JPH0632003B2 - Word registration method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an improvement in a word registration method applied to a specific speaker's word voice recognition device.

(Prior art and its problems) In a specific speaker's word voice recognition device, each standard pattern registered in advance is compared with an input unknown voice pattern to select a word giving the maximum similarity. Recognition is performed by. In registering a standard pattern, it is necessary to cut out a word voice section, and end point detection is performed using an amplitude parameter of a voice signal. However, because of the user's tongue-clicking sound, breathing sound, and ambient noise, the word voice section may be erroneously registered, which causes erroneous recognition.

An object of the present invention is to realize more accurate word detection by detecting the end points of word speech in consideration of the phonetic structure of a word to be registered, and thereby to realize speech recognition with high recognition accuracy. And

(Means for Solving Problems) A word registration method provided by the first invention of the present application in order to solve the above problems is such that when registering a registered voice uttered by a user as a standard pattern, environmental noise is generated. A temporary word standard as a combined standard pattern by connecting the phoneme class standard pattern to a registered word expressed as a phoneme class sequence, having means for storing the pattern and the phoneme class standard pattern of the unspecified speaker. A pattern is generated, registration is performed corresponding to the boundary position between the combined standard pattern and the environmental noise pattern by performing matching based on the time axis correspondence between the pattern in which the environmental noise pattern is connected before and after the combined standard pattern and the registered voice. The feature is that the end point of the registered voice is detected by extracting the position on the voice.

In addition, the word registration method provided by the second invention of the present application, when registering the registered voice uttered by the user as a standard pattern, an unspecified speaker to which the environmental noise pattern and the phoneme class position information are added. Having a means for storing the basic word standard pattern, by performing matching by time axis correspondence between the basic word voice uttered by the user in advance and the pattern in which the environmental noise pattern is connected before and after the basic word standard pattern, The phoneme class standard pattern is obtained by extracting the phoneme class position of the basic word speech corresponding to the phoneme class position of the basic word standard pattern, and the phoneme class standard pattern is concatenated with the registered word expressed as a phoneme class sequence. By generating a temporary standard pattern as a combined standard pattern, and connecting the environmental noise pattern before and after the combined standard pattern. And detecting the end point of the registered voice by extracting the position on the registered voice corresponding to the boundary position between the combined standard pattern and the environmental noise standard pattern by performing matching based on the time axis correspondence between the pattern and the registered voice. Is characterized by.

(Operation) The basic principle of the present invention will be described below.

In order to register a word in the recognition device, it is necessary to detect the start point and end point (end point) of the spoken word voice,
For this reason, generally, the amplitude parameter of the voice signal is used, but before and after the uttered word voice, a fluttering sound, a breath sound,
If background noise is added, the end points may be erroneously detected. The present invention utilizes the phonetic structure of a word in order to accurately detect the end points of the word even in the presence of such noise. Here, as a phonetic expression of a word, a method of expressing it as a series of word phoneme classes is used. Phoneme classes are generally / a /, / i /, / u /, / e /, / o
It refers to phonemes such as /, / p /, / k /, / t /, / s /, / z /, / n /, / m / etc., but it is a group of multiple phonemes. (For example, / p, k, t / can be combined into one phoneme class as unvoiced plosives and / s, z / as fricatives).

If the phoneme class representation of the spoken word is known, the phoneme class standard pattern can be used to detect word intervals as follows. Let the phoneme class representation (ie, the phoneme class sequence) of the spoken word be {s ₁ , s ₂ , ..., S _k , ..., S _k }. Further, it is assumed that a phoneme class standard pattern X (s) represented as a feature vector is given for each phoneme class s. At this time, the phoneme class standard patterns are connected to generate a combined standard pattern X of words.

X = {X (s ₁ ), ..., X (s ₁ ), X (s ₂ ), ..., X (s ₂ , ..., X (s _k ), ..., X (s _k )} where each A specific number of phoneme class standard patterns X (s _k ) are arranged, and the environmental noise pattern is represented by a feature vector N, and an extended joint standard pattern X ^{* in} which the ambient noise pattern is added before and after the joint standard pattern ^. think of.

X ^* = {N, X (s ₁ ), ..., X (s ₁ ), X (s ₂ ), ..., X (S ₂ ), ..., X (s _k ), ..., X (s _k ), N} = {X ^* (1), X ^* (2), ..., X
^* (i), ..., X ^* (1)} The environmental noise pattern is an average pattern of environmental noise,
It can be obtained by observing and averaging inputs for a certain period of time before uttering. On the other hand, the input word speech is set as a feature vector sequence Y = {Y (1), Y (2), ..., Y (J)}. The two feature vector series patterns X ^* and Y can be associated with the time axis non-linearly by DP matching using the well-known dynamic programming (The principle of DP matching is the Acoustical Society of Japan Vol.27, No. 9, p. 483).
That is, the following recurrence formula may be solved for the cumulative amount g (i, j) regarding the inter-vector distance d (i, j) = distance (X ^* (i), Y (j)).

Initial condition: g (0,0) = 0, g (i, 0) = ∞ (i = 1, ..., I), g (0, j) = ∞ (j = 1, ..., J) Recurrence formula : This recurrence formula calculation can be executed along the time axis of the input voice, and the temporary end point (i,
The matching path from j) can be traced back using h (i, j). The tentative end point J of the input voice can be determined by methods such as obtaining a section in which the voice may exist using the amplitude and manually designating the time when the utterance is finished by the user. FIG. 1 shows an example of the time axis correspondence result by DP matching. In the figure, I _S and I _E are the start point and end point of the combined standard pattern, and the positions J _S and J _E corresponding to these are the start point and end point of the word speech. The end points determined in this way are obtained as a result of interpreting the input speech as a noise part preceding a word, a word part consisting of a sequence of phoneme classes following the word, and a noise part following the word. Yes, it is more accurate than the end point determined simply by the magnitude of the amplitude.

There are two possible phoneme class standard patterns: a method using a standard pattern of an unspecified speaker and a method using a standard pattern of a specific speaker. When using the unspecified speaker standard pattern, the phoneme class standard pattern is previously determined by a method of averaging phoneme class patterns extracted from a large number of speaker voice samples or a method of obtaining a plurality of standard patterns by clustering. You can ask for it. When the specific speaker standard pattern is used, it is possible to detect the end points with higher accuracy than the method using the unspecified speaker standard pattern. In this case, since the phoneme class standard pattern cannot be obtained in advance, the speaker phoneme class standard pattern is created by the following procedure prior to the word registration.

First, a basic basic word set is defined in advance. This basic word set should include all phoneme classes, and the standard word patterns of the unspecified speaker corresponding to this should be prepared. The standard pattern of basic words can be created in advance from speech samples of the basic words of many speakers by means of averaging, clustering, or the like.
Here, the position of the phoneme class is specified on the standard pattern of the basic word. FIG. 2 shows an example thereof. The time base correspondence between the basic word voice uttered by the speaker and the standard word standard pattern of the unspecified speaker is executed by DP matching. At this time, as already described, by adding the environmental noise pattern before and after the standard pattern, the detection of the word section becomes reliable. The phoneme class position on the voice uttered by the speaker is determined from the result of the DP matching. The feature vectors at these positions can be the phoneme class standard pattern of the speaker. (When there are a plurality of positions for one phoneme class, methods such as averaging and clustering can be used here as well.) In this description, the phoneme class standard pattern and the environmental noise pattern are both single. Although it is a vector, it can be generally a set of a plurality of vectors. That is, in the calculation of the distance between vectors, a set of standard patterns is given as X (s) = {X ^* ₁ , X ^* ₂ , ,, X ^* _m , ..., X ^* _M } for a certain phoneme class. If there is The vector distance may be calculated as Here, y is the feature vector of the input voice.

(Embodiment) FIG. 3 is a block diagram showing an embodiment of the first invention of the present application. Reference numeral 1 is a voice analysis unit, which converts an input voice signal into a series of feature vectors and outputs the sequence. Here, an arbitrary expression such as a spectrum, a cepstrum, or a bandpass filter is possible as the expression of the feature vector. Reference numeral 2 denotes a changeover switch, which is connected to a terminal A when registering an environmental noise pattern, a terminal C when registering a word for creating a standard pattern, and a terminal D when recognizing.

First, A is selected in the switch 2, and the environmental noise pattern is registered. The feature vector, which is the result of the analysis performed by the voice analysis unit in the silent state, is stored in the environmental noise pattern storage unit 3. A phoneme class standard pattern storage unit 6 stores phoneme class standard patterns created in advance from voice samples of many speakers. A registered word list storage unit 7 stores phoneme class sequence descriptions for each recognition target word. When registering a word, the switch 2 selects C. Here, first, based on the phoneme class sequence of the registered word stored in 7, the phoneme class standard patterns stored in the storage unit 6 are concatenated to create a combined standard pattern, and further, the storage unit 3 The ambient noise pattern read from the above is added before and after the combined standard pattern to create an expanded combined standard pattern, which is stored in the buffer 8. The processing unit 9 executes DP matching between the extended combined standard pattern read from the buffer 8 and the registered word voice sent from the voice analysis unit 1 via the switch 2C, obtains the time axis correspondence, and inputs. The start point and the end point of the voice are determined, and the feature vector series in the section is stored in the storage unit 10 as a standard pattern.

At the time of recognition, D is selected by the switch 2 and the storage unit
The recognition processing is performed in the recognition unit 11 using the standard pattern read from 10.

FIG. 4 is a block diagram showing an embodiment of the second invention of the present application. Reference numeral 21 is a voice analysis unit, which converts the input voice signal into a series of feature vectors and outputs the sequence. Here, an arbitrary expression such as a spectrum, a cepstrum, or a bandpass filter is possible as the expression of the feature vector. Reference numeral 22 denotes a changeover switch. When registering an environmental noise pattern, terminal A is used. When a basic word is input to generate a phoneme standard pattern, terminal B is used. When a standard pattern is created, terminal C is used. , At the time of recognition, it is connected to the terminal D.

First, A is selected by the switch 22 and the environmental noise pattern is registered. A feature vector that is input in a silent state and analyzed by the voice analysis unit 21 is stored in the environmental noise pattern storage unit 23. Reference numeral 24 is a storage unit that holds a standard pattern of basic words of unspecified speakers, and this standard pattern is prepared in advance from a sample of many speakers. Then, B is selected by the switch 22 and a phoneme class standard pattern is generated. Reference numeral 25 denotes a phoneme class standard pattern extraction unit, which is provided before and after the basic word standard pattern of the unspecified speaker read from the storage unit 24.
DP matching of the pattern added with the environmental noise pattern read from 23 and the basic word voice sent from the voice analysis unit 21 via the switch 22B is executed.
The phoneme class position on the input speech is determined from the correspondence of the time axes of the two patterns obtained as a result, and the feature vector on the position is stored in the storage unit 26 as a phoneme class standard pattern. A registered word list storage unit 27 stores a phoneme class sequence description for each recognition target vocabulary.
When registering a word, the switch 22 selects C. Here, first, based on the phoneme class sequence of the registered word stored in 27, a phoneme class standard pattern stored in the storage unit 26 is connected to create a combined standard pattern, and further, the storage unit 23 The environmental noise pattern read from the above is added before and after the combined standard pattern to create an expanded combined standard pattern, which is stored in the buffer 28. Processing unit 29
Is the extended combined standard pattern read from buffer 28,
DP matching is performed with the registered word voice sent from the voice analysis unit 21 via the switch 22C, the time axis correspondence is obtained, the start point and end point of the input voice are determined, and the feature vector series in the section is standardized. Is stored in the storage unit 30 as

At the time of recognition, D is selected by the switch 22 and the storage unit
The recognition processing is performed in the recognition unit 31 using the standard pattern read from 30.

In the above, the recurrence formula in DP matching is not limited to the one described in the present invention. Also, the recognition unit 1
As for 1 and 31, any method other than DP matching can be used.

(Effect of the invention) As described above, according to the present invention, since the end point of the word voice is detected in consideration of the phonetic structure of the word to be registered, more accurate word detection is realized. Thus, it becomes possible to realize voice recognition with high recognition accuracy.

[Brief description of drawings]

FIG. 1 is a conceptual diagram showing an example of time axis correspondence by DP matching, FIG. 2 is a conceptual diagram showing an example of a word standard pattern of an unspecified speaker, and FIGS. 3 and 4 are implementations of the present invention. It is a block diagram which shows an example. In these figures, reference numerals 1 and 21 are voice analysis units and 2,
22 is a changeover switch, 3 and 23 are environmental noise pattern storage units, 24 is an unspecified speaker word standard pattern storage unit, 25 is a phoneme class standard pattern extraction unit, 6 and 26 are phoneme class standard pattern storage units, 7 and 27. Is a registered word list storage unit, 8 and 28 are combined standard pattern buffers, 9 and 29 are processing units (word standard pattern registration units), 10 and 30 are word standard pattern storage units, and 11 and 31 are recognition units.

Claims (2)

[Claims] 1. A registered word expressed as a phoneme class sequence, having means for storing an environmental noise pattern and a phoneme class standard pattern of an unspecified speaker when registering a registered voice uttered by a user as a standard pattern. On the other hand, a temporary word standard pattern is generated as a combined standard pattern by connecting the phoneme class standard patterns, and the time axis correspondence between the pattern in which the environmental noise pattern is connected before and after the combined standard pattern and the registered voice is generated. Match by attaching,
A word registration method in word voice recognition, characterized in that an end point of a registered voice is detected by extracting a position on the registered voice corresponding to a boundary position between a combined standard pattern and an environmental noise pattern. 2. When registering a registered voice uttered by a user as a standard pattern, a means is provided for storing a standard pattern of a basic word of an unspecified speaker, to which information on environmental noise patterns and phoneme class positions is added. , By matching the basic word voice uttered by the user in advance and the pattern in which the environmental noise pattern is connected before and after the basic word standard pattern by time-axis correspondence to correspond to the phoneme class position of the basic word standard pattern. A phoneme class standard pattern is obtained by extracting the phoneme class position of the basic word speech, and the phoneme class standard pattern is concatenated to the registered words expressed as a phoneme class sequence to form a temporary standard pattern as a combined standard pattern. Is generated, and the time axis correspondence between the pattern in which the environmental noise pattern is connected before and after the combined standard pattern and the registered voice is generated. A word registration method in word voice recognition characterized by detecting the end points of the registered voice by extracting the position on the registered voice corresponding to the boundary position between the combined standard pattern and the environmental noise standard pattern by performing matching by .