JPH02134700A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To speed up voice recognition by deciding a voice according to feature parameters from a parameter extracting means which extracts the feature parameters according to a detected formant frequency. CONSTITUTION:When a voice signal S is inputted to a voice signal dividing means 1, the voice signal S is outputted as voice division signals Si which are divided by specific frequency bands. Then adjacent voice division signals Si-1 and Si are compared by a comparing means 2 and converted into binary signals. Its comparison information is inputted to a parameter extracting means 3. Then the parameter extracting means 3 detects the formant frequency according to the input comparison information and extracts the feature parameters according to the detected formant frequency, and then a voice decision means 4 decides the input voice signal S according to the feature parameters. Thus, the time of the detecting operation for the formant frequency is shortened to speed up the voice recognition.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition device, and particularly to an improvement of a speech recognition device of a type that extracts formant frequencies themselves or information based on formant frequencies as feature parameters.

[Prior Art] In general, as a speech recognition method, for example, a vowel to be extracted (in this example, a speech signal of U'' (see FIG. 8(a)) is used.
Figure 8 (b)
) is widely used to extract the absolute values of the frequencies of the first and second formants as characteristic parameters from the characteristic curve (see Figure 8 (C)) obtained by linear predictive analysis (LPG analysis). has been done. Furthermore, Figure 8 (C
), Fl to F4 represent the first to fourth formants.

For example, as shown in FIG. 9, a device for realizing such a conventional voice recognition method includes a microphone 10 that converts a voice such as a vowel into an electrical voice signal, and a device that converts a voice signal input from the microphone 10 into an electric voice signal. A signal adjustment circuit 11 consisting of an amplifier and a low-pass filter that cut unnecessary high frequency components after amplification, and an auto gain control circuit (hereinafter abbreviated as AGC circuit) that adjusts the amplitude level of the audio signal; An AD converter 12 that digitizes an analog signal, a frequency analyzer 13 that samples the digitized audio signal at regular intervals and performs frequency analysis, and a frequency analyzer 13 that performs frequency analysis by sampling the digitized audio signal at regular intervals; A formant detection unit 14 extracts the spectral peak of A device configured with a voice discriminator 15 is already known.

[Problem to be solved by the invention 1 By the way, in such a conventional speech recognition device,
After the audio signal is digitized, the formant frequency is detected by software processing on a CPU board, etc., which inevitably takes time to detect the formant frequency, which makes it difficult to perform high-speed speech recognition. It included the problem of becoming.

Such problems are not limited to those that directly extract formant frequencies as feature parameters;
The same problem may occur in devices that detect formant frequencies and extract feature parameters based on the detected formant frequencies.

The present invention has been made in view of the above-mentioned problems, and it is an object of the present invention to provide a speech recognition device that aims to shorten the formant frequency detection operation time and thereby speed up speech recognition. be.

[Means for Solving the Problems] As shown in FIG. a comparing means 2 that compares and binarizes the output level of the audio divided signal sr < r = 1 to n) for each frequency divided by the means 1 between adjacent channels; and comparison information from the comparing means 2. a parameter extracting means 3 that detects a formant frequency based on the detected formant frequency and extracts a feature parameter based on the detected formant frequency, and a voice discriminating means that discriminates speech based on the feature parameter from the parameter extracting means 3. This is a speech recognition device equipped with 4.

In such technical means, the scope of the present invention can be applied to any device in which formant frequencies can be detected in speech recognition.

In this case, the formant frequency itself and its logarithm information can be selected as appropriate as characteristic parameters for recognizing speech, but from the viewpoint of minimizing individual differences when recognizing speech from unspecified speakers. Therefore, it is preferable to use a relative value of formant frequencies, such as a ratio between formant frequencies or logarithmic difference information of formant frequencies, which does not depend on the formant frequencies themselves, as a feature parameter.

Further, the number of divisions of the audio signal dividing means 1, the range of frequency bands to be divided, etc. can be appropriately selected depending on the formant frequencies of various voices. In this case, if the type uses logarithmic information of the formant frequency as a feature parameter, logarithmic transformation may be performed by the parameter extraction means 3, but if speeding up speech recognition is intended, band pass The audio signal dividing means 1 is configured by arranging filter groups logarithmically or by combining a band filter group and a logarithmic conversion means, and the formant frequency is substantially logarithmically converted in advance by the audio signal dividing means 1. It is preferable to do so.

Furthermore, the design of the comparing means 2 may be changed as appropriate as long as it can compare and binarize the adjacent audio divided signals 3i-1 and 3i, which are analog signals.

Furthermore, the parameter extracting means 3 includes a formant detecting section that detects a formant frequency based on the comparison information from the comparing means 2, and a parameter calculating section that calculates feature parameters based on the detection information from this formant detecting section. As long as it is equipped with the following, the design may be changed as appropriate. In this case, when recognizing voiced consonants,
Since each formant changes over time, it is necessary to design the system so that the changes over time can be stored in memory when extracting feature parameters.

Further, the speech discrimination means 4 is provided with reference information regarding the speech to be recognized, and if it is capable of comparing the feature parameters and the reference information and recognizing the matched speech as the input speech, the design is appropriately changed. be able to.

[Operation] According to the above-mentioned technical means, when the audio signal S is input to the audio signal dividing means 1, the audio signal S is output as the audio divided signal 3i divided into each predetermined frequency band. . Then, the adjacent audio divided signals 3i-1 and Si are compared and binarized by the comparing means 2, and this comparison information is input to the parameter extracting means 3.

Next, the parameter extraction means 3 detects a formant frequency based on the input comparison information, and also extracts a feature parameter based on the detected formant frequency. determines the input audio signal S based on the characteristic parameters.

[Embodiments] Hereinafter, the present invention will be described in detail based on embodiments shown in the accompanying drawings.

FIG. 2 is a block diagram showing an embodiment of the speech recognition device according to the present invention, in which vowels and voiced consonants are recognized as speech using logarithmic differences between adjacent formant frequencies as speech feature parameters. It is something.

In the same figure, 21 is a microphone that converts audio into an electrical audio signal, 22 is an amplifier and low-pass filter that amplifies the audio signal input from the microphone 21 and then cuts unnecessary high frequency components, and an audio signal converter. A signal adjustment circuit consisting of an AGC circuit that adjusts the level, 23
is a bandpass filter (specifically, a channel ch
(i) (i-1 to 15)) (corresponding to the audio signal dividing means 1 of the present invention), 24 is the output between adjacent channels ch (i-1) and ch (i) of the bandpass filter 23 A binarization circuit (corresponding to the comparison means 2 of the present invention) that compares the levels and then binarizes and outputs the result, 25 determines which vowel or voiced consonant the audio signal from the binarization circuit 24 is. This is a CPU board (corresponding to the parameter extraction means 3 and the voice discrimination means 4 of the present invention). Note that the output from the bandpass filter 23 passes through a rectifying and smoothing circuit (not shown) and is converted into a DC signal having only intensity information.

In this embodiment, the binarization circuit 24 has a third
As shown in the figure, 14 comparators 24 (j)
(j = 1 to 14), and each comparator 24 (j) satisfies ch (j+1) > ch (
j), Cj='"1", otherwise Cj-"°
0'' is output.

Further, the CPU board 25 is designed to perform control operations according to the flowchart shown in FIG. Specifically, the CPtJ board 25 receives each comparison signal cj<ch(j+1)-ch(j)) from the binarization circuit 24.
Find out where the sign changes from positive to negative based on (
Step 1), channel ch changing from positive to negative
Number (i) (in this example, it is ml or m3)
is detected as a channel with formant (step 2). Thereafter, the CPU board 25 calculates the difference between the channel number m2 where the second formant is located and the channel number m1 where the first formant is located as a first parameter L f
12, and the difference between the channel number m3 where the third formant is located and the channel number m2 where the second formant is located is extracted as the second parameter Lf23 (step 3.4), and then the - and second parameters After storing the temporal changes of 1-f12 and 1-f23 in the memory, DP matching is performed (steps 5 and 6).

Thereafter, the CPU board 25 compares it with preset reference data based on the first and second parameters 1f12°Lf23 and determines what the input audio signal S is. and output it (step 7).

Next, a case will be described in which, for example, a vowel is recognized by the speech recognition device according to this embodiment.

Nowadays, unspecified speakers (men:
Each vowel “a, i” by MALE, FEMALE)
, u, e, o'''s first and second parameters Lf12
, 1J23 were determined, and the respective values were plotted in the voice feature space with the coordinate axes of the second and second parameters Lf12 and Lf23, and a distribution diagram as shown in FIG. 5 was obtained. In addition, in the same figure, MAL
E is (blacked out).

FEMALE is ○ (white), a is o, i is IU, e is 0, ◇.

According to this distribution map, each vowel is equally spaced in relatively narrow regions 3a, 3i, and 3u, regardless of whether it is male or female.

It is understood that the distribution is Se,3o (encircled by a dotted line in the figure). This means that in the speech feature space, the distribution areas of each vowel almost never overlap with each other, and each vowel distribution area 3a to SO
will be reliably divided by the partition lines 11 to 15.

Therefore, the reference values of the first and second parameters Lf12 and 1J23 corresponding to each vowel "a, i, LJ, e, o" are determined for each fa area S a based on the values of the partition lines 11 to i5. If it is set in advance to include SO or SO, the first and second parameters of the input audio information will be f1.
The audio discrimination unit (step 7) of the CPU board 25 determines in which region 2, 1f23 is included, and the input audio information is accurately determined.

Note that voiced consonants can also be recognized in the same manner as described above by setting time-varying reference data in advance and comparing them with this data.

In such a speech recognition process, since the bandpass filter 23 and the binarization circuit 24 are used in the process of detecting the formant frequency, the CPU board 25 checks the sign of the output signal of the binarization circuit 24. It becomes possible to detect the formant frequency with just this. Therefore, most of the formant frequency detection means can be configured as hardware, and the formant frequency detection operation time is shortened accordingly.

In addition, in this embodiment, the logarithmic information of the formant frequency is determined by hardware by using a special bandpass filter 23, so the logarithmic difference of the formant frequency is calculated within the CPU board 25. There is no need to perform calculations to obtain the result, and furthermore, there is no need to perform multiplication and division cylinders, which is necessary when calculating the ratio of formant frequencies. Therefore, the amount of calculation within the CPU card 25 is reduced, and the extraction time for feature parameters can be shortened accordingly, and integration, such as using an LSI, is facilitated.

In addition, in this example, as a feature parameter,
Since the formant frequency logarithmic difference information 1ft2°L r23 is used, the speech recognition rate is high.

This will be described in relation to Comparative Example 1 and Comparative Example 2 below.

First, Comparative Example 1 uses the formant frequency itself as a characteristic parameter of the voice, and each vowel "a, i, u, e" by an unspecified speaker including a plurality of men and women.
, o", the frequencies of the first formant f1 and the second formant f2 were determined, and the respective frequencies were plotted in the speech feature space with the - and second formants f1 and f2 as the coordinate axes, as shown in Fig. 6. A distribution map like this was obtained.

According to the distribution map according to Comparative Example 1, the distribution areas 3a to SO of each vowel are greatly expanded, and the distribution areas of each vowel are arranged close to each other.
It can be seen that the distribution areas of ゛′ partially overlap.

For this reason, it becomes difficult to clearly demarcate the voice distribution area, which makes it more likely that voice recognition will be misrecognized, and the frequency of voice misrecognition is particularly high between the above-mentioned "゛U" and e. It gets expensive.

Comparative Example 2 uses the ratio of adjacent formant frequencies as a voice characteristic parameter, and each vowel "a, i, u,
e, o'' first formant f1 to third formant f
Find the frequencies of 3 and the ratio f of the -th and second formants.
l /f2, the ratio of second and third formants f3 /f
When each frequency was plotted in the voice feature space with 2 as the coordinate axis, a distribution diagram as shown in FIG. 7 was obtained.

According to the distribution map of Comparative Example 2, the distribution area of each vowel is more clearly demarcated than in Comparative Example 1, but the vowel "1'
It can be seen that the distribution area Si of ' is much wider than the others, and that the size of the distribution area of vowels varies. Therefore, the distribution region 3i of the vowels 11 i II is unnatural compared to the distribution regions of other vowels. Also, if the size of the vowel distribution area varies,
In order to improve speech discrimination accuracy, when weighting is used for processing, the number of counts varies for each vowel, and the problem with weighting is that the processing becomes troublesome. Put it away.

Therefore, the 11 rate of speech in Comparative Examples 1 and 2 is lower than that in the Examples. In addition, in the case of these comparative examples, the method using only subtraction as used in the example cannot be used, and multiplication and division must be performed.
This is disadvantageous in terms of calculation time.

Furthermore, when the logarithmic difference between the first and second formant frequencies and the third formant frequency is used as a characteristic parameter, the distribution diagram in the example is rotated clockwise, and the straight lines dividing each vowel become complicated. Become. For this reason, it is difficult to use a method of dividing each vowel by a straight line parallel to the coordinate axis as in the embodiment, and it is not preferable because it takes time to calculate for discrimination.

In this embodiment, vowels and voiced consonants are recognized as sounds, but if voiceless consonants are also to be recognized, for example, as shown by the phantom lines in FIG. After converting the parallel audio 1.11 signal S1 from the pass filter into a serial signal by the multiplexer 31, it is digitized by the AD converter 32 and sent to the CP.
On the other hand, an unvoiced consonant discriminator (not shown) is provided in the CPLI board 25, and this unvoiced consonant discriminator processes data from the AD converter 32 to also discriminate unvoiced consonants. It is possible to do so.

[Effects of the Invention] As explained above, according to the speech recognition device according to the present invention, the device configuration of the process of detecting the formant frequency of the speech signal is devised to shorten the detection time of the formant frequency. Therefore, speech recognition based on feature parameters based on formant frequencies can be performed at high speed.

Furthermore, in the speech recognition device according to the second aspect, since logarithmic difference information of formant frequencies is used as a feature parameter, it is possible to form a speech feature space corresponding to an auditory space that does not directly depend on formant frequencies. For this reason,
It is possible to distribute the characteristic parameters of the same voice by an unspecified speaker in a relatively narrow range, and it is possible to reliably separate the distribution areas of different voices, thereby reducing misrecognition of the voice. can.

In particular, in the speech recognition device according to claim 3, there is no need to perform logarithmic difference calculation of formant frequencies in the formant extraction means, so when the formant extraction means performs logarithmic difference calculation of formant frequencies by software processing. The extraction time for feature parameters can be further shortened compared to .

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram showing the schematic configuration of a speech recognition device according to the present invention, FIG. 2 is a block diagram showing an embodiment of the speech recognition device according to the invention, and FIG. 3 is a binarization diagram according to the embodiment. An explanatory diagram showing the details of the circuit, FIG. 4 is a flowchart showing specific processing of the CPU board according to the embodiment, FIG. 5 is a graph diagram showing the audio feature space based on the embodiment, and FIGS. Figure C is a graph showing the feature space of speech based on Comparative Example 1 and Comparative Example 2. Figure 8 is a graph showing the formant extraction process in a conventional speech recognition method. Figure 9 is an example of a conventional speech recognition device. FIG. [Explanation of symbols] S...Audio signal Si...Speech division signal 1...Audio signal division means 2...Comparison means 3...Formant extraction means 4...Speech discrimination means

Claims (1)

[Claims] 1) Audio signal dividing means (1) that divides the audio signal (S) into predetermined frequency band components and outputs the divided components, and each frequency divided by the audio signal dividing means (1). a comparison means (2) that compares the output level of each audio division signal (Si) between adjacent channels and binarizes the same, and detects a formant frequency based on the comparison information from the comparison means (2); Parameter extraction means (3) for extracting feature parameters based on the formant frequencies
), and a speech discrimination means (4) for discriminating speech based on the characteristic parameters from the parameter extraction means. 2) The speech recognition device according to claim 1, characterized in that logarithmic difference information of formant frequencies is used as the speech feature parameter. 3) The audio signal dividing means (
1) consists of a group of band-pass filters arranged logarithmically, and the parameter extraction means (3) consists of a formant detection section that detects a formant frequency and an audio signal division means (1) that corresponds to the detected formant frequency. A speech recognition device comprising: a parameter calculation unit that calculates a channel number difference as a feature parameter.