JP2017097188A - Speaker-likeness evaluation device, speaker identification device, speaker collation device, speaker-likeness evaluation method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speaker-likeness evaluation device capable of calculating the similarity in speaker's feature vector capable of properly recognizing a speaker even when the length between a registration speech and recognition speech is remarkably different from each other.SOLUTION: The speaker-likeness evaluation device includes: a divided registered speech's speaker feature vector calculation unit 1001 that generates a divided registered speech which has a segment which is shorter than a segment of a registered speech from a registered speech and calculates a divided registered speech's speaker feature vector from the divided registered speech; a recognized speech speaker's feature vector calculation section 1002 that calculates a recognized speech speaker feature vector from a recognized speech; and a similarity calculation section 150 that calculates the similarity between the recognized speech and the divided registered speech by using the recognized speech speaker feature vector and the divided registered speech speaker feature vector.SELECTED DRAWING: Figure 10

Description

  The present invention relates to a speaker recognition technique using voice, and more particularly to a technique for speaker recognition using the similarity of speaker feature vectors.

  Speaker recognition by voice (hereinafter simply referred to as speaker recognition) is roughly divided into speaker identification and speaker verification. Speaker identification is a technique for determining who corresponds to a pre-registered speaker for an input voice, and is used to search for a criminal voice from voice recording, for example. On the other hand, speaker verification is a technique for determining whether or not a speaker is a pre-registered speaker for input speech, and is used, for example, for identity verification. In some cases, both speaker verification and speaker identification may be provided, and it is determined whether the input speech is included in the pre-registered speaker (verification), and if it is included, which registered speaker is determined (identification). Such usage is also conceivable. In either case, it is necessary to pre-register voices of more than one utterance in the system for each speaker.

  There are text-dependent and text-independent types for speaker recognition. In the text dependent type, the user needs to utter a predetermined word upon recognition. On the other hand, in the text-independent type, the user may utter arbitrary words during recognition.

  In speaker recognition, a technique for calculating one speaker feature vector from the entire input speech signal is used. The method for obtaining the speaker feature vector is shown below. The input audio signal is divided into acoustic analysis frames of several tens of milliseconds, and an acoustic feature vector sequence in which the acoustic feature vectors extracted from each acoustic analysis frame are arranged in time order is created. A speaker feature extraction model that has been learned in advance from an acoustic feature vector sequence, a zero-order statistic calculated for a predetermined mixed normal distribution (UBM model), and a first-order statistic are used for one story. A person feature vector is obtained. These procedures are disclosed in Non-Patent Document 1, for example. The speaker feature vector is easy to learn because it does not require a label indicating the speaker for each speech signal during pre-learning of the speaker feature extraction model. Since the speaker feature vector is calculated, there is an advantage that it can be used for text-independent speaker recognition with different types and lengths of words, and it is widely used in speaker recognition.

  If speaker feature vectors can be calculated from individual input speech, speaker recognition can be easily realized using existing classification techniques and outlier detection techniques. For example, speaker identification is performed by obtaining the cosine similarity between the registered speaker feature vector and the speaker feature vector of the input speech, and returning the speaker name of the registered speaker feature vector having the maximum similarity. realizable. Speaker verification is achieved by determining the cosine similarity between the registered speaker feature vector and the speaker feature vector of the input speech, and determining that the speaker is a registered speaker if the maximum similarity is greater than or equal to a threshold value. it can.

  Hereinafter, an outline of a conventional speaker identification apparatus will be described with reference to FIGS. FIG. 1 is a block diagram showing a configuration of a conventional speaker identification device 800. FIG. 2 is a flowchart showing the operation of the conventional speaker identification device 800. FIG. 3 is a diagram illustrating an example of the estimated speaker identification of the identification utterance by the speaker identification device 800 of the related art. As shown in FIG. 1, the speaker identification device 800 includes an acoustic analysis unit 820-1, a speaker feature vector calculation unit 830-1, a speaker registration unit 840, a registered utterance recording unit 803, and an acoustic analysis unit 820. -2, speaker feature vector calculation unit 830-2, similarity calculation unit 850, and speaker identification unit 870. The speaker identification device 800 is connected to a UBM model recording unit 801 and a speaker feature amount extraction model recording unit 802.

  The i-vector described in Non-Patent Document 1 is used as the speaker feature vector. As described in Non-Patent Document 1, the i-vector calculation requires a UBM model and a speaker feature extraction model. These models are learned in advance, and the UBM model recording unit 801, It is recorded in the person feature quantity extraction model recording unit 802. Note that the UBM model and the speaker feature extraction model correspond to T and Σ of Non-Patent Document 1, respectively.

  First, a process for recording a registered utterance in advance will be described. The registered utterance is an audio signal of the utterance of the speaker that is desired to be registered in advance, and is used to identify the speaker of the identification utterance. It is assumed that there are one or more registered utterances, and the registered utterances are collectively referred to as a registered utterance set. The acoustic analysis unit 820-1 calculates an acoustic feature vector sequence of the registered utterance from the registered utterance (S820-1). As described above, the acoustic feature vector sequence may be calculated by dividing a registered utterance into acoustic analysis frames of several tens of milliseconds and arranging acoustic feature vectors extracted from each acoustic analysis frame in time order. The speaker feature vector calculation unit 830-1 records the registered utterance acoustic feature vector series, which is the output of the acoustic analysis unit 820-1, in the UBM model recording unit 801 and the speaker feature extraction model recording unit 802. A speaker feature vector of the registered utterance is calculated using the placed UBM model and speaker feature extraction model (S830-1). The procedure for calculating the speaker feature vector is as described above. The speaker registration unit 840 sets the registered utterance of the registered utterance recording unit 803 by combining the speaker feature vector of the registered utterance that is the output of the speaker feature vector calculating unit 830-1 and the speaker name corresponding to the registered utterance. Register in the database (S840). The speaker name is information for specifying the speaker of the previous registered utterance, and is given manually.

Next, a process for identifying an identification utterance will be described. An identification utterance is an audio signal of an utterance of a speaker to be identified. The acoustic analysis unit 820-2 calculates an acoustic feature vector sequence of the identified utterance from the identified utterance (S820-2). The calculation procedure of the acoustic feature vector sequence is the same as that in S820-1. The speaker feature vector calculation unit 830-2 records the acoustic feature vector sequence of the identified utterance, which is the output of the acoustic analysis unit 820-2, in the UBM model recording unit 801 and the speaker feature extraction model recording unit 802. The speaker feature vector of the identified utterance is calculated using the placed UBM model and speaker feature extraction model (S830-2). The procedure for calculating the speaker feature vector is the same as that in S830-1. The similarity calculator 850 calculates the similarity between the speaker feature vector of the identified utterance that is the output of the speaker feature vector calculator 830-2 and the speaker feature vector of each registered utterance in the registered utterance database (S850). ). For example, the cosine similarity described in Non-Patent Document 1 may be used for calculating the similarity. If w ₁ and w ₂ are the speaker feature vector of the identified utterance and the speaker feature vector of the registered utterance, respectively, the cosine similarity c is given by the following equation.

Here, ^t represents transposition. In other words, the similarity is an index indicating the likelihood of a speaker, and when the cosine similarity is adopted as the similarity, the higher the similarity, the more likely the speaker to be compared (similar to the speaker) I can say that. The similarity calculation unit 850 outputs the calculated similarity and the registered utterance speaker name as a set. The speaker identifying unit 870 selects the speaker name of the registered utterance as the maximum similarity among the similarities output from the similarity calculating unit 850, and outputs the selected speaker name as an identification result (S870).

  Next, an outline of a conventional speaker verification apparatus will be described with reference to FIGS. In addition, the same number is attached | subjected to the component which has the same function as the component of the speaker identification apparatus 800, and duplication description is abbreviate | omitted. FIG. 4 is a block diagram showing a configuration of a conventional speaker verification apparatus 900. FIG. 5 is a flowchart showing the operation of the conventional speaker verification apparatus 900. FIG. 6 is a diagram illustrating an example of collation of collation utterances performed by a conventional speaker collation apparatus 900. As shown in FIG. 4, the speaker verification device 900 includes an acoustic analysis unit 820-1, a speaker feature vector calculation unit 830-1, a speaker registration unit 940, a registered utterance recording unit 903, and an acoustic analysis unit 820. -2, speaker feature vector calculation unit 830-2, similarity calculation unit 950, and speaker verification unit 970. Similar to the speaker identification device 800, the speaker verification device 900 is connected to the UBM model recording unit 801 and the speaker feature amount extraction model recording unit 802.

  First, a process for recording a registered utterance in advance will be described. The acoustic analysis unit 820-1 and the speaker feature vector calculation unit 830-1 perform processing similar to that of the speaker identification device 800 (S820-1, S830-1). The speaker registration unit 940 registers the speaker feature vector of the registered utterance, which is the output of the speaker feature vector calculation unit 830-1, in the registered utterance database of the registered utterance recording unit 903 (S940).

  Next, a process for collating collation utterances will be described. The collation utterance is an audio signal of the utterance of the speaker to be collated, and is collated with the registered utterance. The acoustic analysis unit 820-2 and the speaker feature vector calculation unit 830-2 also perform processing similar to that of the speaker identification device 800 (S820-2, S830-2). The similarity calculation unit 950 calculates the similarity between the speaker feature vector of the collation utterance output from the speaker feature vector calculation unit 830-2 and the speaker feature vector of each registered utterance in the registered utterance database (S950). ). Unlike the similarity calculation unit 850, the similarity calculation unit 950 outputs only the similarity. The speaker verification unit 970 is a registered utterance speaker when the maximum similarity among the similarities output from the similarity calculation unit 950 is greater than a threshold (or greater than or equal to the threshold). A verification result is generated and output (S970). Alternatively, each similarity degree that is the output of the similarity degree calculation unit 950 is compared with a threshold value, and if there is one that is larger than the threshold value (greater than or equal to the threshold value), a matching result indicating that the speaker is a registered utterance is obtained. It may be generated and output (S970).

Tetsuji Ogawa, Sayaka Shioda, “Speaker Recognition Using i-vector”, Journal of the Acoustical Society of Japan, June 2014, Volume 70, No. 6, pp.332-339.

  It is known that the speaker feature vector has the property that the speaker information appears stronger as the utterance length becomes longer. Due to this property, the similarity of the speaker feature vectors tends to decrease as the utterance length is different even for speech from the same speaker.

  In speaker recognition, the registered utterance is text reading, while the recognition utterance (identification utterance or collation utterance) is often word reading. For this reason, while the registered utterance becomes longer, for example, 10 seconds or more, the recognized utterance becomes shorter, for example, 1.5 seconds or less, and as a result, the recognized utterance becomes extremely short compared to the registered utterance. For this reason, the similarity between the speaker feature vectors of both utterances decreases due to the difference in utterance length. When speaker recognition is performed using such speaker feature vectors, since the similarity is low for any speaker, the difference in similarity for each speaker is difficult to appear, and the recognition accuracy is reduced.

  Therefore, the present invention provides a speaker likeness evaluation device that calculates the similarity of speaker feature vectors for robustly recognizing a speaker even when the lengths of a registered utterance and a recognized utterance are extremely different. For the purpose.

  In one embodiment of the present invention, an utterance voice of a speaker that is a target of speaker identification or speaker verification is a recognition utterance, and an utterance voice used for recognizing the speaker is a registered utterance. A divided registered utterer feature vector calculation unit that generates a divided registered utterance having a section shorter than the divided section and calculates a divided registered utterer feature vector from the divided registered utterance; and a recognition utterance from the recognized utterance A recognition speaker feature vector calculation unit for calculating a speaker feature vector; and a similarity between the recognition utterance and the divided registered utterance using the recognized utterance feature vector and the divided registered utterance feature vector. It is a speaker likeness evaluation apparatus which has a similarity calculation part to calculate.

  According to the present invention, it is possible to calculate the similarity of speaker feature vectors for accurately recognizing a speaker even when the length of a registered utterance and a recognized utterance are extremely different.

The block diagram which shows the structure of the speaker identification apparatus 800 of a prior art. The flowchart which shows operation | movement of the speaker identification apparatus 800 of a prior art. The figure which shows the example of the presumed speaker identification of the identification utterance by the speaker identification apparatus 800 of a prior art. The block diagram which shows the structure of the speaker verification apparatus 900 of a prior art. The flowchart which shows operation | movement of the speaker verification apparatus 900 of a prior art. The figure which shows the example of collation of collation utterance by the speaker collation apparatus 900 of a prior art. 1 is a block diagram illustrating a configuration of a speaker identification device 100 according to a first embodiment. 3 is a flowchart illustrating detailed operations of the speaker identification device 100 according to the first embodiment. The figure which shows the example of the presumed speaker identification of the identification utterance by the speaker identification device 100 of Example 1. FIG. FIG. 4 is a block diagram illustrating a configuration of a speaker identification device 200 according to a second embodiment. 9 is a flowchart showing detailed operations of the speaker identification device 200 according to the second embodiment. FIG. 6 is a block diagram illustrating a configuration of a speaker identification device 300 according to a third embodiment. 10 is a flowchart illustrating detailed operations of the speaker identification device 300 according to the third embodiment. The block diagram which shows the structure of the speaker identification apparatus 300 'of Example 3 (modification). FIG. 9 is a block diagram illustrating a configuration of a speaker verification device 400 according to a fourth embodiment. 10 is a flowchart showing a detailed operation of the speaker verification device 400 according to the fourth embodiment. The figure which shows the example of collation of collation utterance by the speaker collation apparatus 400 of Example 4. FIG. FIG. 10 is a block diagram illustrating a configuration of a speaker verification device 500 according to a fifth embodiment. 10 is a flowchart showing a detailed operation of the speaker verification device 500 according to the fifth embodiment.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the structure part which has the same function, and duplication description is abbreviate | omitted.

  Also, in the following description, an identification utterance or a collation utterance is a recognition utterance, and a set of registered utterances having one or more registered utterances as elements is called a registered utterance set.

<Key points of the present invention>
It is known that the speaker feature vector has not only the characteristic that the speaker information becomes stronger as the utterance length becomes longer, but also the characteristic that it depends not only on the speaker information but also on the word information included in the utterance. . Because of this property, even if the speech length is the same, the similarity of the speaker feature vectors tends to decrease as the words included in the speech differ. For example, the similarity between the speaker feature vectors of the utterances “adult” and “man” tends to be high, and the similarity of the speaker feature vectors of the utterances “adult” and “child” tends to be low.

  The points of the present invention for dealing with the problem that the similarity decreases due to these characteristics of the speaker feature vector are as follows: (1) When registering an utterance, the length of the registered utterance is the same as that of the recognized utterance. Dividing the speech and calculating a speaker feature vector for each of the divided registered utterances that are the utterances after the division; (2) when recognizing the speaker, not the maximum value of the similarity of the speaker feature vectors It is two points that the maximum value of the average value of each degree of similarity is used as a reference for speaker recognition. Point (1) makes it possible to prevent a decrease in the similarity of speaker feature vectors due to a difference in utterance length. Point (2) corresponds to determining a speaker in consideration of the similarity of all speaker feature vectors of a registered utterance and a recognized utterance including various words, and the similarity based on a difference in words included in the registered utterance This makes it possible to reduce the influence of changes in speaker recognition on speaker recognition.

  Hereinafter, the speaker identification device according to the first embodiment will be described with reference to FIGS. FIG. 7 is a block diagram illustrating a configuration of the speaker identification device 100 according to the first embodiment. FIG. 8 is a flowchart illustrating the operation of the speaker identification device 100 according to the first embodiment. FIG. 9 is a diagram showing a state of the estimated speaker identification of the identification utterance by the speaker identification device 100. As shown in FIG. 7, the speaker identification device 100 includes an utterance dividing unit 110, an acoustic analysis unit 820-1, a speaker feature vector calculation unit 830-1, a speaker registration unit 840, and a registered utterance recording unit 803. An acoustic analysis unit 820-2, a speaker feature vector calculation unit 830-2, a normalization similarity calculation unit 150, a similarity averaging unit 160, a speaker identification unit 870, and an acoustic analysis unit 820. -3, a speaker feature vector calculation unit 830-3, a speaker feature vector normalization matrix learning unit 180, and a speaker feature vector normalization matrix recording unit 106. The speaker identification device 100 is connected to the UBM model recording unit 801, the speaker feature extraction model recording unit 802, the normalization matrix learning utterance recording unit 104, and the normalization matrix learning speaker name recording unit 105. doing.

  Note that, as illustrated by a dotted line in FIG. 7, the utterance division unit 110, the acoustic analysis unit 820-1, and the speaker feature vector calculation unit 830-1 are collectively divided into the registered utterance speaker feature vector calculation unit 1001 and the acoustic analysis. The unit 820-2 and the speaker feature vector calculation unit 830-2 are collectively referred to as a recognized speaker feature vector calculation unit 1002.

The speaker identification device 100 performs similarity calculation using a speaker feature vector normalization matrix. This technique using the normalization matrix is called WCCN, and it is known to reduce the variation of the similarity due to the difference in words included in the utterance (Reference Non-Patent Document 2). When WCCN is used, compared with the case where the cosine similarity of Non-Patent Document 1 is used, it is less affected by the difference in words included in the identification utterance / registered utterance, and the speaker identification accuracy is improved.
(Reference Non-Patent Document 2: AO Hatch, S. Kajarekar, A. Stolcke, “Within-Class Covariance Normalization for SVM-based Speaker Recognition”, Proc. Interspeech 2006, pp.1471-1474, 2006.)

  The processing in the speaker identification apparatus 100 is roughly divided into three processes: registration of a registered utterance, learning of a speaker feature vector normalization matrix, and identification of an identified utterance.

The outline of the registration processing flow of the registration utterance is as follows.
1. The voice of the registered utterance is divided every short time (preferably the same length as the identification utterance), and a speaker feature vector is calculated for each divided voice (referred to as a divided registered utterance) (S110, S820). -1, S830-1).
2. The speaker feature vector and the speaker name of the registered utterance are paired and registered in the registered utterance database (S840).

The outline of the learning process flow of the speaker feature vector normalization matrix is as follows.
1. A speaker feature vector is calculated for the speech of the normalization matrix learning utterance (S820-3, S830-3).
2. A variance / covariance matrix of speaker feature vectors is obtained for each speaker, and a speaker feature vector normalization matrix is learned by taking the average of these matrices (S180).

  Note that the normalization matrix learning utterance and the normalization matrix learning speaker name used for learning the speaker feature vector normalization matrix are the normalization matrix learning utterance recording unit 104 and the normalization matrix learning story before the learning starts. It records in person name recording part 105, respectively.

The outline of the identification utterance processing flow is as follows.
1. A speaker feature vector is calculated for the voice of the identified utterance (S820-2, S830-2).
2. The similarity between the speaker feature vector of the identified utterance and the speaker feature vector of each divided registered utterance is calculated (S150).
3.2. The similarities obtained in step 1 are averaged for each speaker, and the average similarity for each speaker is calculated (S160).
4). The speaker name corresponding to the maximum value of the average similarity for each speaker is returned as the speaker (identification result) of the identification utterance (S870).

Hereinafter, input, output, and operation will be described for each component.
<Speech division unit 110>
Input: Registered utterance output: Divided registered utterance Each registered utterance in the registered utterance set is divided every short time to generate divided registered utterances (S110). During division, overlapping of sections shall be allowed. That is, as shown in FIG. 9, there is an overlap in the voices of the divided registered utterances. The window width at the time of division is set to be approximately the same as the identification utterance assumed when the speaker identification is used, for example, 1.5 seconds. The shift width is, for example, 0.5 seconds.

  By allowing overlapping of sections, more divided registered utterances are generated, and it is possible to further suppress the effects resulting from the point (2) of the present invention and the difference in words.

  Note that this processing typically generates a plurality of divided registered utterances having the same length as the identification utterance, but one divided registered utterance that is at least shorter than the length of the registered utterance. It may be generated only.

<Acoustic analysis unit 820-1, 820-2, 820-3>
Input: Divided registered utterance, recognition utterance, utterance for normalization matrix learning output: acoustic feature vector sequence An acoustic feature vector sequence is calculated from each utterance (S820-1, S820-2, S820-3). The calculated acoustic feature vector sequences are referred to as divided registered utterance acoustic feature vector sequences, recognized utterance acoustic feature vector sequences, and normalized matrix learning utterance acoustic feature vector sequences, respectively. The recognition utterance acoustic feature vector corresponding to the input of the identification utterance to the acoustic analysis unit 820-2 of the speaker identification device and the input of the verification utterance to the acoustic analysis unit 820-2 of the speaker verification device The series is referred to as an identification utterance acoustic feature vector series and a verification utterance acoustic feature vector series, respectively.

Here, MFCC is used as the acoustic feature quantity. The MFCC expresses a spectral envelope every short time and is widely used in speech-related technologies such as speech recognition. A vector notation of values of each dimension of the MFCC is referred to as an acoustic feature vector, and an MFCC vector arranged in the time direction is referred to as an acoustic feature vector vector. A method for extracting MFCC is described in Reference Non-Patent Document 3. The frame width for MFCC extraction is, for example, 25 ms, and the shift width is, for example, 10 ms. Also, the MFCC dynamic feature quantity is included in the acoustic feature quantity vector.
(Reference Non-Patent Document 3: Kiyohiro Shikano, Katsunobu Ito, Tatsuya Kawahara, Kazuya Takeda, Mikio Yamamoto, “IT Text Speech Recognition System”, pp.13-14, Ohmsha, 2001)

<Speaker feature vector calculation units 830-1, 830-2, 830-3>
Input: acoustic feature vector series, UBM model, speaker feature extraction model output: speaker feature vector A speaker feature vector is calculated from each acoustic feature vector using the UBM model and speaker feature extraction model (S830). -1, S830-2, S830-3). The calculated speaker feature vectors are referred to as a divided registered speaker feature vector, a recognized speaker feature vector, and a normalization matrix learning speaker feature vector, respectively. Similar to the speaker identification device 800, i-vector is used as the speaker feature vector. Similar to the acoustic analysis unit 820-2, the input to the speaker feature vector calculation unit 830-2 is a recognized utterance acoustic feature vector sequence and a verification utterance acoustic feature vector sequence, and thus a recognized speaker feature vector. Are referred to as an identification speaker feature vector and a collation speaker feature vector, respectively.

<Speaker registration unit 840>
Input: Divided registered utterance feature vector, speaker name Output destination: Registered utterance database Set divided registered utterance feature vector and corresponding speaker name (speaker name of the registered utterance as the division source) And added to the registered utterance database of the registered utterance recording unit 803 (S840). As mentioned above, the speaker name is given manually. That is, in the registered utterance database, one or more divided registered utterer feature vectors are registered for each registered utterance in the registered utterance set.

<Speaker feature vector normalization matrix learning unit 180>
Input: Speaker feature vector for normalization matrix learning, speaker name for normalization matrix learning Output: Speaker feature vector normalization matrix Speaker feature vector normalization matrix is learned (S180). The speaker feature vector normalization matrix calculates the variance covariance matrix of speaker feature vectors for each speaker (that is, from the speaker feature vector group for normalization matrix learning with the same normalization matrix learning speaker name). It is obtained by averaging the obtained variance-covariance matrix among all speakers.

<Normalization similarity calculation unit 150>
Input: Speaker feature vector normalization matrix, identified speaker feature vector, divided registered speaker feature vector, speaker name output: similarity, speaker name Identified speaker feature vector and registered utterance database Similarities with all divided registered speaker feature vectors are calculated (S150). The calculated similarity is output in combination with the speaker name corresponding to the divided registered speaker feature vector.

  Similar to the speaker identification device 800, the cosine similarity is used as the vector similarity. Here, however, a speaker feature vector normalization matrix is used when calculating the cosine similarity. Specifically, the cosine similarity c is given by the following equation.

Here, w ₁ and w ₂ represent an identification speaker feature vector, a divided registered speaker feature vector, and W a speaker feature vector normalization matrix, respectively. In addition, ^t represents transposition and ^-1 represents an inverse matrix.

<Similarity averaging unit 160>
Input: similarity, speaker name output: average similarity for each speaker, speaker name Similarity is averaged for each speaker, and average similarity for each speaker is obtained (S160). The calculated average similarity is output in combination with the speaker name corresponding to the divided registered speaker feature vector. In order to take an average for each speaker, for example, an average may be taken in the range of similarity with the divided registered speaker feature vector having the same division source. Further, when a plurality of registered utterances by the same speaker are input to the speaker identification device 100 and the divided registered utterance feature vector calculated from the registered utterance is registered in the registration database, the same speaker The average similarity may be obtained by using all or a part of these divided registered speaker feature vectors paired with the name.

  When only one divided registered utterance is generated in the process of the utterance dividing unit 110, the speaker identification device 100 is configured without the similarity averaging unit 160.

<Speaker identification unit 870>
Input: Average similarity for each speaker, Speaker name output: Identification result (speaker name)
Of the average similarity for each speaker, the speaker name corresponding to the maximum average similarity (that is, the speaker name of the registered utterance as the division source) is selected and returned as an identification result (S870).

  Further, the input, output, and operation of the divided registered speaker feature vector calculation unit 1001 and the recognized speaker feature vector calculation unit 1002 are summarized as follows.

<Divided Registered Speaker Speaker Feature Vector Calculation Unit 1001>
Input: Registered utterance, UBM model, speaker feature extraction model Output: Divided registered utterance feature vector Divided registration using segmented registered utterance generated from registered utterance using UBM model and speaker feature extraction model The speaker feature vector is calculated (S110, S820-1, S830-1).

<Recognized utterer feature vector calculation unit 1002>
Input: recognized utterance, UBM model, speaker feature extraction model output: recognized utterer feature vector A recognized utterer feature vector is calculated from the recognized utterance using the UBM model and speaker feature extraction model (S820-). 2, S830-2).

  In order to perform speaker identification correctly even when the identification utterance is extremely short compared to the registered utterance, it is necessary to increase the similarity of the speaker feature vector for the same speaker and decrease the similarity for another speaker . The simplest way to deal with the difference in utterance length is, for example, extracting a partial section from the registered utterance that has the same utterance length as the identified utterance, and calculating the speaker feature vector only from that section and calculating the similarity A method of obtaining However, in this method, the degree of similarity changes due to the influence of the word information of the section to be extracted. Therefore, if a section containing words similar to the registered utterance and the identification utterance is extracted from the registered utterance, a different speaker But the similarity is high.

  Therefore, it is necessary to consider changes in the speaker feature vector due to the influence of words included in the utterance as well as dealing with the difference in utterance length. In the invention of the first embodiment, the utterance dividing unit 110 is provided to deal with the difference in utterance length, and the similarity averaging unit 160 is provided to deal with the influence of words included in the utterance.

  This improves the speaker identification accuracy even when the identification utterance is extremely short compared to the registered utterance. Also, speaker identification accuracy is improved for both the text-dependent type and the text-independent type.

  In the first embodiment, the similarity is calculated using the speaker feature vector normalization matrix. However, similar to the speaker identification device 800, the similarity may be calculated by the method described in Non-Patent Document 1.

  Hereinafter, the speaker identification apparatus according to the second embodiment will be described with reference to FIGS. FIG. 10 is a block diagram illustrating a configuration of the speaker identification device 200 according to the second embodiment. FIG. 11 is a flowchart illustrating the operation of the speaker identification device 200 according to the second embodiment. The difference from the speaker identification apparatus 100 of the first embodiment is that there is no component related to learning of the speaker feature vector normalization matrix, and a similarity calculation unit 850 is added instead of the normalization similarity calculation unit 150. It has been done.

  Note that, as illustrated by the dotted line in FIG. 10, as in the first embodiment, the utterance division unit 110, the acoustic analysis unit 820-1, and the speaker feature vector calculation unit 830-1 are collectively calculated to calculate the divided registered speaker feature vectors. The unit 1001, the acoustic analysis unit 820-2, and the speaker feature vector calculation unit 830-2 are collectively referred to as a recognized speaker feature vector calculation unit 1002.

  Hereinafter, input, output, and operation of the similarity calculation unit 850 different from the first embodiment will be described.

<Similarity calculation unit 850>
Input: Identification speaker feature vector, segmented registered speaker feature vector, speaker name output: similarity, speaker name All segmented registered speakers included in the identification speaker feature vector and registered speech database The similarity with the feature vector is calculated (S850). The similarity is the cosine similarity of Non-Patent Document 1. The calculated similarity is output in combination with the speaker name corresponding to the divided registered speaker feature vector.

  Similar to the first embodiment, the speaker identification accuracy is improved even when the identification utterance is extremely short compared to the registered utterance. Also, speaker identification accuracy is improved for both the text-dependent type and the text-independent type.

  In the third embodiment, in addition to the processing performed by the speaker identification device according to the first or second embodiment, registered non-registered speaker determination processing (threshold processing) using a threshold value is performed. When the speaker of the identification utterance is in the registered utterance database (when it is determined that the predetermined requirement is satisfied as a result of the threshold processing), the speaker is the processing result of the speaker identification device of the first or second embodiment. The name is output as the identification result, and when the speaker of the identification utterance is not in the registered utterance database (when it is determined that the predetermined requirement is not satisfied as a result of the threshold processing), it is determined as the non-registered speaker. Here, the registered non-speaker means a speaker whose speaker name is not registered in the registered utterance database.

  Hereinafter, the speaker identification device according to the third embodiment will be described with reference to FIGS. FIG. 12 is a block diagram illustrating a configuration of the speaker verification device 300 according to the third embodiment. FIG. 13 is a flowchart illustrating the operation of the speaker verification device 300 according to the third embodiment. FIG. 14 is a block diagram illustrating a configuration of a speaker verification device 300 ′ according to the third embodiment (modification). The speaker verification device 300 is based on the speaker verification device 100 of the first embodiment, and the speaker verification device 300 'is based on the speaker verification device 200 of the second embodiment. In the speaker verification device 300 and the speaker verification device 300 ′, a registered non-registered speaker determination unit 310 is added.

  As shown in FIG. 12 and FIG. 14 by dotted lines, as in the first embodiment, the utterance division unit 110, the acoustic analysis unit 820-1, and the speaker feature vector calculation unit 830-1 are collectively grouped registered utterers. The feature vector calculation unit 1001, the acoustic analysis unit 820-2, and the speaker feature vector calculation unit 830-2 are collectively referred to as a recognized speaker feature vector calculation unit 1002.

  Hereinafter, input, output, and operation of the speaker identification unit 875 and the non-registered speaker determination unit 310 that are different from those in the first or second embodiment will be described.

<Speaker identification unit 875>
Input: Average similarity for each speaker, speaker name output: Speaker name corresponding to the maximum value of average similarity, maximum value of average similarity Of average similarity for each speaker, the maximum average similarity Is selected, and the average similarity (that is, the maximum average similarity) corresponding to the selected speaker name is output (S875). The speaker identification unit 870 outputs the speaker name corresponding to the maximum value of the average similarity as the identification result, but the speaker identification unit 875 also outputs the maximum value of the average similarity.

<Registered outside speaker determination unit 310>
Input: Speaker name corresponding to the maximum value of average similarity, maximum value of average similarity, threshold output: Identification result It is determined whether the speaker of the identified utterance is a non-registered speaker, and the speaker of the identified utterance is registered utterance If it is in the database, the speaker name is output, and if the speaker of the identified utterance is not in the registered utterance database, the fact that the speaker is a non-registered speaker is output as the identification result (S310). The determination of the registered outside speaker is realized by threshold processing of the maximum value of the average similarity. It is assumed that the threshold is set in advance.

  In general, when determining whether or not the speaker is a registered speaker, a method of thresholding similarity is considered, that is, if the similarity is greater than the threshold, the registered speaker is regarded as the registered speaker. Use a method that considers it not. If registered utterances of multiple speakers are registered in the registered utterance database, threshold processing is performed for all registered speakers, and if there is one registered speaker whose similarity is greater than the threshold, If there is no one, it can be determined that the speaker is a registered outside speaker.

  However, it is also possible to realize the determination of registered outside speakers by performing threshold processing only on the maximum value of the average similarity. If the maximum average similarity is greater than the threshold, it is determined that at least one or more speakers are registered speakers.If the average similarity is smaller than the threshold, other speakers also have an average similarity smaller than the threshold. It is determined that all registered speakers are not the person himself / herself (is a non-registered speaker). Therefore, here, using the maximum value of the average similarity and the threshold value, if the maximum value of the average similarity value is larger than the threshold value, it is determined that the speaker is not a registered outside speaker, and corresponds to the maximum value of the average similarity value. When the maximum value of the average similarity is smaller than the threshold, the speaker name is determined as a registered non-speaker, and information indicating that the speaker is a registered non-speaker is returned as an identification result.

  Note that threshold processing may be performed depending on whether or not the threshold value is exceeded. In addition, the fact that the maximum value of the average similarity is greater than or equal to or greater than the preset threshold is said to be within a predetermined range indicating that the maximum value of the average similarity is large.

  Similar to the first embodiment, the speaker identification accuracy is improved even when the identification utterance is extremely short compared to the registered utterance. Also, speaker identification accuracy is improved for both the text-dependent type and the text-independent type.

  Hereinafter, the speaker verification apparatus according to the fourth embodiment will be described with reference to FIGS. FIG. 15 is a block diagram illustrating a configuration of the speaker verification device 400 according to the fourth embodiment. FIG. 16 is a flowchart illustrating the operation of the speaker verification device 400 according to the fourth embodiment. FIG. 17 is a diagram showing how collation utterances are collated by the speaker collation apparatus 400. As shown in FIG. 15, the speaker verification device 400 includes an utterance dividing unit 110, an acoustic analysis unit 820-1, a speaker feature vector calculation unit 830-1, a speaker registration unit 940, and a registered utterance recording unit 903. An acoustic analysis unit 820-2, a speaker feature vector calculation unit 830-2, a normalized similarity calculation unit 450, a similarity averaging unit 460, a speaker verification unit 970, and an acoustic analysis unit 820. -3, a speaker feature vector calculation unit 830-3, a speaker feature vector normalization matrix learning unit 180, and a speaker feature vector normalization matrix recording unit 106. Speaker verification device 400 is connected to UBM model recording unit 801, speaker feature extraction model recording unit 802, normalization matrix learning utterance recording unit 104, and normalization matrix learning speaker name recording unit 105. doing.

  As illustrated by the dotted line in FIG. 15, similarly to the first embodiment, the utterance dividing unit 110, the acoustic analysis unit 820-1, and the speaker feature vector calculating unit 830-1 are collectively calculated to calculate the divided registered speaker feature vectors. The unit 1001, the acoustic analysis unit 820-2, and the speaker feature vector calculation unit 830-2 are collectively referred to as a recognized speaker feature vector calculation unit 1002.

  In the speaker verification device 400, similar to the first embodiment, the similarity calculation is performed using the speaker feature vector normalization matrix.

  The processing in the speaker verification device 400 is roughly divided into three processes: registration of registered utterance, learning of speaker feature vector normalization matrix, and verification of verification utterance.

The outline of the registration processing flow of the registration utterance is as follows.
1. The voice of the registered utterance is divided every short time (preferably the same length as the identification utterance), and a speaker feature vector is calculated for each divided voice (referred to as a divided registered utterance) (S110, S820). -1, S830-1).
2. The speaker feature vector is registered in the registered utterance database (S940).

  Since the learning process flow of the speaker feature vector normalization matrix is as described in the first embodiment, it is omitted here.

The outline of the verification processing flow for verification utterance is as follows.
1. A speaker feature vector is calculated for the voice of the verification utterance (S820-2, S830-2).
2. The similarity between the speaker feature vector of the verification utterance and the speaker feature vector of each divided registered utterance is calculated (S450).
3.2. The degree of similarity obtained in step 1 is averaged for each registered utterance and the average degree of similarity for each registered utterance is calculated (S460).
4). Based on whether the maximum value of the average similarity is greater than the threshold value, it is determined whether or not the speaker is the registered speaker, and the collation result is returned (S970).

  Below, an input, an output, and operation | movement are demonstrated for every structure part.

<Speaker registration unit 940>
Input: Divided registered utterer feature vector Output destination: Registered utterance database The divided registered utterer feature vector is added to the registered utterance database of the registered utterance recording unit 903 (S940). As in the first embodiment, one or more divided registered utterer feature vectors are registered in the registered utterance database for each registered utterance in the registered utterance set.

  Whether the divided registered utterance feature vector is derived from the utterance of the same speaker so that the similarity averaging unit 460 can perform processing (similarity averaging) (see FIG. 17). Information that can be used to determine whether or not is required. For example, an identifier indicating a registered utterance that is a division source may be used. Of course, like the speaker identification device, the speaker name may be registered in the registered utterance database, and the range in which the average value is calculated using the speaker name as a key may be specified.

<Normalization similarity calculation unit 450>
Input: Speaker feature vector normalization matrix, collated utterer feature vector, divided registered utterer feature vector Output: similarity Collated utterer feature vector and all divided registered utterers included in registered utterance database The similarity with the feature vector is calculated (S450). The difference from the normalized similarity calculation unit 150 is that the speaker name is not input / output.

<Similarity averaging unit 460>
Input: similarity output: average similarity The similarity is averaged for each registered utterance to obtain an average similarity for each registered utterance (S460). The difference from the similarity averaging unit 160 is that the speaker name is not input / output. The range in which the average value is calculated may be a range that can be specified by information that can be used to determine whether or not it is derived from the utterance of the same speaker, instead of the registered utterance unit.

<Speaker verification unit 970>
Input: Average similarity, threshold output: Matching result If the maximum average similarity is greater than the threshold, it is determined that the speaker is the same speaker as the registered utterance. The result is returned (S970). Alternatively, each similarity that is the output of the similarity calculation unit 950 is compared with a threshold value, and if any one is larger than the threshold value, a matching result indicating that the speaker is a registered utterance is generated and output. You may make it (S970). Similar to the registered outside speaker determination unit 310, threshold processing may be performed depending on whether or not the threshold is equal to or greater than the threshold.

  When a speaker feature vector is used in text-independent speaker verification, due to the above-described characteristics of the speaker feature vector, the similarity varies depending on the length of the utterance of the identified utterance and the words included, and the verification accuracy decreases. For example, when a certain word is included in the collation utterance, a situation may occur in which it is always determined that the speaker is a registered speaker. The simplest way to deal with differences in utterance lengths and words in order to perform robust speaker verification against changes in utterance length and words is, for example, one of registered utterances so that the utterance length is comparable to that of verification utterances. A method is conceivable in which a partial section is extracted and a speaker feature vector is calculated only from the section to obtain a similarity. However, in this method, the degree of similarity changes due to the influence of the word information of the section to be extracted. Therefore, if a section containing words similar to the registered utterance and the matching utterance is extracted from the registered utterance, a different speaker But the similarity is high.

  Therefore, it is necessary to consider changes in the speaker feature vector due to the influence of words included in the utterance as well as dealing with the difference in utterance length. In the invention of the fourth embodiment, the difference of the utterance length is dealt with by providing the utterance dividing unit 110, and the influence of words included in the utterance is dealt with by providing the similarity averaging unit 460.

  As a result, even in the text-independent speaker verification, speaker verification using a threshold value that is robust against changes in speech length and words can be performed.

  In the fourth embodiment, the similarity is calculated using the speaker feature vector normalization matrix. However, similar to the speaker verification device 900, the similarity may be calculated by the method described in Non-Patent Document 1.

  Hereinafter, the speaker verification apparatus according to the fifth embodiment will be described with reference to FIGS. FIG. 18 is a block diagram illustrating a configuration of the speaker verification device 500 according to the fifth embodiment. FIG. 19 is a flowchart illustrating the operation of the speaker verification device 500 according to the fifth embodiment. The difference from the speaker verification device 400 of the fourth embodiment is that there is no component related to learning of the speaker feature vector normalization matrix, and a similarity calculation unit 950 is added instead of the normalization similarity calculation unit 450. It has been done.

  As illustrated by the dotted line in FIG. 18, as in the first embodiment, the utterance dividing unit 110, the acoustic analysis unit 820-1, and the speaker feature vector calculating unit 830-1 are combined to calculate the divided registered speaker feature vector. The unit 1001, the acoustic analysis unit 820-2, and the speaker feature vector calculation unit 830-2 are collectively referred to as a recognized speaker feature vector calculation unit 1002.

  Hereinafter, input, output, and operation of the similarity calculation unit 950 different from the fourth embodiment will be described.

<Similarity calculation unit 950>
Input: Verification speaker feature vector, segmented registered speaker feature vector Output: similarity Calculates the similarity between the verification speaker feature vector and all segmented registered speaker feature vectors contained in the registered utterance database (S950). The similarity is the cosine similarity of Non-Patent Document 1. The difference from the similarity calculator 850 is that the speaker name is not input / output.

  Similarly to the invention of the fourth embodiment, speaker verification using a threshold that is robust against changes in speech length and words can be performed even in text-independent speaker verification.

<Supplementary note>
The apparatus of the present invention includes, for example, a single hardware entity as an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected, and a communication device (for example, a communication cable) capable of communicating outside the hardware entity. Can be connected to a communication unit, CPU (may include a central processing unit, cache memory or register), RAM or ROM as a memory, an external storage device as a hard disk, and their input unit, output unit, communication unit , A CPU, a RAM, a ROM, and a bus connected so that data can be exchanged between the external storage devices. If necessary, the hardware entity may be provided with a device (drive) that can read and write a recording medium such as a CD-ROM. A physical entity having such hardware resources includes a general-purpose computer.

  The external storage device of the hardware entity stores a program necessary for realizing the above functions and data necessary for processing the program (not limited to the external storage device, for example, reading a program) It may be stored in a ROM that is a dedicated storage device). Data obtained by the processing of these programs is appropriately stored in a RAM or an external storage device.

  In the hardware entity, each program stored in an external storage device (or ROM or the like) and data necessary for processing each program are read into a memory as necessary, and are interpreted and executed by a CPU as appropriate. . As a result, the CPU realizes a predetermined function (respective component requirements expressed as the above-described unit, unit, etc.).

  The present invention is not limited to the above-described embodiment, and can be appropriately changed without departing from the spirit of the present invention. In addition, the processing described in the above embodiment may be executed not only in time series according to the order of description but also in parallel or individually as required by the processing capability of the apparatus that executes the processing. .

  As described above, when the processing functions in the hardware entity (the apparatus of the present invention) described in the above embodiments are realized by a computer, the processing contents of the functions that the hardware entity should have are described by a program. Then, by executing this program on a computer, the processing functions in the hardware entity are realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape, and the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto-Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable-Programmable-Ready), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program recorded on a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. A configuration in which the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes a processing function only by an execution instruction and result acquisition without transferring a program from the server computer to the computer. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, a hardware entity is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

Claims (8)

Recognize utterance voice of speaker to be speaker identification or speaker verification, and utterance voice used to recognize the speaker as registered utterance,
A divided registered utterer feature vector calculating unit that generates a divided registered utterance having a section shorter than the section of the registered utterance from the registered utterance and calculates a divided registered utterer feature vector from the divided registered utterance; ,
A recognized utterer feature vector calculating unit for calculating a recognized utterer feature vector from the recognized utterance;
A speaker likelihood evaluation apparatus comprising: a similarity calculation unit that calculates a similarity between the recognized utterance and the divided registered utterance using the recognized utterance feature vector and the divided registered utterance feature vector. The speaker quality evaluation device according to claim 1,
The divided registered utterer feature vector calculation unit generates two or more divided registered utterances,
The similarity calculation unit calculates the similarity for each of the two or more divided registered utterances;
further,
A speaker likeness evaluation apparatus comprising: a similarity averaging unit that calculates an average similarity that is an average value of the similarities. A set of registered utterances whose elements are one or more registered utterances is a registered utterance set.
From the similarity or the average similarity calculated using the speaker evaluation apparatus according to claim 1 or 2, for each of the identification utterance and the registered utterance of the registered utterance set, the identification utterance A speaker identification device for generating an identification result,
A speaker identification device comprising: a speaker identification unit that selects a speaker name of a registered utterance corresponding to the maximum value of the similarity or the average similarity, and uses the speaker name as the identification result. A set of registered utterances whose elements are one or more registered utterances is a registered utterance set.
From the similarity or the average similarity calculated using the speaker evaluation apparatus according to claim 1 or 2, for each of the identification utterance and the registered utterance of the registered utterance set, the identification utterance A speaker identification device for generating an identification result,
A speaker name of a registered utterance corresponding to the maximum value of the similarity or the average similarity and a speaker identification unit for selecting the maximum value;
A non-registered speaker determination unit that uses the speaker name as the identification result when the maximum value is within a predetermined range indicating that the maximum value is large. A set of registered utterances whose elements are one or more registered utterances is set as a registered utterance set.
From the similarity or the average similarity calculated using the speaker evaluation device according to claim 1 or 2, for each of the verification utterance and the registered utterance of the registered utterance set, the verification utterance A speaker verification device that generates a verification result,
A speaker verification device comprising: a speaker verification unit that generates a verification result that the speaker is the registered utterance when the maximum value is within a predetermined range indicating that the maximum value is large. Recognize utterance voice of speaker to be speaker identification or speaker verification, and utterance voice used to recognize the speaker as registered utterance,
A divided registered utterer feature vector calculation step of generating a divided registered utterance having a section shorter than the section of the registered utterance from the registered utterance, and calculating a divided registered utterer feature vector from the divided registered utterance; ,
A recognized utterer feature vector calculating step for calculating a recognized utterer feature vector from the recognized utterance;
A method for evaluating the likelihood of a speaker that performs a similarity calculation step of calculating a similarity between the recognized utterance and the divided registered utterance using the recognized utterance feature vector and the divided registered utterance feature vector. The speaker quality evaluation method according to claim 6,
The divided registered utterer feature vector calculation step is to generate two or more divided registered utterances,
The similarity calculation step calculates the similarity for each of the two or more divided registered utterances,
further,
A similarity level averaging step of calculating an average similarity level that is an average value of the similarity levels.   A program for causing a computer to execute the speaker quality evaluation method according to claim 6 or 7.

Images