JPH11338492A - Speaker recognition unit - Google Patents

Abstract

PROBLEM TO BE SOLVED: To enhance a speaker recognition performance for a speaking content assigning type. SOLUTION: A speaker's character weighting factor addition part 1-5 and a weighting similarity calculating part 1-8 are provided in a text assigned type speaker recognition unit. A distance between a speaker of a recognition object and many-and-unspecified speakers is compared to add a weighting factor proportional to the largeness of the distance to a phoneme model, assuming that a model having a large value thereof shows a speaker's character intensively. The weighted phoneme models are connected at the recognition to prepare a utterance-assigned word model, and similarity between the model and a characteristic vector time series of an speaking voice is calculated by the similarity calculating part 1-8 to determine a result of speaker recognition based on the calculated value.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an acoustic analysis of a voice uttered by a speaker, comparing the voice feature pattern with a speaker feature pattern registered in advance, and determining the speaker from the result. The present invention also relates to a speaker recognition device for determining whether or not a speaker is the claimed person.

[0002]

2. Description of the Related Art In speaker recognition, a text-dependent system in which a voice pattern in which a specific speaker utters a password is held as a standard pattern is generally used. In this method, at the time of recognition, the distance between the input voice and the user's standard pattern after time axis matching is calculated, and the value is compared with a certain threshold to determine whether or not the user is the user. However, in the text-dependent type, if a person's voice is recorded by a tape recorder or the like and is reproduced in front of the recognition device, there is a risk that the device will accept it.

[0003] In recent years, there has been proposed a text designation type speaker recognition system in which an arbitrary text can be designated from the device every time recognition is performed. In this method, only when the user can determine that the specified text has been correctly uttered, the text is accepted. This makes it possible to prevent spoofing due to the recorded voice since the utterance content is different each time.

In the text designation type, in order to determine whether or not the user has uttered the correctly specified text, it is necessary to generate a model of the voice when the user utters the text. In order for the system side to be able to specify an arbitrary text, all basic phonemic models are learned for each speaker, and connected to generate a speech model of the specified text.

[0005] To create a phoneme model for each speaker, learning is performed by directly extracting HMM (Hidden Markov Model) parameters of each phoneme model from speech data (hundreds of sentences) produced by one speaker in a very large amount. This is ideal, but it is practically impossible to force the user to make such a large amount of utterances. Therefore, an unspecified speaker phoneme model created from data of many speakers is used as an initial model, and 100
Using the word-level training data, the initial model is adapted to the speaker. In this way, a speaker phonemic model is generated.

[0006]

Generally, the characteristics of a speaker's voice do not appear in all voice sections, but, for example, as partial features of the voice section such as the average pitch of vowels and the nasal spectrum. Observed. However, in the conventional method, all phonemes of a registered speaker are evaluated equally to calculate the similarity of words, and the calculated similarity is used for speaker recognition. However, according to this method, a portion which expresses the above-mentioned speaker characteristics well and a portion which does not express the same are evaluated equally, so that good speaker recognition performance is not always obtained.

Accordingly, it is an object of the present invention to provide a speaker recognition apparatus capable of improving speaker recognition by largely reflecting a characteristic utterance of a registered speaker in a recognition result.

[0008]

SUMMARY OF THE INVENTION The present invention is a text designation type speaker recognition device which determines whether or not a speaker is himself / herself by uttering a designated word every time.
A speech input unit that captures a speaker's utterance, a feature vector calculation unit that calculates a feature vector sequence of a speech signal, an unspecified speaker phoneme model storage unit that stores a phoneme model of an unspecified speaker, A speaker adaptation unit that adapts the phoneme model of the unspecified speaker to the registered speaker during the registration operation, a registered speaker phoneme model storage unit that stores the phoneme model of the registered speaker, and a specified text during the speaker recognition operation A specified word model synthesizing unit for synthesizing the word model, a recognition result judging unit for judging the speaker recognition result by comparing the similarity between the speaker voice and the word model with a judgment threshold value, And a recognition result output unit that outputs a recognition result output unit that outputs an unspecified speaker phoneme model to a registered speaker. Weighted coefficients to the phonological model. A weighting similarity that calculates a similarity between the feature vector sequence of the input speech and the word model as a cumulative distance obtained by accumulating a value obtained by weighting a partial distance of each phoneme model unit in a time-series direction. A speaker recognition device including a calculation unit.

With this configuration, it is possible to realize a speaker recognition apparatus capable of improving speaker recognition by largely reflecting the characteristic utterance of the registered speaker in the recognition result.

[0010]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The invention according to claim 1 is a text designation type speaker recognition device which determines whether or not a speaker is himself / herself by uttering a designated word every time. A speech input unit that captures a speaker's utterance, a feature vector calculation unit that calculates a feature vector sequence of a speech signal, an unspecified speaker phoneme model storage unit that stores a phoneme model of an unspecified speaker, A speaker adaptation unit that adapts the phoneme model of the unspecified speaker to the registered speaker during the registration operation, a registered speaker phoneme model storage unit that stores the phoneme model of the registered speaker, and a specified text during the speaker recognition operation A specified word model synthesizing unit for synthesizing the word model, a recognition result judging unit for judging the speaker recognition result by comparing the similarity between the speaker voice and the word model with a judgment threshold value, Text with a recognition result output unit to output A talk in which a weighting coefficient proportional to the moving amount of the average feature vector when the unspecified speaker phoneme model is adapted to the registered speaker by the speaker adaptation unit is added to the designated speaker recognition device. The user feature weighting coefficient adding unit, and the similarity between the feature vector sequence of the input speech and the word model,
A speaker recognition device comprising a weighted similarity calculating unit that calculates a value weighted to a partial distance in phoneme model units as a cumulative distance accumulated in a time-series direction.

In this configuration, a weighting factor proportional to the moving distance of the average feature vector at the time of speaker adaptation is added to the phoneme model and stored, and at the time of speaker recognition, the weighted phoneme model is connected to the designated text. Create a word model of
In the similarity calculation, the partial distance of each phoneme model is weighted by the above-mentioned weighting coefficient, and the cumulative distance added over the vocal section is compared with the judgment threshold value of the registered speaker. Determine the recognition result. As a result, the similarity of the characteristic utterance part of the registered speaker is strongly reflected in the speaker recognition, so that improvement in the speaker recognition accuracy can be expected.

According to the second aspect of the present invention, when the voice data of a specific speaker is sufficiently obtained and the phoneme model of the specific speaker can be directly created without adapting the phoneme data of the unspecified speaker,
Alternatively, when the phoneme model of the registered speaker is prepared in advance, the distance between the average feature vector of the same phoneme model of the unspecified speaker and the registered speaker is calculated, and the weighting coefficient proportional to the value is calculated as the phoneme model. Is provided with a speaker characteristic weighting coefficient adding unit to be added.

With this configuration, when sufficient speech data of a specific speaker can be obtained and a phoneme model of a specific speaker can be directly created without adapting phoneme data of an unspecified speaker, unspecified / registered speech The distance between the average feature vectors is calculated using the phoneme model of the same phoneme of the speaker, and a weight coefficient is added to the phoneme model of the registered speaker. Then, at the time of speaker recognition, a weighted similarity is calculated. Thus, even when speaker adaptation is not performed, a weight coefficient can be added to a characteristic phoneme model of a registered speaker.

According to a third aspect of the present invention, when a word model is used in a text-dependent manner, correspondence between the same phoneme portion of the word model of the unspecified speaker and that of the specific speaker is obtained, and the distance between the individual average vectors is obtained. , A speaker characteristic weighting coefficient adding unit for partially adding weighting to the word model in proportion to the value.

With this configuration, when a word model is used for speaker recognition, time axis matching is performed between the word model of an unspecified speaker and the word model of a specific speaker, and the average vector of the corresponding phoneme portion is obtained. The distance is calculated, and a weight coefficient proportional to the distance is added to the word model in phoneme units. As a result, a word model similar to the case where the weighted phoneme models are connected is obtained. By calculating the weighted similarity at the time of speaker recognition, speaker recognition emphasizing the characteristic phoneme of the registered speaker can be realized by the conventional fixed text speaker recognition device.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a speaker recognition apparatus according to an embodiment of the present invention, FIG. 2 is a circuit block diagram of the speaker recognition apparatus, and FIG. 3 is a conceptual diagram of speaker feature weighting of the speaker recognition apparatus. FIG. 4 is a flowchart at the time of the speaker registration operation of the speaker recognition device, and FIG. 5 is a flowchart at the time of the speaker recognition operation of the speaker recognition device.

In FIG. 1, reference numeral 1-1 denotes a voice input unit for taking in a utterance of a speaker, 1-2 denotes a feature vector calculation unit for calculating a feature vector sequence of a voice signal, and 1-3 denotes an unspecified speech. Is a speaker-independent speaker model storage unit that stores a phoneme model of a speaker. 1-4 is a speaker adaptation unit that adapts the phoneme model of the speaker-independent speaker to a registered speaker during a speaker registration operation. There is a speaker characteristic weight 1-5 for adding a weighting factor proportional to the moving amount of the average feature vector when the speaker-independent adaptation unit adapts the unspecified speaker phoneme model to the registered speaker. A coefficient adding unit 1-6 is a registered speaker phoneme model storage unit for storing an adapted registered speaker phoneme model, and 1-7 is a designation for synthesizing a word model of a designated text at the time of speaker recognition operation. A word model synthesizing unit 1-8 is a feature vector of the input speech. The degree of similarity between the Le series and the word model,
A weighted similarity calculation unit that calculates a value weighted to a partial distance in phoneme model units as a cumulative distance accumulated in a time-series direction, and 1-9 denotes a similarity between a speaker's voice and a word model as a determination threshold. And a recognition result output unit 1-10 for outputting the speaker recognition result.

FIG. 2 is a circuit block diagram of the speech recognition apparatus, wherein 2-1 is a microphone, 2-2 is a speaker, 2-3 is a central processing unit (CPU), 2-4 is a writable memory (RAM), 2-5 is a read-only memory (ROM).

In the block diagram of FIG. 1, the voice input unit 1-1 is provided by the microphone 2-1; the unspecified speaker phoneme model storage unit 1-3 is provided by the ROM 2-5; and the registered speaker phoneme model storage unit 1-6. Is a speaker adaptation unit 1-
4, a speaker characteristic weighting coefficient adding unit 1-5, a designated word model synthesizing unit 1-7, a weighting similarity calculating unit 1-8, and a recognition result determining unit 1-9. By executing the program and accessing the ROM 2-5 and the RAM 2-4, the recognition result determination unit 1-9 is executed by the speaker 2-2 or other output means.

The operation of the speaker recognition device configured as described above will be described separately for a speaker registration operation and a speaker recognition process.

1. The operation of creating a weighted phoneme model of a registered speaker during the speaker registration operation will be described with reference to the flowchart of FIG. This process is executed when the speaker to be recognized registers his / her own voice pattern in the apparatus.

Step (1): A voice signal uttered by the user is fetched from the voice input unit 1-1.

Step (2): The speech signal is converted into a digital signal by the feature vector calculation section 1-2, and a speech feature vector is calculated.

Step (3): Adapts the phoneme model of the unspecified speaker to the specific speaker (user) using the user's speech feature vector. Specifically, the speaker adaptation unit 1-4
In the above, registration is performed by moving the average feature vector of the phoneme model of the unspecified speaker stored in the unspecified speaker phoneme model storage unit 1-3 in the direction of the average feature vector of the same phoneme of the registered speaker. Create a speaker phonemic model.

Step (4): The moving distance of the average feature vector when the phoneme model of the unspecified speaker is adapted to the registered speaker is calculated by the speaker characteristic weighting coefficient calculating unit.

Step (5): The speaker characteristic weighting coefficient adding unit 1-5 adds a distance-proportional weighting coefficient obtained by multiplying the moving distance by a certain number to the phoneme model of the registered speaker.

Step (6): The phoneme model to which the weight coefficient has been added is recorded in the registered speaker phoneme model storage unit 1-6.

According to the second aspect, when the speaker adaptation is not performed and the recognition model of the specific speaker is directly obtained, the registered speaker and the unspecified speaker are identical without using the vector moving distance at the time of speaker adaptation. A vector distance between the average feature vectors of the phoneme model is calculated, and a weight coefficient proportional to the value is calculated. Then, the phoneme model to which the weight coefficient has been added is recorded in the registered speaker phoneme model storage unit 1-6.

According to the third aspect, when a speech model in word units is obtained from the beginning without using a recognition model in phoneme units,
The time axis of both word models is normalized, the distance between corresponding average feature vectors of the same phoneme portion is calculated, and a weighting coefficient proportional to the value is added to each phoneme section of the registered speaker's word model.

2. Speaker recognition processing The operation at the time of speaker recognition processing will be described with reference to the flowchart of FIG.

Step (1): The device designates the utterance word to the speaker. Step (2): A speaker-weighted phoneme model is connected in accordance with a dictionary to create a word model of a designated word.

Step (3): The voice signal of the word uttered by the speaker is fetched from the voice input section 1-1.

Step (4): feature vector calculation unit 1
2, a feature vector time series of the input voice is calculated.

Step (5): Weighted similarity calculating section 1
According to −8, time axis matching is performed between the time series of the speech feature vector and the word model, and the correspondence between the time series of the feature vector and each phoneme of the word model is obtained.

Step (6): A weighted cumulative distance is calculated in accordance with the correspondence obtained in step (5).

More specifically, the distance between the time series of the speech feature vector and the corresponding partial phoneme model between the word models is calculated. Then, the vector distance in phoneme units is accumulated over the utterance section. At this time, the speaker characteristics are emphasized by multiplying the distance by the weight coefficient of the phoneme model.

Cumulative distance = (distance per phoneme × weight coefficient)
The cumulative distance is normalized by dividing the above cumulative distance by the sum of the values obtained by multiplying the weighting coefficient of the phoneme model by the number of appearances, and the cumulative distance is normalized, and the judgment threshold of the recognition result is weighted. Can be used as is regardless of the presence or absence of.

Normalized cumulative distance = cumulative distance ÷ (sum of utterance sections of (weighting coefficient × cumulative number of phoneme models)) By this weighting process, the phoneme portion that well reflects the speaker characteristics of the speaker is emphasized. Is done. As a result, the cumulative distance when the speaker is the person is not much affected, while the cumulative distance when the speaker is the impostor is increased, the contrast between the person and the impostor becomes clearer, and the speaker recognition accuracy is improved. Can be expected to improve.

When a discrete distribution HMM (Hidden Markov Model) is used as the phoneme model, a weighting coefficient is calculated from the average vector distance of each state of the phoneme model before and after the adaptation, and is added to the model. The distance between the HMM models is k
A probabilistic distance measure such as ullback may be used.
The weight coefficient Zj of the state j is based on (Equation 1).

[0040]

(Equation 1)

At the time of speaker recognition, an optimal state transition sequence is obtained by the Viterbi algorithm while ignoring the weighting factor, and the output likelihood on the path is weighted and accumulated. For the time series y1T of the feature vectors, the weighted cumulative likelihood P (y) is given by (Equation 2). To normalize the cumulative likelihood, the duration of each state of the optimal state transition sequence is divided by the product sum of the weighting factor.

[0042]

(Equation 2)

As described above, the normalized weighted likelihood in the case of the HMM phoneme model is obtained. Step (7): The recognition result determination unit 1-9 compares the accumulated distance with a threshold value for determining whether or not the speaker is the person,
Determine the recognition result.

[0044]

According to the present invention, the inter-speaker distance of a phoneme that expresses the speech characteristics of a registered speaker well is emphasized.
Improvement of recognition accuracy of registered speakers can be expected. Since no prior phonetic knowledge about the characteristics of the speaker is used in the weighting process, it is possible to flexibly cope with individual differences between speakers.

According to the second aspect of the present invention, when a phoneme model can be directly created from the speech data of a specific speaker, the characteristics of the speaker are analyzed by comparison with a general (unspecified) speaker, and the phoneme model is converted to a phoneme model. By performing weighting, speaker recognition can be performed with higher accuracy.

According to the third aspect of the present invention, even in the conventional general text-dependent speaker recognition apparatus, the word models between the registered speaker and the general (unspecified) speaker are compared, and the characteristic of the speaker is determined. By adding weights to the appropriate phonemes, it is possible to perform high-accuracy speaker recognition in which speaker characteristics are emphasized.

[Brief description of the drawings]

FIG. 1 is a configuration block diagram of a speaker recognition device according to an embodiment of the present invention.

FIG. 2 is a circuit block diagram of a speaker recognition device according to an embodiment of the present invention.

FIG. 3 is a conceptual diagram of speaker feature weighting of the speaker recognition device in one embodiment of the present invention.

FIG. 4 is a flowchart at the time of a speaker registration operation of the speaker recognition device according to the embodiment of the present invention;

FIG. 5 is a flowchart at the time of a speaker recognition operation of the speaker recognition device according to the embodiment of the present invention;

[Explanation of symbols]

 1-1 Speech input unit 1-2 Feature vector calculation unit 1-3 Unspecified speaker phoneme model storage unit 1-4 Speaker adaptation unit 1-5 Speaker feature weight coefficient addition unit 1-6 Registered speaker phoneme model Storage unit 1-7 Designated word model synthesis unit 1-8 Weighted similarity calculation unit 1-9 Recognition result determination unit 1-10 Recognition result output unit 2-1 Microphone 2-2 Speaker 2-3 Central processing unit (CPU) 2 -4 Writable memory (RAM) 2-5 Read-only memory (ROM)

Claims (3)

[Claims] 1. A text designation type speaker recognition device for determining whether or not a speaker is himself / herself by uttering a designated word every time, and a voice input unit for taking in a utterance of the speaker. A feature vector calculating unit for calculating a feature vector sequence of a voice signal, an unspecified speaker phoneme model storage unit storing a phoneme model of an unspecified speaker, and a phoneme model of an unspecified speaker during a speaker registration operation. Speaker adaptation unit that adapts a registered speaker, registered speaker phoneme model storage unit that stores registered speaker phoneme models, and specified word model synthesis that synthesizes the word model of the specified text during speaker recognition operation A recognition result determination unit that determines a speaker recognition result by comparing a similarity between a speaker voice and a word model with a determination threshold value, and a recognition result output unit that outputs a speaker recognition result. In the text-specified speaker recognition device, A speaker feature weighting factor adding unit that adds a weighting factor proportional to the amount of movement of the average feature vector when the unspecified speaker phonemic model is adapted to the registered speaker by the adapting unit; A speaker characterized by comprising a weighted similarity calculation unit that calculates a similarity between a feature vector sequence and a word model as a cumulative distance obtained by accumulating a value obtained by weighting a partial distance in a unit of a phoneme model in a time-series direction. Recognition device. 2. A method in which sufficient speech data of a specific speaker is obtained and a phoneme model of a specific speaker can be directly created without adapting phoneme data of an unspecified speaker, or a phoneme model of a registered speaker. Is prepared in advance, the distance between the average feature vector of the same phoneme model of the unspecified speaker and the registered speaker is calculated, and the speaker feature weight for adding a weighting coefficient proportional to the value to the phoneme model The speaker recognition device according to claim 1, further comprising a coefficient adding unit. 3. In the case where a word model is used in a text-dependent type, the correspondence between the same phoneme portion of the word model of the unspecified speaker and the word model of the specific speaker is determined, and the distance between the respective average vectors is obtained. 2. The speaker recognition apparatus according to claim 1, further comprising a speaker characteristic weighting coefficient adding unit for partially adding weight to the word model.