JP2864821B2 - Voice recognition device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a feature pattern of an input voice.
The present invention relates to a speech recognition device that derives a recognition result by time-normalized matching between a pattern and a feature pattern of a speech to be recognized which is registered in advance.

[0002]

2. Description of the Related Art In a conventional word speech recognition apparatus, a similarity between two speech patterns is generally obtained by time-normalized matching between spectral parameters in a time series. However, since this method does not take energy information into consideration, similarity may be high even between patterns having different energy shapes. In order to solve this, a method has been proposed in which a distance scale that takes into account not only the vector distance but also the energy distance is used, and the difference in energy shape is reflected in the distance to improve the recognition rate (for example, a method has been proposed). JP-B-2-41760).

[0003]

However, in the above-described speech recognition apparatus, it is possible to control the range of the time-normalized matching path to some extent by using the energy distance, It is not guaranteed that exact pattern matching will be performed. For this reason, the similarity increases between patterns having different energy shapes, and erroneous recognition may occur.

In recognition between words having similar phonemes, it is important to accurately compare the difference between the stationary part and the transient part of the speech. It can be considered that the stationary part and the transient part of the voice almost coincide with the stationary part and the transient part of the energy change. Since the stationary part has a large proportion in the speech, the difference can be distinguished by the conventional method. However, the transient part has a small proportion in the speech, so that two patterns must be compared by accurate matching. . For example, personal names "SATO" and "KATO"
In the case of, the difference between the two is the part of the consonants "S" and "K" at the beginning of the word, which substantially corresponds to the transient part of the voice. If these cannot be matched correctly, improvement in recognition performance cannot be expected. However, in the conventional speech recognition device, it is difficult to perform accurate matching between the stationary part and the transient part, and it is difficult for the difference between the transient parts to be reflected in the final similarity. Had become.

[0005] Further, in the case of a speech accompanied by a syllable having a small energy at the beginning and end of the speech, if the speech section detection is erroneous, the energy pattern of the registered speech differs from that of the registered speech, which causes erroneous recognition.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned conventional problems, and erroneous recognition between words having different energy shapes and words having similar phonemes can be greatly reduced by simple control. It is an object of the present invention to provide a speech recognition device capable of accurately recognizing a speech in which the beginning and ending are likely to be lost.

[0007]

In order to solve the above-mentioned problems, a speech recognition apparatus according to the present invention provides a speech spectral parameter.
A voice analysis unit for extracting a time series of a logarithmic energy parameter and a normalized logarithmic energy parameter, a smoothing unit for smoothing a minute variation of the normalized energy sequence, and a unit other than the beginning and end of the voice.
Is the energy in response to the energy change between adjacent frames, i.e., a transition pattern with no rise, no change,
Energy to determine the degree of change - the change degree determination unit, the audio of the start
Always set the end energy change degree to no change
The matching path is restricted by the initial setting unit and the rules of association arbitrarily determined by the energy change degree of the input pattern and the standard pattern, and the path is weighted according to the value of the energy change degree. And a pattern matching unit for executing time-normalized matching.

Further, the speech recognition apparatus of the present invention includes an initial setting unit for setting the degree of energy change at the beginning and end of the speech as if there is no energy change.

[0009]

According to the present invention, with the above-described configuration, a rough form of an energy change is obtained by smoothing processing for absorbing a minute change in an energy series, and a time regularization matching the state of the energy change of the voice pattern is performed. Can perform morphological matching,
The stationary part of the voice without being affected by the minute change of energy,
Transient parts can be accurately associated with each other,
It is possible to provide a speech recognition device that can significantly reduce erroneous recognition between words having different shapes and between similar words.

In addition, the present invention performs time-normalized matching between two patterns assuming that there is no energy change at the beginning and end of speech. It is possible to provide a speech recognition device capable of accurately recognizing speech even when a small mountain exists and speech section detection is apt to be erroneous.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A speech recognition apparatus according to one embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a speech recognition apparatus according to an embodiment of the present invention. In FIG. 1, reference numeral 1 denotes a speech analysis unit for extracting speech feature parameters, and a spectrum parameter extraction for extracting a time series of speech spectrum feature parameters at regular intervals from an input speech signal. It comprises a unit 11 and an energy parameter extracting unit 12 for extracting a normalized value of the logarithmic energy value of the audio signal. Reference numeral 2 denotes a smoothing unit for smoothing a small change in energy. Numeral 3 denotes an energy change degree determining unit which determines an energy change state from the energy difference value, and an energy change code calculating unit 31 which determines the energy change from the transition state of the energy difference code value at each time. It comprises an energy-degree-of-change calculator 32 for calculating the degree. Reference numeral 4 denotes an initial setting unit for setting the degree of energy change at the beginning and end of the voice as no energy change. Reference numeral 5 denotes an input pattern memory for storing a time series of characteristic parameters of an input voice, and reference numeral 6 denotes a standard pattern memory for storing a time series of characteristic parameters of a voice to be recognized. Reference numeral 7 denotes a pattern matching unit for calculating the similarity between the input pattern and the standard pattern, which limits the correspondence between the input pattern and the standard pattern according to the energy change state. The control unit 72 includes a weighting unit 71 and a time-normalized matching that is weighted according to the degree of change in energy while limiting the range of matching according to the restrictions of the association unit 71. Reference numeral 8 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 the above embodiment will be described in detail with reference to FIG. The audio signal input through a microphone or the like is converted into spectral parameters such as linear prediction coefficients at regular time intervals by, for example, linear prediction analysis or the like in the spectral parameter extraction unit 11. The voice signal is also sent to an energy parameter extraction unit 12, where a logarithmic energy value is calculated at regular intervals. This logarithmic energy value is further converted into a value normalized between the maximum value and the minimum value of the energy value in the voice section. This energy normalization is performed in order to absorb variations in the magnitude of energy due to speech or words. This normalized logarithmic energy-value series is
In the subsequent processing, the smoothing processing is performed by the smoothing unit 2 in order to reduce an adverse effect due to a small change in energy. As the smoothing processing, for example, a moving average value between one frame before and after each frame as shown in (Equation 1) is calculated.

[0013]

(Equation 1)

Here, Pi represents the normalized logarithmic energy value in frame i, and Ei represents the energy value after smoothing.
By this smoothing process, small irregularities in the energy sequence are absorbed, and a rough outline of the energy of the sound can be obtained. Note that several frames before and after
The same effect can be obtained by performing a so-called median smoothing process for finding the median value of the energy value between programs. Spectral parameters and energy parameters calculated by the spectral parameter extracting unit 11 and the smoothing unit 2
The time series of the data is stored in the input pattern memory 5.

Using the time series of the energy values calculated by the smoothing unit 2, the energy-difference code calculation unit 31 performs an energy change pattern for each frame, ie, no rise, fall, or change. To determine. Here, an arbitrary frame of the input voice pattern is defined as i (i = 1... I), and the frame i
Let Ei be the energy value after smoothing.

First, the frame i and the preceding frame (i-
The difference value ΔEi of the energy value in 1) is obtained by the following equation.

[0017]

(Equation 2)

The sign of the difference value ΔEi obtained by (Equation 2)
That is, + (plus) and-(minus) are obtained, and the value 1 is added to +.
And − is given the value −1. If the absolute value of the difference value is smaller than an arbitrarily determined constant a, that is, if the following condition is satisfied, the value 0 is given assuming that the change in energy is small.

[0019]

(Equation 3)

The change pattern (-1, 0, 1) of the energy difference of an arbitrary frame i obtained here is defined as Fi. Using the time series of Fi, the energy change degree calculator 32 calculates the energy change degree Si in each frame by the following equation.

[0021]

(Equation 4)

This value is defined as an amount indicating the tendency of the change in energy between an arbitrary frame i and one frame before and after the arbitrary frame i. However, the energy change degree at the beginning and end of the voice, that is, the values of S1 and SI are forcibly set to 0 by the initial setting unit 4.
Is set to The time series of the energy change degree Si thus calculated is stored in the input pattern memory 5.

It should be noted that the time series of the spectral parameters of each standard pattern to be recognized and the energy change degree S are obtained in advance in the same manner as described above and stored in the standard pattern memory 6. I do.

Next, the operation of the pattern matching section 7 will be described. Here, an arbitrary frame of the standard pattern to be processed is j, and the degree of energy change in the frame j is Sj. The pattern matching between the input pattern and the standard pattern is performed by time-normalized matching (called DP matching) by dynamic programming according to the gradient restriction as shown by the recurrence formula of Equation (5). Is executed by

[0025]

(Equation 5)

Where dij is the input pattern i-frame,
The vector distance gij between the spectral feature patterns of the standard pattern j frame is the cumulative distance. Furthermore, in order to clarify the difference between the transient parts of the voice, the following equation is obtained by weighting the slope limit of (Equation 5) according to the magnitude of the energy change in each frame according to the magnitude thereof. (Refer to (FIG. 2)).

[0027]

(Equation 6)

(Equation 6) indicates that the matching path is an input pattern.
When moving in the normal direction, S i is used in the standard pattern direction, and when moving in the oblique direction, S i + S j is added to Equation 5 as a weight. The input pattern and the standard pattern are used. A frame having a large energy change is weighted in accordance with the degree of energy change of the frame. By this processing, the difference between the transient parts of the voice is largely reflected on the accumulated distance. Here, when the energy change degree S is negative, the value 0 is substituted for S in (Equation 6) so that the energy change portion in the negative direction is not weighted.

Now, it is assumed that the matching calculation is performed for the i-frame of the input pattern and the j-frame of the standard pattern. The control unit 72 first sets the frame of the input pattern.
Whether the recurrence formula of (Equation 6) is calculated for the value of the energy change degree Si of the unit i according to the limitation of the associating unit 71 which defines the limitation of the association as shown in (Equation 7), for example. Judge whether or not.

[0030]

(Equation 7)

For example, when Si is 2, according to the restriction of (Equation 7), when Sj is 1 or 2, matching calculation (Equation 4) is executed, and otherwise, it is not executed. As described above, the pattern matching calculation is performed for each standard pattern according to the restriction of the association unit 71 to calculate the similarity. FIG. 3 shows an example of the result of performing a matching calculation by the pattern matching unit 7 between two patterns.
The hatched portion in FIG. 3 indicates a range in which matching calculation can be performed. Here, the portion surrounded by the solid line can be used to calculate the route to the end point, and the accumulated distance can be obtained. However, in other portions, the route calculation is interrupted halfway. In FIG. 3, the path (1) is such that the peaks A1 and A2 of the energy of the input pattern are regarded as one peak and correspond to the peak B1 of the energy of the standard pattern, and A2 and B2 are similarly associated. This indicates the route at the time of making, and the route is interrupted on the way because there is no mountain of the input pattern corresponding to the mountain B3. Route (2) is A
1 represents a route when B1 corresponds to B2, A2 corresponds to B2, and A3 corresponds to B3. Paths (3) and (4) correspond to the case where mountain B1 is associated with the starting point of the input pattern because the degree of energy change of mountain B1 of the standard pattern is small. This is due to the effect of setting the degree of energy change at the start point to 0 in the initial setting unit 4, and is a matching path in consideration of the lack of the beginning of speech and the addition of noise. The same effect is obtained for the end point. In this manner, stable time-normalized matching between the transient part and the steady part of the voice is performed while taking into account a possible path.

If the pattern matching unit 7 performs pattern matching between two patterns having clearly different energy shapes, the matching path cannot reach the end point, and at this stage, the standard pattern Can be excluded from the recognition candidate voice.

Finally, the similarity comparing section 8 outputs a standard pattern giving the maximum similarity among the obtained similarities with each standard pattern as a recognition candidate.

As described above, according to the present embodiment, the smoothing unit 2 performs a smoothing process to eliminate a small change in energy, and the energy change degree determining unit 3 performs a smoothing process before and after each frame. The degree of change in energy is determined by a numerical value from the tendency of energy change between frames, and the numerical value is determined by the control unit 72 in accordance with the restriction on the matching of the matching unit 71 predetermined by the numerical value. By performing time-normalized matching in accordance with the slope limitation by weighting according to the degree of change, optimal matching between the transient part and the steady part of the energy change is realized without being affected by small fluctuations in energy. As a result, erroneous recognition between words having different energy shapes and erroneous recognition between words having similar phonemes can be significantly reduced.

Further, the initial setting unit 4 sets the degree of energy change at the beginning and end of the voice to 0 to allow a degree of freedom in matching at the beginning and end, so that even when a part of the voice is likely to be missing. In addition, matching path calculation is performed in consideration of the case of missing, and the recognition accuracy can be improved.

[0036]

As described above, according to the present invention, the smoothing unit for smoothing the shape of the energy and the rising and falling of the energy pattern before and after each frame of the voice pattern. The degree of energy change is numerically determined by an energy change degree determination unit, and the matching range is limited by a predetermined limit in accordance with the degree of change of the energy so that parts having similar change tendency are matched. A pattern matching unit that performs time-normalized matching using a recurrence formula that is limited and weighted according to the degree of energy change, and energy change at the beginning and end of speech
Initial setting to set energy change as if there is no
By providing the section, energy - energy without affected by the small changes - transition portion of the change, it is possible to accurately match the constant region comrades, recognition performance can be greatly improved. In addition, the effect of weighting according to the degree of energy change emphasizes differences in transient parts of speech, and in particular reduces false recognition between words having similar phonemes.
Can be. Furthermore, there is no change in energy at the beginning and end
Initial setting to allow the degree of freedom of association
In the case where a voice that is easy to
, Matching calculation considering missing initial endings
Is implemented and can provide a high-performance speech recognition device.
Wear.

The energy at the beginning and end of speech
By providing the initial setting unit that sets the energy-degree of change assuming no change, even when a speech that is apt to cause an error in speech section detection is input, a matching calculation is performed in consideration of the case where the initial ending is missing, A high-performance speech recognition device can be provided.

[Brief description of the drawings]

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

FIG. 2 is an explanatory diagram illustrating a slope limitation of time-normalized matching according to the present embodiment.

FIG. 3 is an explanatory diagram illustrating an example of a result of pattern matching according to the present embodiment.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 Voice analysis part 2 Smoothing part 3 Energy change degree determination part 4 Initial setting part 5 Input pattern memory 6 Standard pattern memory 7 Pattern matching part 8 Similarity comparison part 11 Spectrum parameter extraction part 12 Energy Parameter extraction unit 31 Energy difference code calculation unit 32 Energy change degree calculation unit 71 Correlation unit 72 Control unit

Continuation of front page (56) References JP-A-1-283599 (JP, A) JP-A-58-211199 (JP, A) JP-A-59-97200 (JP, A) JP-A-60-125899 (JP) JP-A-58-139200 (JP, A) JP-A-4-295895 (JP, A) JP-B-61-52478 (JP, B2) JP-B-2-23876 (JP, B2) (58) Field surveyed (Int.Cl. ⁶ , DB name) G10L 3/00 533 G10L 3/00 531 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. An audio signal comprising a speech characteristic spectrum and a positive
Speech component to extract time series of normalized logarithmic energy values
Analysis unit and an energy value series obtained from the voice analysis unit
And an energy for each frame at points other than the beginning and end of the speech using an energy difference value between adjacent frames from the output sequence of the smoothing unit. Energy-degree-of-change determining unit that determines the degree of change, and the beginning and end of speech
Always set the degree of energy change at the end to no change
An initial setting unit, the energy change degree determining unit,
In accordance with the degree of change in the energy of the input voice pattern set by the setting unit, the calculation range of matching is limited in accordance with a predetermined restriction on the association with the standard pattern, and the standard pattern and the input pattern are restricted. And a pattern matching unit for performing time-normalized matching between two patterns by a recurrence formula weighting the matching path according to the value of the energy change degree of the pattern. Voice recognition device.