Voice processing device and voice authentication system
The integration of audio data and input device frequency characteristics with a DNN-based feature extraction method enhances speaker verification accuracy by accounting for device-specific sensitivities, ensuring reliable identification across varying input devices.
Patent Information
- Application Number
- JP2022545269
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2020-08-31
- Publication Date
- 2024-11-13
- Estimated Expiration
- 2040-08-31
AI Technical Summary
Speaker verification systems face a significant drop in identity identification rate when different input devices are used for registration and verification due to varying frequency sensitivities, leading to potential verification failures.
An audio processing device integrates audio data and frequency characteristics of the input device to extract speaker identification features, using a deep neural network (DNN) to combine acoustic and device features, ensuring accurate identification regardless of the input device used.
This approach enables highly accurate speaker verification by incorporating device-specific frequency characteristics, allowing consistent identification across different input devices.
Smart Images

Figure 0007582314000001 
Figure 0007582314000002 
Figure 0007582314000003
Abstract
Description
[Technical field]
[0001] The present invention relates to a voice processing device, a voice processing method, a recording medium, and a voice authentication system, and more particularly to a voice processing device, a voice processing method, a recording medium, and a voice authentication system that verify a speaker based on voice data input via an input device. [Background technology]
[0002] In a related technique, a speaker is identified by comparing the voice characteristics contained in the first voice data with the voice characteristics contained in the second voice data. This related technique is called identity verification by voice authentication or speaker verification. In recent years, speaker verification has been used in an expanding range of situations, particularly in tasks that require remote conversation, such as construction sites and factories.
[0003] Patent Document 1 describes a method for performing speaker verification by obtaining time-series features by frequency analysis of voice data and comparing the obtained feature pattern with pre-registered feature patterns.
[0004] In the related technology described in Patent Document 2, voice characteristics input using an input device such as a microphone for calls or a headset microphone equipped on a smartphone are compared with voice characteristics registered using another input device. For example, voice characteristics registered using a tablet in an office are compared with voice characteristics input from a headset microphone at the site. [Prior art documents] [Patent documents]
[0005] [Patent Document 1] Japanese Patent Application Publication No. 07-084594 [Patent Document 2] JP 2016-075740 A Summary of the Invention [Problem to be solved by the invention]
[0006] When the input device used for enrollment is different from the input device used for verification, the frequency ranges to which the input devices are sensitive differ. In such a case, the identification rate decreases compared to when the same input device is used for both enrollment and verification. As a result, the possibility of failure in speaker verification increases.
[0007] The present invention has been made in view of the above problems, and has an object to realize highly accurate speaker verification regardless of the input device. [Means for solving the problem]
[0008] A voice processing apparatus according to one embodiment of the present invention includes an integration means for integrating voice data input using an input device with the frequency characteristics of the input device, and a feature extraction means for extracting speaker identification features for identifying the speaker of the voice data from integrated features obtained by integrating the voice data with the frequency characteristics.
[0009] A voice processing method according to one embodiment of the present invention includes integrating voice data input using an input device with frequency characteristics of the input device, and extracting speaker identification features for identifying the speaker of the voice data from integrated features obtained by integrating the voice data with the frequency characteristics.
[0010] A recording medium according to one embodiment of the present invention stores a program for causing a computer to execute a process of integrating voice data input using an input device with the frequency characteristics of the input device, and a process of extracting speaker identification features for identifying the speaker of the voice data from integrated features obtained by integrating the voice data with the frequency characteristics.
[0011] A voice authentication system according to one embodiment of the present invention includes a voice processing device according to one embodiment of the present invention, and a verification device that verifies whether the speaker is a registered person based on speaker identification features output from the voice processing device. Effect of the Invention
[0012] According to one aspect of the present invention, highly accurate speaker verification can be achieved regardless of the input device. [Brief description of the drawings]
[0013] [Figure 1] 1 is a block diagram showing a configuration of a voice authentication system common to all embodiments. [Diagram 2] 1 is a block diagram showing a configuration of a voice processing device according to a first embodiment. [Diagram 3] 1 is a graph showing an example of frequency dependency (frequency characteristics) of sensitivity for an input device. [Figure 4] 1 shows a characteristic vector obtained from an example of the frequency characteristics of an input device. [Diagram 5] FIG. 2 is a diagram for explaining a flow in which the feature extraction unit according to the first embodiment obtains features for speaker identification from integrated features using DNN. [Figure 6] 4 is a flowchart showing the operation of the voice processing device according to the first embodiment. [Figure 7] FIG. 11 is a block diagram showing the configuration of a voice processing device according to a second embodiment. [Figure 8] 10 is a flowchart showing the operation of the voice processing device according to the second embodiment. [Figure 9] FIG. 1 is a diagram illustrating a hardware configuration of a voice processing device according to a first or second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0014] [Common to all embodiments] First, an example of the configuration of a voice authentication system that is commonly applied to all the embodiments described below will be described.
[0015] (Voice Authentication System 1) An example of the configuration of a voice authentication system 1 will be described with reference to Fig. 1. Fig. 1 is a block diagram showing an example of the configuration of the voice authentication system 1.
[0016] As shown in Fig. 1, the voice authentication system 1 includes a voice processing device 100 (200) and a matching device 10. The voice authentication system 1 may also include one or more input devices. The voice processing device 100 (200) is the voice processing device 100 or the voice processing device 200.
[0017] The processes and operations executed by the voice processing device 100 (200) will be described in detail in the first and second embodiments described later. The voice processing device 100 (200) acquires voice data (hereinafter referred to as registered voice data) of a speaker (person A) registered in advance from a database (DB) on a network or from a DB connected to the voice processing device 100 (200). The voice processing device 100 (200) also acquires voice data (hereinafter referred to as collation voice data) of a subject (person B) to be collated from an input device. The input device is used to input voice to the voice processing device 100 (200). In one example, the input device is a microphone or a headset microphone for calls provided on a smartphone.
[0018] The voice processing device 100 (200) generates a feature A for speaker identification based on the registered voice data. The voice processing device 100 (200) also generates a feature B for speaker identification based on the matching voice data. The feature A for speaker identification is obtained by integrating the registered voice data registered in the DB with the frequency characteristics of the input device used to input the registered voice data. The acoustic feature is a feature vector whose elements are one or more feature amounts (hereinafter, sometimes referred to as first parameters) that are numerical values that quantitatively represent the characteristics of the registered voice data. The device feature is a feature vector whose elements are one or more feature amounts (hereinafter, sometimes referred to as second parameters) that are numerical values that quantitatively represent the characteristics of the input device. The feature B for speaker identification is obtained by integrating the matching voice data input using the input device with the frequency characteristics of the input device used to input the matching voice data.
[0019] The following two-step process is called "integration" of voice data (registered voice data or matching voice data) and frequency characteristics of an input device. Hereinafter, the registered voice data or matching voice data will be described as registered voice data / matching voice data. The first step is to extract an acoustic feature related to the frequency characteristics of the registered voice data / matching voice data, and also to extract a device feature related to the frequency characteristics of the sensitivity of the input device used for input. The second step is to combine both the acoustic feature and the device feature. The combination is to decompose the acoustic feature into a first parameter that is its element, and to decompose the device feature into a second parameter that is its element, and to generate a feature vector that includes both the first parameter and the second parameter as elements of independent dimensions. As described above, the first parameter is a feature extracted from the frequency characteristics of the registered voice data / matching voice data. The second parameter is a feature extracted from the frequency characteristics of the sensitivity of the input device used for inputting the registered voice data / matching voice data. In this case, the combination means generating an (n+m)-dimensional feature vector whose elements are n features that are the first parameters constituting the acoustic feature and m features that are the second parameters constituting the device feature (n and m are integers).
[0020] This results in a feature (hereafter referred to as an integrated feature) that depends on both the frequency characteristics of the registered voice data / matching voice data and the frequency characteristics of the sensitivity of the input device used to input the registered voice data / matching voice data. The integrated feature is a feature vector whose elements are multiple (n+m in the above example) feature quantities.
[0021] The meaning of "integration" in each embodiment described later is the same as that described here.
[0022] The acoustic features are extracted from the enrollment voice data and the matching voice data. Meanwhile, the device features are extracted from data related to the input device (for example, data indicating the frequency characteristics of the sensitivity of the input device). Then, the voice processing device 100 (200) transmits the speaker identification features A and B to the matching device 10.
[0023] The matching device 10 receives speaker identification features A and speaker identification features B from the voice processing device 100 (200). The matching device 10 confirms whether the speaker is a registered person based on the speaker identification features A and speaker identification features B output from the voice processing device 100 (200). More specifically, the matching device 10 compares the speaker identification features A and B and outputs a result of the identity confirmation. That is, the matching device 10 outputs information indicating whether person A and person B are the same person.
[0024] In addition, the voice authentication system 1 may be equipped with a control device (control function) that controls the electronic lock on the door for entering the office, automatically starts up or logs on to an information terminal, and permits access to information on an intranet based on the identity verification result output by the verification device 10.
[0025] The voice authentication system 1 may be realized as a network service. In this case, the voice processing device 100 (200) and the verification device 10 may be on a network and may be capable of communicating with one or more input devices via a wireless network.
[0026] The following describes a specific example of the voice processing device 100 (200) included in the voice authentication system 1. In the following description, "voice data" refers to both "enrollment voice data" and "verification voice data."
[0027] [Embodiment 1] A voice processing device 100 will be described as a first embodiment with reference to FIGS.
[0028] (Speech processing device 100) The configuration of the voice processing device 100 according to the first embodiment will be described with reference to Fig. 2. Fig. 2 is a block diagram showing the configuration of the voice processing device 100. As shown in Fig. 2, the voice processing device 100 includes an integration unit 110 and a feature extraction unit 120.
[0029] The integrating unit 110 integrates the voice data input using one or more input devices with the frequency characteristics of the input devices. The integrating unit 110 is an example of an integrating means.
[0030] In one example, the integration unit 110 acquires voice data (registered voice data or collated voice data in FIG. 1) and information identifying an input device used to input the voice data. The integration unit 110 extracts acoustic features from the voice data. For example, the acoustic features may be Mel-Frequency Cepstrum Coefficients (MFCC) or linear predictive coding (LPC) coefficients, or may be a power spectrum or a spectral envelope. Alternatively, the acoustic features may be a feature vector of any dimension (hereinafter, referred to as an acoustic vector) composed of feature amounts obtained by frequency analysis of the voice data. In one example, the acoustic vector indicates the frequency characteristics of the voice data.
[0031] Furthermore, the integration unit 110 acquires data related to the input device from the DB (FIG. 1) using information identifying the input device. Specifically, the integration unit 110 acquires data indicating the frequency dependency (called frequency characteristics) of the sensitivity of the input device.
[0032] Fig. 3 is a graph showing an example of the frequency characteristics of an input device. In the graph shown in Fig. 3, the vertical axis represents sensitivity (dB) and the horizontal axis represents frequency (Hz). The integration unit 110 extracts device features from the data of the frequency characteristics of the input device.
[0033] Fig. 4 shows an example of a device feature. In the example shown in Fig. 4, the device feature is a characteristic vector F (an example of a device feature) indicating the frequency characteristics of the sensitivity of an input device. The characteristic vector F has elements (f1, f2, f3, ..., f32) that are the average values obtained by integrating the sensitivity of the input device (Fig. 3) in a frequency band (a band of a predetermined width including the frequency bin) for each frequency bin and dividing the integral value by the bandwidth.
[0034] The integration unit 110 combines the acoustic feature and the device feature thus obtained to obtain an integrated feature based on the matching voice data and an integrated feature based on the registered voice data. As described with respect to the voice authentication system 1, the integrated feature is a feature vector that depends on both the frequency characteristics of the registered voice data / matching voice data and the frequency characteristics of the sensitivity of the input device used to input the registered voice data / matching voice data. As described above, the integrated feature includes a first parameter related to the frequency characteristics of the registered voice data / matching voice data and a second parameter related to the frequency characteristics of the sensitivity of the input device used to input the registered voice data / matching voice data. An example of the process related to the details of the integration and the integrated feature will be described in the second embodiment. The integration unit 110 outputs the integrated feature thus obtained to the feature extraction unit 120.
[0035] The feature extraction unit 120 extracts speaker identification features (speaker identification features A and B) for identifying the speaker of the voice from the integrated features obtained by integrating the voice data and the frequency characteristics. The feature extraction unit 120 is an example of a feature extraction means.
[0036] An example of a process in which the feature extraction unit 120 extracts features for speaker identification from the integrated features will be described with reference to Fig. 5. As shown in Fig. 5, the feature extraction unit 120 includes a DNN (Deep Neural Network).
[0037] In the learning phase, the feature extraction unit 120 inputs training data and updates each parameter of the DNN based on an arbitrary loss function so that the output result matches the correct answer data. The correct answer data is data that indicates the correct answer of the speaker. The DNN completes learning so that it can identify the speaker based on the integrated features before the phase for extracting features for speaker identification.
[0038] The feature extraction unit 120 inputs the integrated features to the trained DNN. The DNN of the feature extraction unit 120 uses the input integrated features to identify a speaker (e.g., person A or person B). The feature extraction unit 120 also extracts speaker identification features that the trained DNN focuses on.
[0039] Specifically, the feature extraction unit 120 extracts speaker identification features focused on for identifying the speaker from the intermediate layer of the DNN. In other words, the feature extraction unit 120 extracts speaker identification features for identifying the speaker of the voice using an integrated feature obtained by integrating the voice data and the frequency characteristics and the DNN. Therefore, since the speaker identification features are acquired based on the acoustic features and the device features, the speaker identification features do not depend on the frequency characteristics of the input device. Therefore, the matching device 10 can identify the speaker based on the speaker identification features regardless of whether the same input device (with frequency characteristics) or different input devices (with frequency characteristics) are used at the time of registration and at the time of matching.
[0040] (Operation of the voice processing device 100) The operation of the voice processing device 100 according to the first embodiment will be described with reference to Fig. 6. Fig. 6 is a flowchart showing the flow of processes executed by each unit of the voice processing device 100.
[0041] 6, the integration unit 110 integrates the voice data input using the input device with the frequency characteristics of the input device (S1). The integration unit 110 outputs the integrated feature data obtained as a result of step S1 to the feature extraction unit 120.
[0042] The feature extraction unit 120 receives data of the integrated feature obtained by integrating the voice data and the frequency characteristic from the integration unit 110. The feature extraction unit 120 extracts features for speaker identification from the received integrated feature (S2).
[0043] The feature extraction unit 120 outputs the data of the speaker identification features obtained as a result of step S2. In one example, the feature extraction unit 120 transmits the data of the speaker identification features to the matching device 10 (FIG. 1). Note that when training the above-mentioned DNN, the voice processing device 100 also obtains the data of the speaker identification features according to the procedure described here, and stores the data of the speaker identification features linked to the information for identifying the speaker as training data in a training DB (training database) not shown. The above-mentioned DNN performs training to identify speakers using the training data stored in the training DB.
[0044] This is the end of the operation of the voice processing device 100 according to the first embodiment.
[0045] (Effects of this embodiment) According to the configuration of this embodiment, the integration unit 110 integrates the voice data input using the input device with the frequency characteristics of the input device, and the feature extraction unit 120 extracts speaker identification features for identifying the speaker of the voice from the integrated features obtained by integrating the voice data with the frequency characteristics. The speaker identification features include not only information related to the acoustic features of the voice input using the input device, but also information related to the frequency characteristics of the input device. Therefore, the matching device 10 of the voice authentication system 1 can perform speaker matching with high accuracy based on the speaker identification features, regardless of whether the input device used to input the voice at the time of registration is the same as or different from the input device used to input the voice at the time of matching.
[0046] However, it is desirable that the input device used for voice input during registration has sensitivity over a wider range than the input device used for voice input during verification. More specifically, the bandwidth used by the input device used for voice input during registration (bandwidth to which it has sensitivity) should preferably include the bandwidth used by the input device used for voice input during verification.
[0047] [Embodiment 2] A voice processing device 200 will be described as a second embodiment with reference to FIGS.
[0048] (Speech processing device 200) The configuration of a voice processing device 200 according to the second embodiment will be described with reference to Fig. 7. Fig. 7 is a block diagram showing the configuration of the voice processing device 200. As shown in Fig. 7, the voice processing device 200 includes an integration unit 210 and a feature extraction unit 120.
[0049] The integration unit 210 integrates the voice data input using the input device with the frequency characteristics of the input device. The integration unit 210 is an example of an integration means. As shown in FIG. 7, the integration unit 210 includes a characteristic vector calculation unit 211, a voice conversion unit 212, and a combination unit 213.
[0050] The characteristic vector calculation unit 211 calculates an average value of the sensitivity of the input device in one frequency band (a band of a predetermined width including the frequency bin) for each frequency bin, and sets the average value calculated for each frequency bin as an element of a characteristic vector (an example of a device feature). The characteristic vector indicates frequency characteristics unique to the input device. The characteristic vector calculation unit 211 is an example of a characteristic vector calculation means.
[0051] In one example, the characteristic vector calculation unit 211 of the integration unit 210 acquires data related to the input device from a DB (FIG. 1) or an input unit (not shown). The data related to the input device includes information for identifying the input device and data indicating the sensitivity of the input device. The characteristic vector calculation unit 211 calculates an average value of the sensitivity of the input device in one frequency band (a band of a predetermined width including the frequency bin) for each frequency bin from the data indicating the sensitivity of the input device. Next, the characteristic vector calculation unit 211 calculates a characteristic vector having the average value of the sensitivity for each frequency bin as an element. Then, the characteristic vector calculation unit 211 transmits the calculated characteristic vector data to the combination unit 213.
[0052] The speech conversion unit 212 converts the speech data from the time domain to the frequency domain to obtain an acoustic vector sequence (an example of an acoustic feature). Here, the acoustic vector sequence represents a time series of acoustic vectors for each predetermined time width. The speech conversion unit 212 is an example of a speech conversion means.
[0053] In one example, the voice conversion unit 212 of the integration unit 210 receives collated voice data from an input device and acquires registered voice data from a DB. The voice conversion unit 212 converts the voice data into amplitude spectrum data for each predetermined time width by fast Fourier transform (FFT).
[0054] Furthermore, the voice conversion unit 212 may use a filter bank to divide the amplitude spectrum data for each predetermined time width into each predetermined frequency band.
[0055] The speech conversion unit 212 obtains a plurality of features from the amplitude spectrum data for each predetermined time width (or from the data obtained by dividing the data for each predetermined frequency band using a filter bank). The speech conversion unit 212 then generates an acoustic vector composed of the obtained plurality of features. In one example, the feature is the acoustic intensity for each predetermined frequency range. In this way, the speech conversion unit 212 obtains a time series of acoustic vectors for each predetermined time width (hereinafter, referred to as an acoustic vector sequence). The speech conversion unit 212 then transmits the calculated acoustic vector sequence data to the combination unit 213.
[0056] The combining unit 213 obtains a feature-acoustic vector sequence (an example of an integrated feature) by "combining" an acoustic vector sequence (an example of an acoustic feature) with a feature vector (an example of a device feature).
[0057] In one example, the combining unit 213 of the integration unit 210 receives data of the characteristic vector from the characteristic vector calculation unit 211. In addition, the combining unit 213 receives data of the acoustic vector sequence from the voice conversion unit 212.
[0058] The combining unit 213 then expands the dimensions of each acoustic vector in the acoustic vector sequence, and adds the elements of the characteristic vector as elements of the acoustic vectors whose dimensions have been expanded in the acoustic vector sequence.
[0059] The combining unit 213 outputs the data of the characteristic-acoustic vector sequence thus obtained to the feature extracting unit 120.
[0060] The feature extraction unit 120 extracts speaker identification features for identifying the speaker of the voice from a characteristic-acoustic vector sequence (an example of an integrated feature) obtained by combining an acoustic vector sequence (an example of an acoustic feature) with a characteristic vector (an example of a device feature). The feature extraction unit 120 is an example of a feature extraction means.
[0061] In one example, the feature extraction unit 120 receives data of the feature-acoustic vector sequence from the combination unit 213 of the integration unit 210. The feature extraction unit 120 inputs the data of the feature-acoustic vector sequence to the trained DNN (FIG. 5). The feature extraction unit 120 obtains integrated features based on the feature-acoustic vector sequence from an intermediate layer of the trained DNN. The integrated features are features extracted from the feature-acoustic vector sequence.
[0062] The feature extraction unit 120 outputs data of integrated features based on the characteristic-acoustic vector sequence to the matching device 10 (FIG. 1).
[0063] (Modification) In this modified example, the acoustic vector at the time of enrollment (speaker identification feature A) is matched with the acoustic vector at the time of matching (speaker identification feature B) in the common part of the effective band where both the input device used at the time of matching and the input device used at the time of enrollment have sensitivity.
[0064] The characteristic vector calculation unit 211 in this modified example obtains a third characteristic vector by combining (described later) a first characteristic vector indicating the frequency characteristics of the sensitivity of input device A and a second characteristic vector indicating the frequency characteristics of the sensitivity of input device B.
[0065] The characteristic vector calculation section 211 according to this modification outputs the data of the third characteristic vector calculated in this manner to the combination section 213.
[0066] The combining unit 213 multiplies the third feature vector obtained by combining the two feature vectors by the acoustic vector at the time of enrollment (an example of feature A for speaker identification) and the acoustic vector at the time of matching (an example of feature B for speaker identification).
[0067] In a band where at least one of the input device used during matching and the input device used during registration has no sensitivity, the value of the third characteristic vector is zero, and therefore the value of the acoustic vector multiplied by the third characteristic vector also becomes zero in a portion other than the common portion of the effective band where the two input devices have sensitivity.
[0068] In this way, the effective band of the speaker identification feature A is the same as the effective band of the speaker identification feature B. This allows the matching device 10 (FIG. 1) to match the speaker identification feature A and the speaker identification feature B that have the same effective band.
[0069] The synthesis of two characteristic vectors in this modified example will be described in more detail. The characteristic vector calculation unit 211 compares the n-th element (fn) of the first characteristic vector with the corresponding element (gn) of the second characteristic vector. Then, the characteristic vector calculation unit 211 sets the smaller of these two elements (fn, gn) as the corresponding element of the third characteristic vector. Alternatively, the characteristic vector calculation unit 211 may set the geometric mean √(fn×gn) of the n-th element (fn) of the first characteristic vector and the corresponding element (gn) of the second characteristic vector as the n-th element of the third characteristic vector. Alternatively, the characteristic vector calculation unit 211 may input the first characteristic vector and the second characteristic vector to a DNN (not shown), and extract a third characteristic vector from the intermediate layer of the DNN in which components other than the common part of the effective bands of both the first characteristic vector and the second characteristic vector are weighted with a value of 0.
[0070] (Operation of the voice processing device 200) The operation of the voice processing device 200 according to the second embodiment will be described with reference to Fig. 8. Fig. 8 is a flowchart showing the flow of the process executed by the voice processing device 200.
[0071] 8, the characteristic vector calculation unit 211 of the integration unit 210 acquires data on an input device from a DB (FIG. 1) or an input unit (not shown) (S201). The data on the input device includes information for identifying the input device and data indicating the frequency characteristics of the input device (FIG. 3).
[0072] The characteristic vector calculation unit 211 calculates an average value of the sensitivity of the input device in one frequency band (a band of a predetermined width including the frequency bin) for each frequency bin from data indicating the frequency characteristics of the input device. The characteristic vector calculation unit 211 calculates a characteristic vector having the calculated average value of the sensitivity for each frequency bin as an element (S202). Then, the characteristic vector calculation unit 211 transmits data of the calculated characteristic vector to the combination unit 213.
[0073] The speech conversion unit 212 uses a filter bank to perform frequency analysis on the speech data, and obtains amplitude spectrum data for each predetermined time width. Furthermore, the speech conversion unit 212 calculates the above-mentioned acoustic vector sequence from the amplitude spectrum data for each predetermined time width (S203). Then, the speech conversion unit 212 transmits the calculated acoustic vector sequence data to the combination unit 213.
[0074] The combining unit 213 calculates a characteristic-acoustic vector sequence (an example of an integrated feature) by combining an acoustic vector sequence (an example of an acoustic feature) based on the voice data input using the input device with a characteristic vector (an example of a device feature) related to the frequency characteristics of the input device (S204). The combining unit 213 outputs the data of the characteristic-acoustic vector sequence thus obtained to the feature extracting unit 120.
[0075] The feature extraction unit 120 receives data of the feature-acoustic vector sequence from the combination unit 213 of the integration unit 210. The feature extraction unit 120 extracts features for speaker identification from the feature-acoustic vector sequence (S205). Specifically, the feature extraction unit 120 extracts a feature A for speaker identification (FIG. 1) from the feature-acoustic vector sequence based on the enrollment voice data, and extracts a feature B for speaker identification (FIG. 1) from the feature-acoustic vector sequence based on the collation voice data.
[0076] The feature extraction unit 120 outputs the data of the speaker identification features thus obtained. In one example, the feature extraction unit 120 transmits the data of the speaker identification features to the matching device 10 (FIG. 1).
[0077] This is the end of the operation of the voice processing device 200 according to the second embodiment.
[0078] (Effects of this embodiment) According to the configuration of this embodiment, the integration unit 210 integrates the voice data input using an input device with the frequency characteristics of the input device, and the feature extraction unit 120 extracts speaker identification features for identifying the speaker of the voice from the integrated features obtained by integrating the voice data with the frequency characteristics. The speaker identification features include not only information related to the acoustic features of the voice input using the input device, but also information related to the frequency characteristics of the input device. Therefore, the matching device 10 of the voice authentication system 1 can perform speaker matching with high accuracy based on the speaker identification features, regardless of whether the input device used to input the voice at the time of registration is the same as or different from the input device used to input the voice at the time of matching.
[0079] More specifically, the integration unit 210 includes a characteristic vector calculation unit 211 that calculates an average value of the sensitivity of the input device for each frequency bin and sets the average value calculated for each frequency bin as an element of a characteristic vector. The characteristic vector indicates the frequency characteristics of the input device.
[0080] The integration unit 210 also includes a voice conversion unit 212 that uses a filter bank to perform a Fourier transform of the voice from the time domain to the frequency domain to obtain an acoustic vector sequence. The integration unit 210 also includes a combination unit 213 that combines the acoustic vector sequence with a feature vector to obtain a feature-acoustic vector sequence. This makes it possible to obtain a feature-acoustic vector sequence in which the acoustic vector sequence, which is an acoustic feature, and the feature vector, which is a device feature, are combined.
[0081] Furthermore, the feature extraction unit 120 can obtain features for speaker identification based on the characteristic-acoustic vector sequence. Therefore, as described above, the matching device 10 of the voice authentication system 1 can perform speaker matching with high accuracy based on the features for speaker identification.
[0082] [Hardware configuration] Each of the components of the voice processing devices 100 and 200 described in the first and second embodiments is represented by a functional block. Some or all of these components are realized by an information processing device 900 as shown in Fig. 9. Fig. 9 is a block diagram showing an example of the hardware configuration of the information processing device 900.
[0083] As shown in FIG. 9, an information processing device 900 includes, for example, the following configuration.
[0084] ·CPU(Central Processing Unit)901 ROM (Read Only Memory) 902 ·RAM(Random Access Memory)903 Program 904 loaded into RAM 903 A storage device 905 for storing a program 904 A drive device 907 for reading and writing data from the recording medium 906 A communication interface 908 for connecting to a communication network 909 An input / output interface 910 for inputting and outputting data Bus 911 connecting each component Each of the components of the audio processing devices 100 and 200 described in the first and second embodiments is realized by the CPU 901 reading and executing a program 904 that realizes these functions. The program 904 that realizes the functions of each component is, for example, stored in advance in the storage device 905 or the ROM 902, and is loaded into the RAM 903 by the CPU 901 and executed as necessary. The program 904 may be supplied to the CPU 901 via the communication network 909, or may be stored in advance in the recording medium 906, and the drive device 907 may read out the program and supply it to the CPU 901.
[0085] According to the above configuration, the sound processing devices 100 and 200 described in the first and second embodiments are realized as hardware, and therefore the same effects as those described in the first and second embodiments can be achieved. [Industrial Applicability]
[0086] In one example, the present invention can be used in a voice authentication system that performs identity verification by analyzing voice data input using an input device. [Explanation of symbols]
[0087] 1. Voice authentication system 10. Collation device 100 Audio processing device 110 Integration Department 120 Feature Extraction Unit 200 Audio processing device 210 Integration Department 211 Characteristic Vector Calculation Unit 212 Voice conversion unit
Claims
1. an integration means for calculating an integrated feature by combining an acoustic vector sequence representing an acoustic feature extracted from voice data input using an input device with a characteristic vector representing characteristic information indicating a frequency characteristic of the input device; a feature extraction means for extracting a speaker identification feature for identifying a speaker of the voice data from the integrated feature; Equipped with The integration means includes: a voice conversion means for converting a frequency of the voice data to obtain the acoustic vector sequence, which is a time series of acoustic vectors indicating the frequency characteristics of the voice data inputted from the input device; a characteristic vector calculation means for calculating an average value of the sensitivity of the input device for each frequency bin, and setting the average value of the sensitivity calculated for each frequency bin as an element of the characteristic vector indicating the frequency characteristic of the input device; The feature vector calculation means obtains the feature vector by combining two feature vectors for two input devices used during speaker enrollment and verification, respectively. Audio processing device.
2. the integrated feature is a feature-acoustic vector sequence in which the acoustic vector sequence and the feature vector are combined; The integration means includes a combination means for combining the acoustic vector sequence and the characteristic vector to obtain the characteristic-acoustic vector sequence.
2. The audio processing device according to claim 1 .
3. The feature extraction means inputs the integrated features to a deep neural network (DNN) and obtains the speaker identification features from an intermediate layer of the DNN.
3. The audio processing device according to claim 1, wherein the audio processing device is a digital audio processing device.
4. A voice processing device according to any one of claims 1 to 3, a verification device that verifies whether the speaker is a registered person based on the speaker identification features output from the speech processing device; A voice authentication system.
Citation Information
Patent Citations
Voice recognition device
JP1990029100A
Speech recognition device
JP1995084594A
Speech recognition device and microphone frequency characteristic converting method
JP1998105191A
Voice processing device, voice processing method, and program
JP2016075740A
Acoustic score calculation device, and method and program therefor
JP2016122110A
