JPH0323920B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition processing device, and particularly to a speech recognition processing device that converts feature parameter vectors for each frame of uttered speech into subclassified sound type information, which is subclassified phonemes. In a speech recognition device that performs speech recognition processing using a time series of sound type information,
The conversion to subclassified note type information (note type conversion) is performed for each of the plurality of feature parameter vectors, and by using the time series of the obtained multiple subclassified note type information, the types of subclassified note types can be determined. The present invention relates to a speech recognition processing device that achieves an increased synergistic effect.

In conventional word speech recognition processing devices, when performing multi-word level recognition, the speech patterns of all registered words are held in a dictionary, and the speech patterns of all registered words are matched with the input speech to be recognized. Therefore, the memory capacity of the word dictionary and the amount of calculation become enormous, the restrictions on the device are quite severe for practical use, and the recognition rate is still not sufficient.

In addition, in the case of Japanese, devices that recognize input speech monosyllable by monosyllable have to store speech patterns of about 100 syllables in a dictionary and match them with the input speech, so dictionary memory and computational complexity are not a problem. However, it is difficult to recognize syllables that include consonants, especially plosives, etc. with current technology, and although there have been several attempts, they are still at a stage where they are not yet practical due to the recognition rate.

On the other hand, there is a method that converts the feature parameter vector for each frame of uttered speech into subclassified sound type information, which is subclassified phonemes, and performs speech recognition processing using the time series of the subclassified sound type information. Proposed. (Proceedings of the Acoustical Society of Japan, 1972)
(October, pp. 399-400) As shown in the paper, subclassified sound types are:
It classifies spectra using physical scales without being aware of phonemes. For example, the spectral patterns of all frames analyzed with a 13-channel filter (200 to 3400 Hz) are classified using the k-means clustering method, and each A spectral pattern that is the center of a cluster can be used as a subclassified sound type.

An object of the present invention is to use two or more different feature parameter vectors when converting uttered speech into subclassified sound type information, which is phonemes subdivided into frame units. By obtaining information on more than one type of sub-classified note type, it is possible to reduce the loss of voice feature information due to the synergistic effect of the two or more sub-classified note types. In order to achieve this object, the present invention includes a speech analysis processing section that analyzes input speech for each frame at a predetermined time interval, a feature extraction processing section that extracts speech features from the analysis results, and a phoneme extraction processing section that extracts speech features from the analysis results. a time series conversion processing unit that converts the input speech into a time series of subclassified note types, a word dictionary that stores in advance the time series of subclassified note types of a plurality of registered voices, and a time series of subclassified note types of input speech. and a matching processing unit that calculates the distance between the input speech and a plurality of registered speeches by matching the time series of the subclassified phonetic types of the registered speeches in the word dictionary; In the speech recognition processing device, the time series conversion processing unit uses a plurality of feature parameters,
A subclassified note type dictionary unit configured to perform conversion into subclassified note types for each feature, and further coupled to the above-mentioned time series conversion processing unit, and holds pre-created subclassified note type information for each feature. In addition to providing
The word dictionary is composed of a word dictionary that holds subclassification sound type information for each feature of speech registered words, and during speech recognition processing, the time series conversion processing section uses the subclassification sound type dictionary section to is converted into a time series of subclassified note type information for each feature, and the matching processing unit converts the time series of subclassified note type information for each feature obtained by the conversion into a time series of subclassified note type information for each feature into a time series of subclassified note type information for each feature. It is characterized by matching processing with subclassified note type information.

First, the principle of the present invention will be explained below. now
Let P ₁ be a set of subclassified note types obtained by feature I. The time series of feature parameter vectors within the voice section extracted by the voice feature extractor is converted into a time series _VI of the number of this subclassified note type by the time series converter.

V _I =P _a1 ⁽¹⁾ P _a2 ⁽²⁾ ...P _ao ⁽ⁿ⁾ ...Equation 1 ∀P _ai ⁽ⁱ⁾ ∈P _I (1<i<M) (i): Frame number M of input audio :Total number of subclassified note types of feature I. Also, let P _II be the set of subclassified note types obtained by the feature. The time series of the feature parameter vectors in the voice section extracted by the voice feature extraction section is converted into a time series of this subclassification note type number _VII by the time series conversion section.
is converted to

V _II = Q _b1 ⁽¹⁾ Q _b2 ⁽²⁾ ...Q _bo ⁽ⁿ⁾ ...Formula 2 Q _bi ⁽ⁱ⁾ ∈P _II (1<i<N) (i): Frame number N of input audio: Total number of subclassified sound types of the feature Here, the subclassified sound type of each frame in the time series V _I obtained by the feature is one of the elements of the set P _I ,
P _ai ⁽ⁱ⁾ of the i-th frame describes information regarding the feature I of the i-th frame of the input audio. Then, if we add the time series VI _II obtained by another feature, one of the set of N subclassified note types P _II will be selected for the i-th frame of VI _II . , when the number of the subclassified note type of the i-th frame of the input voice obtained by the feature is used together with the number of the subclassified note type obtained by the feature, the number of combinations of each note type number is (M × N). Become.

From this, it can be seen that using the subclassified note types obtained from the features has the same effect as having (M×N) subclassified note types. For example, if the number of subclassified sound types obtained by the features and characteristics is respectively M = 100 and N = 100, there will be 10,000 subclassified sound types, and the uttered voice will be 10,000. It can be expressed as a time series of selected sounds from subclassified note types, and the loss of information about voice features is significantly reduced compared to the conventional case where both features and characteristics are described using 100 subclassified note types. It has the effect of Also, compared to the case where 10,000 subclassified note types are retained, in this example, 200 (=M
It is sufficient to hold +N) subclassified note types, and the amount of memory is reduced to 1/50.

Hereinafter, the present invention will be explained with reference to the drawings.

The figure is a block diagram of a speech recognition processing device according to an embodiment of the present invention, in which 1 is an input medium section, 2 is a speech analysis processing section, 3 is a feature extraction processing section, 4 is a time series conversion processing section, and 5 6 is a matching processing unit, 6 is a determination processing unit, 7-1 to 7-n are sub-classified note type dictionary units that hold sub-classified note types, 8-1 to 8-n are word dictionary units, 9
is the result output.

First, the uttered voice is input from the input medium section 1 (microphone, etc.), and the voice analysis section 2 performs LPC analysis on every frame at intervals of several ms to several tens of ms.
The audio is analyzed using methods such as FFT analysis.

Next, in the feature extraction processing unit 3, voice feature parameters are extracted or obtained according to the characteristics of a plurality of bandpass filters arranged at equal intervals on a frequency scale based on auditory characteristics based on the above analysis results. Several peak frequencies (formant frequencies) of the spectral shape are extracted. Furthermore, the feature extraction processing unit 3 extracts n feature information different from n bands such as a high frequency band and a low frequency band each consisting of a plurality of parameters, thereby generating n feature parameter vectors. Output the time series.

Also, based on the voice data (registered data) uttered by the speaker in advance, the voice analysis processing section 2 and the feature extraction processing section 3 create a time series of n feature parameter vectors, and A subclassified note type in frame units is created for each feature from the series data, and is registered in the subclassified note type dictionary sections 7-1 to 7-n together with identification numbers. For example, a method using cluster analysis is known for creating subclassified note types.

Using the n types of subclassified note types registered above, the voice is converted into subclassified note type numbers for each n type of feature parameter vector series. This conversion is performed by the time series conversion processing section 4, and a time series of subclassified note type numbers is output for each of the n types of feature information.

The above-mentioned processing is applied to the utterance of the word to be recognized, and it is converted into n types of time series of subclassified phonetic numbers, and each of them is registered in advance in the word dictionary sections 8-1 to 8-n. I'll keep it. That is, the word dictionary units 8-1 to 8-n each register a time series of numbers of n types of subclassified sounds related to n types of features of registered speech uttered by a speaker.

The matching processing unit 5 performs matching between the n word dictionary units 8-1 to 8-n and the time series of the numbers of the n types of subdivided sounds related to the n types of features of the input speech to be recognized using the DP method, etc. Calculated using the method. for example
During matching processing using the DP method, the distance between each frame of the input audio and each frame of the registered audio required to execute the DP method is calculated based on the characteristics of the corresponding frame of the input audio and the corresponding frame of the registered audio. It can be calculated by adding up all the distances. The distance for each feature is determined by the number of the subclass note type of the relevant frame of the input audio and the number of the subclass note type of the relevant frame of the registered audio, and the distance between the subclass note types regarding that feature. It can be obtained by searching a distance matrix in which . Finally, the determination processing unit 6 outputs the name of the registered word with the minimum distance as a recognition result.

The present invention will be explained below based on an operation example.

To facilitate understanding, there are two types of feature parameter vectors to be analyzed: the first feature parameter consists of 50 subclassified note types, and the second feature parameter consists of 80 subclassification note types. shall be taken as a thing. It is also assumed that the input voice is uttered with a length of 1 second, and the voice analysis processing section 2 analyzes the voice in frames at intervals of 10 ms. Furthermore, there are three words to be recognized: “Osaka”, “Kanagawa”, and “Tokiyo”.
Let it be a word. Furthermore, it is assumed that the matching process uses a linear matching method.

It is now assumed that a time series of subclassified note type information (specifically, a time series of subclassified note type numbers) for 100 frames (=1 second/10ms) shown below has been obtained for the unknown input voice.

Time series information regarding the first feature (obtained by the subclassification note type dictionary section 7-1): (1) (2) (3) (4) (100) [43, 5, 18, 36,... , 24] Time series information regarding the second feature (obtained by the subclassification sound type dictionary section 7-2): (1) (2) (3) (4) (100) [14, 59, 64, 75,...,33] On the other hand, the word dictionary section stores the time series of subclassified sound type information shown below for the three words "Osaka", "Kanazawa", and "Tokiyo" that have been uttered in advance. shall be.

Time series information regarding the first feature stored in the word dictionary 8-1: (1) (2) (3) (4) (100) Osaka [20, 50, 40, 25,..., 49] (1) (2) (3) (4) (100) Kanazawa [9, 24, 31, 15,..., 30] (1) (2) (3) (4) (100) Tokyo [40, 7, 15, 38] ,..., 29] Time series information regarding the second feature stored in the word dictionary 8-2: (1) (2) (3) (4) (100) Osaka [25, 18, 78, 25,..., 59] (1) (2) (3) (4) (100) Kanazawa [70, 28, 31, 42,..., 11] (1) (2) (3) (4) (100) Tokyo [12, 63, 58, 73,..., 36] The matching processing unit 5 compares the time series of the subclassified note type information obtained regarding the unknown input speech with the time series of the subclassified note type information stored in the word dictionary unit. Find the distance by searching the distance matrix.

Now, as the distance regarding the first feature, (1) (2) (3) (4) (100) Osaka [23, 45, 22, 11,..., 25] (1) (2) (3) (4) ( 100) Kanazawa [34, 19, 13, 21, ..., 6] (1) (2) (3) (4) (100) Tokyo [3, 2, 3, 2, ..., 5] is obtained, and the As the distance regarding the two features, (1) (2) (3) (4) (100) Osaka [11, 41, 14, 50,..., 26] (1) (2) (3) (4) (100) Suppose that Kanazawa [56, 31, 33, 33, ..., 22] (1) (2) (3) (4) (100) Tokyo [2, 4, 6, 2, ..., 3] are obtained. In this case, by adding up all the distances for each feature and each frame for the word to be matched, it is found that the word "Tokyo" has the smallest distance from the unknown input voice,
"Tokyo" is output as the recognition result.

As explained above, according to the present invention, it is possible to create a situation in which a large number of subclassified sound types actually exist by combining a small number of subclassified sound types, thereby reducing the loss of characteristic information of uttered speech. can be suppressed. Furthermore, the amount of dictionary memory can also be reduced.

[Brief explanation of the drawing]

The figure is a block diagram of a speech recognition processing device according to an embodiment of the present invention. In the figure, 1 is an input medium section, 2 is a speech analysis processing section,
3 is a feature extraction processing unit, 4 is a time series conversion processing unit, 5
6 is a matching processing section, 6 is a judgment processing section, 7-1 to 7-n
8-1 to 8-n are word dictionary sections, and 9 is a result output.

Claims (1)

[Claims] 1. A speech analysis processing unit 2 that analyzes input speech for each frame at a predetermined time interval; a feature extraction processing unit 3 that extracts speech features from the analysis results; and segmentation of input speech into phonemes. a time series conversion processing unit 4 that converts the input speech into a time series of subclassified note types, a word dictionary 8 that stores in advance a time series of subclassified note types of a plurality of registered voices, and a time series of subclassified note types of input speech. a matching processing unit 5 that calculates the distance between the input speech and a plurality of registered speeches by matching the time series of subclassified phonetic types of the registered speeches in the word dictionary 8; In a speech recognition processing device having a determination processing unit 6 that determines the one with the smallest distance from the other, the time series conversion processing unit 4 uses a plurality of feature parameters and converts each feature into subclassified sound types. It is further provided with subclassification note type dictionary units 7-1 to 7-n which are coupled to the time series conversion processing unit 4 and hold subclassification note type information for each feature formed in advance. , the word dictionary 8 is a word dictionary 8 that holds detailed classification sound type information for each feature of voice registered words.
-1 to 8-n, and during speech recognition processing, the time series conversion processing unit 4
converts the input voice into a time series of subclassification sound type information for each feature using the subclassification sound type dictionary units 7-1 to 7-n, and the matching processing unit 5 converts the input speech into a time series of subclassification sound type information for each feature. A speech recognition processing device characterized in that a time series of subclassified sound type information for each feature is matched with corresponding subclassified sound type information in word dictionaries 8-1 to 8-n.