greatly reflected in the overall similarity.

Finally, the similarity comparing section 6 outputs the standard pattern that provides the maximum similarity among the obtained similarities with each standard pattern as a recognition candidate.

As described above, according to this embodiment, the energy change degree determination unit 2 numerically determines the degree of energy change from the tendency of energy change between frames before and after each frame, and uses this value to determine the degree of energy change in advance. table 5
By performing time-normalized matching according to the matching rule in 1, it is possible to achieve optimal matching between the transient and steady parts of energy changes.
Misrecognition between words with different energy shapes can be reduced.

Furthermore, by providing the pattern matching unit 5 with a weighting unit 53 that emphasizes the transitional part of energy change, differences in the transitional part of speech are greatly reflected in the overall similarity, and errors between words with similar phonemes are reduced. Recognition can be minimized.

[0033]

[Effects of the Invention] As described above, according to the present invention, the degree of energy change, such as rising or falling, is determined numerically from the energy change tendency before and after each frame of the audio pattern, and the degree of energy change, such as rising or falling, is determined in advance according to the degree of change. By configuring pattern matching to be controlled according to a set limit so that parts with similar change trends are matched, accurate matching of transient and steady parts of energy changes is possible, and recognition performance is improved. can be significantly improved.

[0034] Furthermore, by weighting parts with large energy changes during pattern matching, differences in transient parts of speech are emphasized, and speech recognition can particularly reduce misrecognition between words with similar phonemes. equipment can be provided.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a speech recognition device according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram showing slope restrictions in pattern matching.

FIG. 3 is an explanatory diagram showing an example of the result of pattern matching in this embodiment.

FIG. 4 is an explanatory diagram showing an example of weighted slope restriction.

[Explanation of symbols]

1 Speech analysis section 2 Energy change degree determination section 3 Input pattern memory 4 Standard pattern memory 5 Pattern matching section 6 Similarity comparison section 11 Spectral parameter extraction section 12 Energy parameter extraction section 21 Energy difference code calculation unit 22 Energy change degree calculation section 51 Table 52 Control section 53 Weighting section

Claims (4)

[Claims] 1. A voice analysis unit that extracts a voice feature spectrum and a time series of normalized logarithmic energy values from a voice signal; and a voice analysis unit that extracts a time series of a voice feature spectrum and a normalized logarithmic energy value; an energy change degree determination unit that determines the degree of energy change between adjacent frames; and an energy change degree determination unit that determines the degree of energy change between adjacent frames; A speech recognition device comprising: a pattern matching unit that executes time-normalized matching between two patterns while limiting the range of matching according to rules for association with standard patterns. [Claim 2] The energy change degree determining unit determines, for each frame, an increase from an energy difference value with an adjacent frame.
2. The speech recognition apparatus according to claim 1, wherein three energy change patterns are determined, ie, a downward trend and no change, and the degree of energy change is determined numerically from a transition state of the change pattern. [Claim 3] The pattern matching section is configured to perform a pattern matching process according to the degree of energy change between adjacent frames of the input audio pattern.
The matching range is limited according to rules for association with an arbitrarily defined standard pattern, and time-normalized matching between two patterns is performed while increasing the weight of the matching path in areas where the degree of energy change between adjacent frames is large. The speech recognition device according to claim 1, characterized in that: 4. The speech recognition apparatus according to claim 2, wherein the energy change degree determination unit determines that there is no energy change when the absolute value of the energy difference value with an adjacent frame is smaller than a certain value.