JPH0558557B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a speech recognition device, and particularly to a speech recognition device using pattern matching technology.

(Prior Art) When performing speech recognition, it is very important to stably and accurately extract the features of the constant vowel part of input speech in order to improve recognition performance. This is because the constant portion of vowels in human speech occupies a larger proportion in time than the transitional portion (non-stationary portion), which is the portion that transitions from consonant or vowel to vowel, or from vowel to consonant, etc. In addition, since the duration is relatively long, there is little variation due to the influence of utterance timing, etc., and features can be extracted stably. Therefore, a recognition method that mainly uses the features of the vowel stationary part is effective. This is because.

A local peak extraction technique has been proposed as an effective technique for extracting features of vowel stationary parts in conventional devices. This technique attempts to detect the formant frequency band of the vowel stationary part.

FIGS. 3A to 3C are diagrams for explaining this technique. According to this technology, the center frequency (channel (numbering of center frequencies) number k corresponding to each center frequency) is
Frequency analysis and logarithmic transformation using different bandpass filters (marked with k is a positive integer) are performed sequentially at predetermined time intervals (hereinafter referred to as frames), and the resulting frequency spectrum is calculated. (Figure 3A), normalize the spectrum by subtracting the least squares approximation straight line of these spectra from these frequency spectra (Figure 3B), and in the channel where the value of the normalized spectrum is greater than "0". The local peak pattern is extracted as a 1-bit feature by setting the local peak value of the channel where the value of the output signal is maximum to ``1'' and setting all the local peak values of the remaining channels to ``0''. Figure 3C).

Calculate the similarity between the local peak pattern extracted above and the standard pattern prepared in advance, calculate the similarity for each recognition target category, and select the category name that gives the maximum similarity among all recognition target categories. Output as recognition result.

(Problems to be Solved by the Invention) The local peak is a feature that stably extracts the formant band of the vowel stationary part, and highly stable recognition can be performed.

However, it is difficult to stably extract features for consonant parts, such as fricatives such as /s/ and /ch/. This is because local peaking is a technique to extract the band where the normalized spectrum is maximum, and in the vowel stationary part, the normalized spectrum is maximum in the channel corresponding to the formant frequency band, which is the main feature of the vowel stationary part. However, the formant frequency can be extracted stably, but on the other hand, consonant parts such as fricatives do not have the property that the normalized spectrum becomes maximum in a specific channel (frequency band) like the formant frequency in the constant vowel part. Therefore, the appearance position of the local peak in the consonant part is unstable and difficult to be determined uniquely.

Therefore, when recognizing and judging speech in which both vowel stationary parts are the same, such as "ichi" and "shichi," it is difficult to accurately identify the two because the features of the consonant part cannot be stably extracted using local peaks alone. This has the problem of making it difficult to judge, leading to a decline in recognition performance.

This invention eliminates the above-mentioned problems and extracts the local peak features and consonant features of input speech as time-series patterns, and uses them to calculate the similarity with each standard pattern. The present invention aims to provide a speech recognition device that outputs accurate similarity and, as a result, has excellent recognition performance.

(Means for Solving the Problem) In order to achieve this objective, according to the present invention, a time-series pattern of features from the voice start time to the voice end time (speech interval) is extruded from the input voice. , calculate the similarity between this time series pattern and a standard pattern prepared in advance, calculate the similarity for each recognition target category, and select the category name that has the maximum similarity among all recognition target categories. In a speech recognition device whose recognition result is a. Frequency analysis using multiple channels (numbering of center frequencies), logarithmic transformation, frequency spectrum extraction, and normalization of the vocal fold sound source characteristics to the frequency spectrum. a spectrum normalization unit that calculates a normalized spectrum pattern, and b) a consonant in a frame determined to have consonant properties based on the magnitude relationship of the normalized spectrum values in the high-frequency and low-frequency channel regions within the speech interval; Consonant pattern extraction is performed, and in frames that are determined not to have consonant characteristics, consonant pattern extraction is not performed (the value is set to "0" in all channel components), and a consonant pattern is created. c. A consonant pattern extraction unit that A local peak pattern extraction unit that creates a local peak pattern for each frame calculates the similarity between the consonant pattern calculated in step d and b) and a consonant standard pattern prepared in advance, and calculates the similarity between each recognition target category. a consonantal similarity calculation unit that calculates the consonantal similarity for; e a consonantal standard pattern storage unit; and f the similarity between the local peak pattern calculated in section c) and the local peak standard pattern prepared in advance. a local peak similarity calculation unit that performs calculations and calculates the local peak similarity for each recognition target category; and a determining unit that determines the category name for which the total similarity is the largest among all recognition target categories as a recognition result.

Furthermore, in carrying out the present invention, the consonantity pattern extraction unit performs processing on all frames within the speech interval to create consonance patterns, so the normalized spectrum value in the a frame is high. frame determination means for determining that a frame in which the normalized spectrum value in the low frequency channel region is larger than the normalized spectrum value in the low frequency channel region has consonant properties; b. when determining that the frame has consonant properties; , the value of the consonant pattern of the channel component whose normalized spectrum value is larger than a predetermined threshold in this frame is set as the value of the normalized spectrum in this channel, and the value of the consonant pattern of the other channel components is set to "0". a first consonantity pattern value determination means which sets the value of the consonance pattern in this frame to "0" for all channel components if it is not determined that the frame has consonant properties;
Preferably, the method includes a consonant pattern value determining means.

(Operation) As described above, according to the speech recognition device of the present invention, the local peak similarity is added to the determining section, and the consonant pattern extraction section and the consonant similarity The consonantal similarity of the consonantal frame obtained based on the magnitude relationship of the normalized spectral values in the high-frequency and low-frequency channel regions is sequentially processed by the degree calculation unit, and this consonantal similarity is In addition to the determination section, this determination section performs recognition by calculating the total similarity obtained by adding the local peak similarity and the consonant similarity, so that accurate and stable speech recognition can be performed.

(Embodiments) Hereinafter, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a block diagram showing an embodiment of this invention.
FIG. 2A is a functional block diagram of the consonant pattern extraction section, and FIG. 2B is a flowchart for explaining the operation of the consonant pattern extraction section. Figures 1 and 2A
The configuration of an embodiment of the speech recognition device of the present invention will be explained using FIGS. and B along with its operation.

Input audio D1 is input to frequency analysis section 10.

The frequency analysis unit 10 performs bandpass filter analysis of a predetermined number of bands (hereinafter, the numbering of bands is referred to as a channel), and calculates the frequency spectrum D2, which is the output thereof, at each predetermined time interval (frame). , is output to the spectrum normalization section 11 and the speech section detection section 12.

The voice section detection unit 12 determines the start time and end time of the input voice from the magnitude of the value of the frequency spectrum D2, etc., generates a start time signal D3 and an end time signal D4, and sends both signals to the local peak pattern extraction unit 13. and output to the consonant pattern extraction section 14. The spectrum normalization unit 11 calculates a normalized spectrum D5 by subtracting the least square approximation straight line of the frequency spectrum from the frequency spectrum,
It is output to the local peak pattern extraction section 13 and the consonant pattern extraction section 14.

The local peak pattern extraction unit 13 sets the value of the local peak pattern of the channel in which the value of the normalized spectrum is maximum among the channels in which the value of the normalized spectrum in the frame is a positive value to "1",
The process of setting the value of the local peak pattern of all other channels to "0" is sequentially performed on all frames from the start frame to the end frame, and output to the local peak similarity calculation unit 15 as a local peak pattern D6. do.

The local peak similarity calculation unit 15 calculates the similarity between the local peak pattern D6 and all local peak standard patterns stored in the local peak standard pattern storage unit 16 in advance, and calculates the local peak similarity D8 for each recognition target category. is output to the determination unit 19.

Note that the frequency analysis unit 10, spectrum normalization unit 11, speech interval detection unit 12, local peak pattern extraction unit 13, local peak similarity calculation unit 15, local peak standard pattern storage unit 16, and determination unit 19 described above have already been implemented. Since this is a component of a speech recognition device based on the proposed local peak extraction technique, a detailed explanation thereof will be omitted unless it has a special function.

In the speech recognition device of the present invention, in addition to the local peak similarity D8 described above, the consonant similarity is added to determine the overall similarity in the determining section 19, so that the consonant pattern The extraction unit 14, the consonant similarity calculation unit 17, and the determination unit 19 are configured to operate as described below.

The consonant pattern extraction unit 14 is configured to create a consonant pattern D7 by a method described later in the explanation section of FIG. 2, and output it to the consonant similarity calculation unit 17.

The consonant similarity calculation unit 17 calculates the similarity between the consonant pattern D7 and all the consonant standard patterns stored in advance in the consonant standard pattern storage unit 18,
The consonant similarity degree D9 for each recognition target category is output to the determination unit 19.

The determination unit 19 calculates the sum of local peak similarity and consonant similarity for each recognition target category, and outputs the category name for which the total similarity value is the largest among all recognition target categories as recognition result D10. Configure it as follows.

By the way, the above-mentioned consonant pattern extraction unit 14
For example, as shown in FIG. 2A, frame determining means 2 determines whether the frame has consonant properties
0, the first consonant pattern value determining means 21 when the frame has consonant properties, and the second consonant pattern value determining means 2 when the frame does not have consonant properties.
2.

The frame determining means 20 determines that a frame in which the normalized spectral value in the high channel region is larger than the normalized spectral value in the low channel region has consonant properties. .

When it is determined that the first consonantity pattern value determining means 21 has consonant properties, the value of the consonance pattern of the channel component whose normalized spectrum value is larger than a predetermined threshold value in the frame is normalized in this channel. The value of the spectrum is determined, and the value of the consonant pattern of the other channel components is determined to be "0".

The second consonant pattern value determining means 22 determines that the consonant pattern values in the frame are "0" for all channel components when it is not determined that the frame has consonant properties.

Next, the operation of the consonant pattern extraction section 14, which is a main part of the embodiment of the present invention, will be explained in detail using the flowchart shown in FIG. 2B. In the following description, each processing step will be represented by S. Further, the operation procedure described here is only one preferred example, and therefore any other suitable procedure may be used.

The starting frame number is SFR, the ending frame number is EFR, and the normalized spectrum is NSP (i, j).
(i: channel number, j: frame number), consonant pattern by CMP (i, j) (i: channel number,
j; frame number), the number of frequency analysis channels
CHNNO. Furthermore, the frame number for which consonant pattern extraction is performed is assumed to be j.

First, in the frame determination means 20, j=
Perform initial settings for SFR (S1).

Next, the relative magnitude of the normalized spectrum output of the high frequency channel component to the low frequency channel component in the frame (frame number j) is calculated using the following equation (S2).

SUB(j)={ _CHNNO 〓 ^i=HS NSP(i,j)}/(CHNNO−HS+1)−{ _LE 〓 ⁱ⁼¹ NSP(i,j)}/LE……(1) However, HS and LE Empirically, LE=CHNNO/3, HS=2・
Set to about CHNNO/3. Then, it is determined whether or not the condition (A): SUB(j)>THL1 (THL1 is set at a predetermined threshold value of about 0) is satisfied (S3).

Next, in the first consonant pattern value determining means 21,
When the above condition (A) is satisfied, that is, when it is determined that it has consonant properties, the normalized spectrum output NSP (i, j) > THL2 Condition (B) (THL2 is set to about 0 with a predetermined threshold) (S4), and the value of the consonant pattern in the channel that satisfies condition (B) is set as CMP (i, j) = NSP (i, j) (S5).

The value of the consonant pattern in channels that do not satisfy the above condition (B) is set to CMP (i, j) = 0 (S6).

On the other hand, in the second consonant pattern value determining means 22,
When the above condition (A) is not satisfied, that is, when it is not determined that the frame has consonant properties, the value of the consonant pattern in all channels in the frame is set to CMP (i, j) = 0 (S6) .

After the first and second consonant pattern value determining means 21 and 22 finish extracting the consonant pattern in the frame, 1 is added to the frame number j (S7).

Next, it is checked whether the above-mentioned processes have been completed for all frames using the following condition: j≦EFR When this condition is satisfied, the process from step S2 is sequentially repeated, and when it is not satisfied, the process is The consonant pattern extraction in the input speech ends.

Next, FIGS. 4 and 5 are diagrams mainly for explaining the extraction results of local peak patterns and consonant patterns when the utterances are "ichi" and "shichi".

FIG. 4A and FIG. 5A are audio power diagrams corresponding to "ichi" and "shichi", respectively, with the frame number on the horizontal axis and the audio power on the vertical axis. Figure 4 B and C and Figure 5 B and C are local peak pattern diagrams and consonant pattern diagrams corresponding to "ichi" and "shichi", respectively, with the frame number on the horizontal axis and the channel number on the vertical axis, respectively. It is shown here. Further, FIGS. 5D and 5E are frequency spectrum diagrams and normalized spectrum diagrams corresponding to the consonant part of "shichi", and the channel numbers are shown on the horizontal axis.

For the uttered sound of “ichi” shown in FIG. 4A, the local peak pattern extraction unit (indicated by 13 in FIG. 1)
The region where the local peak pattern obtained in is "1" appears as shown by the black part in FIG. 4B. Furthermore, for the pronunciation of "ichi", the value of the consonant pattern obtained by the consonant pattern extraction unit (indicated by 14 in Figure 1) becomes equal to the value of the normalized spectrum at the corresponding channel number and frame number. The area is shown in FIG. 4C by the black area CO1.

On the other hand, in the case of the vocal sound "shichi" shown in FIG. 5A, the region where the local peak pattern is "1" and the value of the consonant pattern are equal to the values of the normalized spectrum at the corresponding channel number and frame number. The regions appear as the black areas CO2 and CO3 shown in FIGS. 5B and 5C, respectively. In this case, the spectral output of the initial part of the word "shichi" shown in FIG. 5A appears as shown in FIG. 5D, and its normalized spectral output becomes as shown in FIG. 5E, satisfying condition (A). Since it is satisfied, it is determined to be a consonant frame, and further, if the threshold THL2 of condition (B) is set to 0,
The region where NSP (i, j) > THL2 is satisfied and the value of the consonant pattern is equal to the value of the normalized spectrum at the corresponding channel number and frame number is
Denoted as CO2.

In this way, both "ichi" and "shichi" can be accurately identified based on the difference in the consonant patterns at the beginning of the word.

FIGS. 6A and 6B are explanatory diagrams of the overall similarity indicating the contribution of consonant patterns to recognition. FIG. 6A is a diagram showing the overall similarity of the uttered sound "ichi" having the speech pattern of FIG. 4A to the category name "ichi" and the category name "shichi", and FIG. FIG. 3 is a diagram showing the overall similarity of the uttered sound "shichi" having the voice pattern with respect to the category name "ichi" and the category name "shichi". In these figures 〓〓
The part indicated by 〓 indicates the local peak similarity, and 〓
The part indicated by 〓〓 indicates the degree of consonant similarity. This sixth
As can be understood from Figure A, the total similarity of the vocal sound "ichi" and the total similarity of the standard pattern of "shichi" are significantly different. Therefore, in either case, the two can be accurately identified based on the difference in consonant patterns.

(Effects of the Invention) As is clear from the above description, the speech recognition device of the present invention distinguishes between the time-series local peak pattern, which is the result of stably extracting the features of the vowel stationary part, and the consonant Since we adopted a method that performs recognition judgment based on both the time-series consonantity pattern, which is the result of stably extracting the features of
As a result, it is possible to perform accurate and stable speech recognition that also takes consonant features into account.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the speech recognition device of the present invention, FIG. 2A is a functional block diagram of the consonant pattern extraction section of the present invention, and FIG. 2B is a consonant pattern extractor of the present invention. FIG. 3 is a diagram showing the conventional local peak pattern extraction used to explain the present invention. FIG. 4 is an explanation of the local peak pattern and consonant pattern for the voice power of the uttered sound "ichi". FIG. 5 is an explanatory diagram of the local peak pattern and consonant pattern with respect to the voice power of the vocalization sound "shichi", and FIG. 6 is an explanatory diagram of the overall similarity indicating the contribution of the consonant pattern to recognition. 10...Frequency analysis section, 11...Spectrum normalization section, 12...Speech interval detection section, 13...Local peak pattern extraction section, 14...Consonant pattern extraction section, 15...Local peak similarity calculation section ,
16...Local peak standard pattern storage section, 18
... Consonance standard pattern storage unit, 19... Judgment unit,
20... Frame determining means, 21... First consonant pattern value determining means, 22... Second consonant pattern value determining means.

Claims (1)

[Claims] 1. Extracting a time-series pattern of features from an input voice from a voice start time to a voice end time (speech interval), and comparing the time-series pattern with a standard pattern prepared in advance. In a speech recognition device that calculates the degree of similarity for each recognition target category, and uses the category name with the highest degree of similarity among all recognition target categories as the recognition result, a. a spectrum normalization unit that calculates a normalized spectrum pattern by performing frequency analysis and logarithmic transformation (numbering of frequencies), extracting a frequency spectrum, and normalizing vocal cord sound source characteristics with respect to the frequency spectrum; Based on the magnitude relationship of the normalized spectrum values in the high and low channel regions, consonantity patterns are extracted for frames that are determined to have consonant properties, and frames that are determined not to have consonant properties are extracted. c. a consonant pattern extraction unit that sequentially performs a process of not extracting consonant patterns (setting the value to "0" in all channel components) in the frames in which the consonant pattern is extracted, and creates a consonant pattern; c. A local peak pattern that creates a local peak pattern by performing processing on all frames within a voice section to set the channel component whose value is a positive value and local maximum value to "1" and all other channel components to "0". an extraction unit; d) a consonant similarity calculation that calculates the similarity between the consonant pattern calculated in the above b) and a consonant standard pattern prepared in advance, and calculates the consonant similarity for each recognition target category; e a storage unit for the consonant standard pattern; a local peak similarity calculation unit that calculates the degree of similarity; 1. A speech recognition device comprising: a determination unit that determines the largest category name among target categories as a recognition result. 2. The consonant pattern extraction unit performs processing on all frames in the speech interval to create a consonant pattern, so that: a) the normalized spectrum in the high channel region for the normalized spectral value in the frame; frame determination means for determining that a frame whose value is larger than the normalized spectral value in the low channel region has consonant properties; b. when it is determined that the frame has consonant properties, the frame is subjected to the normalization in the frame; A first consonantity in which the value of the consonance pattern of a channel component whose spectrum value is larger than a predetermined threshold value is the value of the normalized spectrum in the channel, and the value of the consonance pattern of other channel components is set to "0". c) second consonantity pattern value determination means for setting the value of the consonantity pattern in the frame to "0" for all channel components if the frame is not determined to have consonant properties; A speech recognition device according to claim 1, characterized in that: