A method and device for voiceprint recognition speech evaluation

Abstract

The present invention discloses a voiceprint recognition speech evaluation method and device. The method comprises the following steps: establishing a voiceprint speech test database, establishing a voiceprint speech registration set and a test set corresponding to each speaker according to the user's voiceprint speech test requirements and the test database, sending the voiceprint speech registration set to at least one voiceprint recognition device to be tested one by one, generating a corresponding voiceprint recognition model, sending the voiceprint speech test set to at least one voiceprint recognition device to be tested one by one, completing the voiceprint recognition test, and receiving the voiceprint recognition results fed back by each voiceprint recognition device to be tested, judging whether each voiceprint recognition device to be tested has completed the voiceprint recognition test for all speech test sets, and if completed, calculating the performance index of voiceprint recognition according to the voiceprint recognition result, otherwise, continuing to wait for receiving the voiceprint recognition results fed back by each voiceprint recognition device to be tested.

Description

Technical Field

[0001] The present invention relates to a voiceprint recognition speech evaluation method and also to a corresponding voiceprint recognition speech evaluation device, belonging to the technical field of speech analysis. Background Art

[0002] At present, intelligent voice technology is widely used in the field of public security, especially the use of voiceprint recognition (also called speaker recognition) technology to confirm criminal suspects and remote voiceprint recognition and login verification for APPs in the financial field.

[0003] The existing voiceprint recognition method and device based on actual application scenarios obtain voiceprint information and extract the voiceprint features of the voiceprint information, then use a pre-established pattern recognition model to perform pattern matching on the voiceprint features to obtain label information corresponding to the voiceprint features, and store the label information and the voiceprint information in a voiceprint database in correspondence, so that the voiceprint information can be screened according to the label information during voiceprint recognition.

[0004] However, existing voiceprint recognition methods and devices can only realize the voiceprint recognition function, but cannot effectively evaluate the technical indicators of voiceprint recognition, and thus cannot confirm the voiceprint recognition effects of different methods and devices. Summary of the invention

[0005] The primary technical problem to be solved by the present invention is to provide a voiceprint recognition speech evaluation method.

[0006] Another technical problem to be solved by the present invention is to provide a voiceprint recognition speech evaluation device.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] According to a first aspect of an embodiment of the present invention, a voiceprint recognition speech evaluation method is provided, comprising the following steps:

[0009] Step S1, establishing a voiceprint speech test database;

[0010] Step S2: according to the user's voiceprint voice test requirements and test database, establish a voiceprint voice registration set and a test set corresponding to each speaker;

[0011] Step S3, sending the voiceprint speech registration sets to at least one voiceprint recognition device to be tested one by one to generate a corresponding voiceprint recognition model;

[0012] Step S4, sending the voiceprint speech test set to at least one voiceprint recognition device to be tested one by one, completing the voiceprint recognition test, and receiving the voiceprint recognition results fed back by each voiceprint recognition device to be tested;

[0013] Step S5, judging whether each of the voiceprint recognition devices to be tested has completed the voiceprint recognition test for all speech test sets; if completed, calculating the performance index of voiceprint recognition according to the voiceprint recognition result; otherwise, continuing to wait for receiving the voiceprint recognition result fed back by each of the voiceprint recognition devices to be tested.

[0014] According to a second aspect of an embodiment of the present invention, a voiceprint recognition speech evaluation method is provided, comprising the following steps:

[0015] Step S10, establishing a voiceprint speech test database;

[0016] Step S20: establishing voiceprint voice registration sets and test sets corresponding to each speaker according to the user's voiceprint voice test requirements and test database;

[0017] Step S30, playing the voiceprint speech registration set one by one, so that at least one voiceprint recognition device to be tested can pick up the voiceprint speech and generate a corresponding voiceprint recognition model;

[0018] Step S40, playing the voiceprint speech test set one by one, so that at least one of the voiceprint recognition devices to be tested can pick up the voiceprint recognition test, complete the voiceprint recognition test, and feed back the voiceprint recognition result;

[0019] Step S50, determining whether each of the voiceprint recognition devices to be tested has completed the voiceprint recognition test for all speech test sets; if completed, calculating the performance index of voiceprint recognition according to the voiceprint recognition result; otherwise, continuing to wait for receiving the voiceprint recognition result fed back by each of the voiceprint recognition devices to be tested.

[0020] Preferably, in steps S1 and S10, the voiceprint speech test database includes speech files of several speakers and annotation files corresponding to each speech file, and each speech file of each speaker corresponds to a speech content collected by a certain data collection channel.

[0021] Preferably, in steps S2 and S20, when the user's channel test requirement is a same-channel test, after selecting the voice files of each speaker that meet the user's voiceprint voice test requirements and belong to the same data acquisition channel according to the annotated files in the voiceprint voice test database, a preset number of voice files are selected from the voice files of each speaker to form the voiceprint voice registration set of the speaker, and a preset number of voice files are selected from the remaining voice files of the speaker to form the voiceprint voice test set of the speaker.

[0022] Preferably, in steps S2 and S20, when the user's channel test requirement is a cross-channel test, after selecting the voice files of each speaker that meet the user's voiceprint voice test requirements and belong to different data acquisition channels according to the annotated files in the voiceprint voice test database, a preset number of voice files are selected from the voice files of one data acquisition channel of each speaker to form the voiceprint voice registration set of the speaker, and a preset number of voice files are selected from the voice files of one of the remaining data acquisition channels of the speaker to form the voiceprint voice test set of the speaker.

[0023] Preferably, in steps S5 and S50, the optimal recognition number ZYRE in the performance index of voiceprint recognition is an index that simultaneously evaluates multiple voiceprint recognition devices to be tested from two dimensions of timeliness and accuracy, and is calculated according to the following formula:

[0024] ZYRE=T ZC ×C ZC +T RE ×A RE

[0025] In the above formula, T ZC is the normalized voiceprint registration time, C ZC is the voiceprint registration completion rate, T RE is the normalized voiceprint recognition time, A RE It is the voiceprint recognition accuracy.

[0026] Preferably, in steps S5 and S50, the voiceprint recognition accuracy A in the performance index of voiceprint recognition is RE , according to the following formula:

[0027]

[0028] E FR is the false rejection rate of each voiceprint recognition device to be tested, E FA The false acceptance rate of each voiceprint recognition device under test.

[0029] According to a third aspect of an embodiment of the present invention, there is provided a voiceprint recognition speech evaluation device, comprising a speech processor and a speech memory, wherein the speech processor reads a computer program or instruction in the speech memory to perform the following operations:

[0030] Establish a voiceprint speech test database;

[0031] According to the user's voiceprint voice test requirements and test database, establish voiceprint voice registration sets and test sets corresponding to each speaker;

[0032] Sending the voiceprint speech registration sets one by one to at least one voiceprint recognition device to be tested to generate a corresponding voiceprint recognition model;

[0033] Sending the voiceprint speech test set to at least one voiceprint recognition device to be tested one by one, completing the voiceprint recognition test, and receiving the voiceprint recognition results fed back by each voiceprint recognition device to be tested;

[0034] Determine whether each of the voiceprint recognition devices to be tested has completed the voiceprint recognition test for all speech test sets. If completed, calculate the performance index of voiceprint recognition according to the voiceprint recognition result. Otherwise, continue to wait for receiving the voiceprint recognition result fed back by each of the voiceprint recognition devices to be tested.

[0035] According to a fourth aspect of an embodiment of the present invention, there is provided a voiceprint recognition speech evaluation device, comprising a speech processor, a speech memory, a multi-channel speech conversion controller, an artificial mouth and a surround sound box; the speech processor is connected to the speech memory on one hand, and is connected to the multi-channel speech conversion controller on the other hand, and the multi-channel speech conversion controller is respectively connected to the artificial mouth and the surround sound box;

[0036] The voice processor is used to read the computer program or instruction in the voice memory to execute the establishment of a voiceprint voice test database, establish a voiceprint voice registration set and a test set corresponding to each speaker according to the user's voiceprint voice test requirements and the test database, send the voiceprint voice registration set, the test set and the environmental noise file to the multi-channel voice conversion controller according to the test requirements, and receive the voiceprint recognition result fed back by the voiceprint recognition device to be tested, and when it is determined that each of the voiceprint recognition devices to be tested has completed the voiceprint recognition test of all voice test sets, calculate the performance index of voiceprint recognition according to the voiceprint recognition result;

[0037] The multi-channel speech conversion controller is used to convert the voiceprint speech registration set, the test set and the environmental noise file into corresponding simulated audio and send them to at least one of the artificial mouths and the surround speakers respectively;

[0038] The artificial mouth is used to simulate the human voice production mode to output the audio corresponding to the received voiceprint speech registration set and test set in a full anechoic chamber environment;

[0039] The surround sound box is used to simulate the environmental noise in a specific scene according to the received environmental noise file in a full anechoic room environment.

[0040] Preferably, the multi-channel voice conversion controller includes a main control chip, a first voice conversion chip, a second voice conversion chip, a first voice audio decoding chip, a second voice audio decoding chip and a power chip; the main control chip is respectively connected to the first voice conversion chip, the second voice conversion chip, the first voice audio decoding chip, the second voice audio decoding chip and the power chip, the input ends of the first voice conversion chip and the second voice conversion chip are connected to the output end of the voice processor, the output ends of the first voice conversion chip and the second voice conversion chip are correspondingly connected to the input ends of the first voice audio decoding chip and the second voice audio decoding chip, and the output ends of the first voice audio decoding chip and the second voice audio decoding chip are correspondingly connected to the artificial mouth and the surround speakers.

[0041] The voiceprint recognition speech evaluation method and device provided by the present invention provide an operational, scientific and fair evaluation technology method for the current different voiceprint recognition methods and devices. The present invention can be applied to the evaluation of technical indicators of voiceprint recognition devices in different application scenarios and different implementation methods based on file data streams and voice data streams, thus expanding its application scope. In addition, the present invention uses an intelligent method to automatically perform statistical analysis and calculation on the voiceprint test results of the voiceprint recognition device to be tested, which greatly improves the test efficiency and test accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flow chart of a voiceprint recognition speech evaluation method provided by an embodiment of the present invention;

[0043] Figure 2 A flowchart of another voiceprint recognition speech evaluation method provided by an embodiment of the present invention;

[0044] Figure 3 A schematic diagram of the structure of a voiceprint recognition speech evaluation device provided by an embodiment of the present invention;

[0045] Figure 4 A schematic diagram of the structure of another voiceprint recognition speech evaluation device provided by an embodiment of the present invention;

[0046] Figure 5 A schematic diagram of the working process of another voiceprint recognition speech evaluation device provided by an embodiment of the present invention;

[0047] Figure 6 A circuit schematic diagram of a multi-channel voice conversion controller in another voiceprint recognition speech evaluation device provided by an embodiment of the present invention;

[0048] Figure 7A circuit schematic diagram of a main control chip of a multi-channel voice conversion controller in another voiceprint recognition speech evaluation device provided by an embodiment of the present invention;

[0049] Figure 8 A schematic diagram of the structure of an artificial mouth in another voiceprint recognition speech evaluation device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The technical content of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0051] In order to effectively evaluate the technical indicators of the voiceprint recognition method and device, an embodiment of the present invention provides a voiceprint recognition speech evaluation method, which is used to effectively evaluate the technical indicators of a voiceprint recognition device to be tested (such as a server-type product) that does not have a sound pickup function, and includes the following steps:

[0052] Step S1: Establish a voiceprint speech test database.

[0053] When establishing a voiceprint speech test database for different application scenarios, the impact of personnel, data collection channels, environmental noise, speech duration and speech content on the evaluation results is fully considered. Personnel includes the impact of gender ratio, age distribution ratio, geographical distribution, number of speakers and emotions. Data collection channels include wireless telephone channels and Internet channels. Environmental noise involves environmental noise speech data in different specific scenarios, such as railway stations, airport halls, banks, etc. Voice content includes Chinese text related to text (fixed content) and for different application scenarios and digital text unrelated to text (non-fixed content) (including 6-digit strings, 8-digit strings, 11-digit mobile phone numbers, 18-digit ID numbers). Voice duration includes short and long voices. In order to establish a voiceprint speech test database, several speakers from different regions of China were invited to participate in voice recording and obtain several voice files. The recording process was carried out in a quiet indoor environment (including background noise that does not affect voiceprint recognition), and all voice files involved the collection of several different voice collection devices.

[0054] Therefore, the voiceprint speech test database includes voice files of several speakers and annotation files corresponding to each voice file. Among them, each voice file of each speaker corresponds to a voice content collected by a certain data collection channel, so multiple voice contents collected by different data collection channels correspond to multiple voice files.

[0055] Each voice file is annotated to obtain an annotation file corresponding to the voice file. Before annotating each voice file, the voice quality of each voice file needs to be evaluated to ensure the effectiveness of the voiceprint recognition evaluation. The voice quality evaluation comprehensively evaluates the voice quality characteristics of the voice file through the existing voice quality evaluation method, mainly from the signal-to-noise ratio, voice loudness, effective voice distortion and other dimensions to perform weighted threshold evaluation. Voice files with a value less than the threshold are not allowed to enter the voiceprint voice test database. Among them, the annotation information of the annotation file corresponding to each voice file includes but is not limited to the speaker's gender, number, region, name, voice type (digital text or Chinese text), voice content, data acquisition channel, and whether the voice is synthesized. The annotation information is defined by the file name, for example: 000001-001-M-01-02-03-000004.wav, where 000001 represents the speaker's ID number, 001 represents the data collection channel number, M represents male, 01 represents single speech, 02 represents the region code, 03 represents the voice type is a digital string, and 000004 represents the voice sequence number. It should be emphasized that the voice files used for voiceprint recognition testing can be preliminarily classified and archived by annotating the file name. More detailed information can be written in the annotation file, such as: voice quality score, data validity period, acquisition equipment hardware classification and model, etc., so as to facilitate more detailed classification of voice files according to the annotation file.

[0056] Step S2: according to the user's voiceprint speech test requirements and the test database, establish a voiceprint speech registration set and a test set corresponding to each speaker.

[0057] The user's voiceprint speech test requirements include the application scenario of voiceprint recognition, the age range of the speaker, the gender ratio, the geographical scope, the channel test requirements, and the voice duration. Among them, according to the channel test requirements, the voice files for establishing the voiceprint speech registration set and the test set of each speaker are selected from the test database. Specifically, when the user's channel test requirement is the same channel test, after selecting the voice files of each speaker that meet the user's voiceprint speech test requirements and belong to the same data acquisition channel according to the annotation files in the test database, a preset number of voice files are selected from the voice files of each speaker to form the voiceprint speech registration set of the speaker, and a preset number of voice files are selected from the remaining voice files of the speaker to form the voiceprint speech test set of the speaker. For example, taking the establishment of a voiceprint speech registration set and a test set of a certain speaker as an example, assuming that the speaker has 20 voice files collected from the Internet channel, then 5 voice files are selected from the 20 voice files to form the voiceprint speech registration set of the speaker, and 2 voice files are selected from the remaining 15 voice files to form the voiceprint speech test set of the speaker.

[0058] When the user's channel test requirement is a cross-channel test, after selecting the voice files of each speaker that meet the user's voiceprint voice test requirements and belong to different data acquisition channels according to the annotated files in the test database, a preset number of voice files are selected from the voice files of one data acquisition channel of each speaker to form the voiceprint voice registration set of the speaker, and a preset number of voice files are selected from the voice files of the remaining other data acquisition channel of the speaker to form the voiceprint voice test set of the speaker.

[0059] Step S3: Send the voiceprint speech registration sets one by one to at least one voiceprint recognition device to be tested to generate a corresponding voiceprint recognition model.

[0060] When it is necessary to evaluate the technical indicators of a single voiceprint recognition device to be tested, the voiceprint speech registration sets corresponding to each speaker established in step S2 are sent to the voiceprint recognition device to be tested one by one, and the test timing is started. The voiceprint recognition device to be tested generates a voiceprint recognition model for each speaker based on all the voice files in each voiceprint speech registration set. Among them, the voiceprint recognition device to be tested generates a corresponding voiceprint recognition model based on the voiceprint speech registration set of the speaker, which is an existing mature technology and will not be repeated here.

[0061] When it is necessary to evaluate the technical indicators of multiple voiceprint recognition devices to be tested, the voiceprint speech registration set corresponding to each speaker established in step S2 is sent to each voiceprint recognition device to be tested in parallel one by one, and the test timing is started. Each voiceprint recognition device to be tested simultaneously generates a voiceprint recognition model for each speaker based on all the voice files in each voiceprint speech registration set.

[0062] Step S4: Send the voiceprint speech test set to at least one voiceprint recognition device to be tested one by one, complete the voiceprint recognition test, and receive the voiceprint recognition results fed back by each voiceprint recognition device to be tested.

[0063] Similarly, when it is necessary to evaluate the technical indicators of a single voiceprint recognition device to be tested, the voiceprint speech test sets corresponding to each speaker established in step S2 are sent to the voiceprint recognition device to be tested one by one. The voiceprint recognition device to be tested matches the voice files in each voiceprint speech test set with the voiceprint recognition models of all speakers one by one, and obtains the matching scores of each voice file used for testing of each speaker for all voiceprint recognition models as the voiceprint recognition result of the corresponding speaker and provides feedback.

[0064] When it is necessary to evaluate the technical indicators of multiple voiceprint recognition devices to be tested, the voiceprint speech test sets corresponding to each speaker established in step S2 are sent to each voiceprint recognition device to be tested in parallel one by one. Each voiceprint recognition device to be tested simultaneously matches the voice files in each voiceprint speech test set with the voiceprint recognition models of all speakers one by one, and obtains the matching scores of each voice file used for testing of each speaker for all voiceprint recognition models, which are used as the voiceprint recognition results of the corresponding speakers and are fed back.

[0065] Step S5, judging whether each voiceprint recognition device to be tested has completed the voiceprint recognition test for all speech test sets, if completed, calculating the performance index of voiceprint recognition according to the voiceprint recognition result, otherwise, continuing to wait for receiving the voiceprint recognition result fed back by each voiceprint recognition device to be tested.

[0066] If no voiceprint recognition results are received from the voiceprint recognition devices to be tested within the preset time after the last voiceprint recognition result is received from the voiceprint recognition devices to be tested, it is considered that the voiceprint recognition tests of the entire speech test set are completed by the voiceprint recognition devices to be tested, and the timing is ended at the same time, and the time for each voiceprint recognition device to be tested to complete the voiceprint recognition model and the voiceprint recognition test is calculated.

[0067] The technical indicators of the voiceprint recognition device to be tested for voiceprint recognition mainly include false rejection rate, false acceptance rate, equal error rate, error loss, minimum error loss, voiceprint registration time, voiceprint recognition time, optimal recognition number and voiceprint recognition accuracy. The calculation process of each technical indicator is described in detail below.

[0068] The voiceprint recognition device to be tested can be divided into a real speaker test case and an impersonation test case according to whether the speaker test speech and the voiceprint recognition model belong to the same speaker. In the real speaker test case, whether the speaker test speech and the voiceprint recognition model belong to the same speaker is tested, and in the impersonation test case, whether the speaker test speech and the voiceprint recognition model belong to different speakers is tested.

[0069] The false rejection rate is the proportion of the real speaker test case that the voiceprint recognition device to be tested incorrectly recognizes under a set threshold. Formula 1 is the calculation formula for the false rejection rate.

[0070]

[0071] In the above formula, E FR is the false rejection rate, N target is the total number of all real speaker test examples; N frIn order to determine the real speaker test case as the total number of the impersonation test cases, the total number of the impersonation test cases is divided according to a set threshold (set by the user, such as 95 points) and the matching scores of all voiceprint recognition models for each voice file of the speaker used for the test are obtained.

[0072] For example, 100 voiceprint recognition models in the voiceprint recognition device to be tested are tested using 100 real speakers' test voices. The test results show that 5 real speakers are incorrectly identified as impersonators. Then the false rejection rate of the voiceprint recognition device to be tested is 5 / 100=5%.

[0073] The false acceptance rate is the proportion of false recognition errors of the test case that the test device will recognize under the set threshold. Formula 2 is the calculation formula of the false acceptance rate.

[0074]

[0075] In the above formula, E FA is the false acceptance rate, N imposter is the total number of all impersonation test cases, N fa In order to determine the total number of impersonation test cases as real speaker test cases, the matching scores of all voiceprint recognition models for each voice file of the speaker used for testing are divided according to a set threshold (set by the user, such as 95 points).

[0076] For example, 100 impersonation test voices are used to test the voiceprint recognition model in the voiceprint recognition device to be tested. The test results show that 5 impersonation test voices are incorrectly recognized as real speakers. Then the false acceptance rate of the voiceprint recognition device to be tested is 5 / 100=5%.

[0077] When the false rejection rate and the false acceptance rate are equal under a certain set threshold, the equal value is the equal error rate. Under the threshold corresponding to the equal error rate, the false rejection rate, the false acceptance rate and the equal error rate have the relationship of formula 3.

[0078] EER=E FR =E FA (3)

[0079] The error loss is the total error loss cost of the voiceprint recognition device to be tested after setting the error cost weighting coefficients of the false rejection rate and the false acceptance rate. Formula 4 is the calculation formula for the error loss.

[0080] DCF=C FR *E FR *P target +C FA *E FA *(1-P target ) (4)

[0081] In the above formula, DCF is the error loss, C FR is the error cost weighting coefficient of the set false rejection rate, C FA is the error cost weighting coefficient of the set false acceptance rate, P target is the ratio of real speaker test cases to the total test cases (the matching scores of all voiceprint recognition models for each voice file of the speaker used for testing are divided according to the set threshold), (1-P target ) is the ratio of impersonation test cases to the total test cases.

[0082] The minimum error loss is the minimum value of the error losses calculated by adjusting and setting different decision thresholds.

[0083] The voiceprint registration time is the time interval from when the voiceprint recognition evaluation device sends the registration voice data to the voiceprint recognition device to be tested to when the voiceprint recognition evaluation device receives the valid registration result returned by the voiceprint recognition device to be tested.

[0084] The voiceprint recognition time is the time interval from when the voiceprint recognition evaluation device sends the test voice data to the voiceprint recognition device to be tested, to when the voiceprint recognition evaluation device receives the valid recognition completion result returned by the voiceprint recognition device to be tested.

[0085] The optimal recognition number is an indicator for simultaneously evaluating multiple voiceprint recognition devices to be tested from the two dimensions of timeliness and accuracy. Formula 5 is the calculation formula for the optimal recognition number.

[0086] ZYRE=T ZC ×C ZC +T RE ×A RE (5)

[0087] In the above formula, ZYRE is the optimal recognition number, T ZC is the normalized voiceprint registration time, C ZC is the voiceprint registration completion rate, T RE is the normalized voiceprint recognition time, A RE It is the voiceprint recognition accuracy.

[0088] When testing multiple voiceprint recognition devices to be tested, the voiceprint recognition evaluation device sends the same registration voice and test voice to the multiple voiceprint recognition devices to be tested, and obtains the voiceprint registration time, voiceprint recognition time, false rejection rate and false acceptance rate of the multiple voiceprint recognition devices to be tested. The voiceprint registration time of the multiple voiceprint recognition devices to be tested is normalized, and the voiceprint recognition registration time with the smallest value among the multiple voiceprint recognition devices to be tested is used as the reference value, and the reference value is divided by the other voiceprint recognition registration times to obtain the normalized voiceprint registration time T of each voiceprint recognition device to be tested. ZC Similarly, the normalized voiceprint recognition time T of each voiceprint recognition device to be tested can be obtained: RE .

[0089] C ZC It is the ratio of the total number of voiceprint registrations actually completed to the total number of voiceprint registrations in that time.

[0090] Under the same judgment threshold, the E of each voiceprint recognition device to be tested is FA and E FR Substitute into formula 6 to get the voiceprint recognition accuracy A RE .

[0091]

[0092] Thus, the optimal recognition number ZYRE of each voiceprint recognition device to be tested is finally obtained, and the performance of each voiceprint recognition device to be tested is determined by comparing the sizes, and the largest optimal recognition number ZYRE is the best.

[0093] like Figure 2 As shown, an embodiment of the present invention further provides a voiceprint recognition speech evaluation method, which is used to effectively evaluate the technical indicators of a voiceprint recognition device to be tested with a sound pickup function (such as a voiceprint lock, a voiceprint clock-in machine, etc.), and includes the following steps:

[0094] Step S10: Establish a voiceprint speech test database.

[0095] Step S20: According to the user's voiceprint speech test requirements and the test database, a voiceprint speech registration set and a test set corresponding to each speaker are established.

[0096] The specific implementation of steps S10 and S20 is the same as that of steps S1 and S2, which will not be repeated here.

[0097] Step S30: Play the voiceprint speech registration sets one by one, so that at least one voiceprint recognition device to be tested can pick them up and generate a corresponding voiceprint recognition model.

[0098] When it is necessary to evaluate the technical indicators of a single voiceprint recognition device to be tested, the voiceprint speech registration set corresponding to each speaker established in step S20 can be played one by one in a fully anechoic room environment, while simulating the ambient noise in a specific application scenario, and the test timing is started. The voiceprint recognition device to be tested generates a voiceprint recognition model for each speaker based on all the voice files picked up in each voiceprint speech registration set.

[0099] When it is necessary to evaluate the technical indicators of multiple voiceprint recognition devices to be tested, it is also possible to evaluate the technical indicators of multiple voiceprint recognition devices to be tested in a fully anechoic room environment, and at the same time simulate the environmental noise in a specific application scenario. For the voiceprint speech registration set corresponding to each speaker established in step S20, play the voice files in each voiceprint speech registration set one by one, and start the test timing. Each voiceprint recognition device to be tested generates a voiceprint recognition model for each speaker based on all the voice files picked up in each voiceprint speech registration set.

[0100] Step S40: Play the voiceprint speech test sets one by one, so that at least one voiceprint recognition device to be tested can pick them up, complete the voiceprint recognition test, and feed back the voiceprint recognition result.

[0101] Similarly, when it is necessary to evaluate the technical indicators of a single voiceprint recognition device to be tested, the voiceprint speech test set corresponding to each speaker established in step S20 can be evaluated in a fully anechoic room environment, while simulating the ambient noise in a specific application scenario. The voice files in each voiceprint speech test set are played one by one, and the voiceprint recognition device to be tested matches the picked-up voice files in each voiceprint speech test set with the voiceprint recognition models of all speakers one by one, and obtains the matching scores of each voice file used for testing of each speaker for all voiceprint recognition models, which are used as the voiceprint recognition results of the corresponding speakers and are fed back.

[0102] When it is necessary to evaluate the technical indicators of multiple voiceprint recognition devices to be tested, the voiceprint speech test set corresponding to each speaker established in step S20 can be played one by one in a fully anechoic room environment, in conjunction with the simulated environmental noise in a specific application scenario. Each voiceprint recognition device to be tested will match the picked-up voice files in each voiceprint speech test set with the voiceprint recognition models of all speakers one by one, and obtain the matching scores of each voice file used for testing of each speaker for all voiceprint recognition models, as the voiceprint recognition result of the corresponding speaker and provide feedback

[0103] Step S50, determine whether each voiceprint recognition device to be tested has completed the voiceprint recognition test for all speech test sets. If completed, calculate the performance index of voiceprint recognition according to the voiceprint recognition result. Otherwise, continue to wait for receiving the voiceprint recognition result fed back by each voiceprint recognition device to be tested.

[0104] If no voiceprint recognition results are received from the voiceprint recognition devices to be tested within the preset time after the last voiceprint recognition result is received from the voiceprint recognition devices to be tested, it is considered that the voiceprint recognition tests of the entire speech test set are completed by the voiceprint recognition devices to be tested, and the timing is ended at the same time, and the time for each voiceprint recognition device to be tested to complete the voiceprint recognition model and the voiceprint recognition test is calculated.

[0105] The calculation process of the relevant technical indicators is consistent with the calculation method in S5 and will not be repeated here.

[0106] It should be emphasized that in the present invention, since the process of processing data by the voiceprint recognition device to be tested takes the longest time during the test, it seriously affects the test efficiency. At the same time, the setting of the judgment threshold and weighted value has an impact on the performance index results. In order to improve the timeliness of the test, after the data processing of the voiceprint recognition device to be tested is completed, the judgment threshold and weighted scheme value can be directly changed to dynamically generate the corresponding evaluation index results.

[0107] In addition, if Figure 3 As shown, the embodiment of the present invention also provides a voiceprint recognition speech evaluation device, including a speech processor 100 and a speech memory 200, which are connected via a data bus. The device is used to effectively evaluate the technical indicators of the voiceprint recognition device to be tested without a sound pickup function. The speech processor 100 reads the computer program or instructions in the speech memory 200 to perform the following operations:

[0108] Establish a voiceprint speech test database.

[0109] According to the user's voiceprint speech test requirements and test database, establish voiceprint speech registration sets and test sets corresponding to each speaker.

[0110] The voiceprint speech registration sets are sent one by one to at least one voiceprint recognition device to be tested to generate a corresponding voiceprint recognition model.

[0111] The voiceprint speech test sets are sent to at least one voiceprint recognition device to be tested one by one to complete the voiceprint recognition test, and the voiceprint recognition results fed back by each voiceprint recognition device to be tested are received.

[0112] It is determined whether each voiceprint recognition device to be tested has completed the voiceprint recognition test for all speech test sets. If completed, the performance index of voiceprint recognition is calculated according to the voiceprint recognition result. Otherwise, the voiceprint recognition result fed back by each voiceprint recognition device to be tested is continued to be received.

[0113] In addition, if Figure 4 As shown, the embodiment of the present invention also provides a voiceprint recognition speech evaluation device, including a voice processor 300, a voice memory 400, a multi-channel voice conversion controller 500, an artificial mouth 600 and a surround sound box 700; the voice processor 300 is connected to the voice memory 400 on one hand, and is connected to the multi-channel voice conversion controller 500 on the other hand, and the multi-channel voice conversion controller 500 is respectively connected to the artificial mouth 600 and the surround sound box 700. This device is used to effectively evaluate the technical indicators of the voiceprint recognition device to be tested with a sound pickup function.

[0114] The speech processor 100 is used to read the computer program or instructions in the speech memory 200 to execute the establishment of a voiceprint speech test database, establish a voiceprint speech registration set and a test set corresponding to each speaker according to the user's voiceprint speech test requirements and the test database, send the voiceprint speech registration set, the test set and the environmental noise file to the multi-channel speech conversion controller 500 according to the test requirements, and receive the voiceprint recognition result fed back by the voiceprint recognition device to be tested (such as Figure 5 As shown), when it is determined that each voiceprint recognition device to be tested has completed the voiceprint recognition test for all speech test sets, the performance index of the voiceprint recognition is calculated according to the voiceprint recognition result.

[0115] The multi-channel speech conversion controller 500 is used to convert the registration set, the test set and the environmental noise file into corresponding simulated audio and send them to at least one artificial mouth 600 and the surround sound box 700 respectively.

[0116] The artificial mouth 600 is used to simulate the human voice in the anechoic room environment to output the audio corresponding to the received registration set and the test set, so as to simulate the human voice for testing. Figure 5 shown.

[0117] The surround sound box 700 is used to simulate the environmental noise in a specific scene according to the received environmental noise file in a full anechoic room environment.

[0118] like Figure 6As shown, the multi-channel voice conversion controller 500 includes a main control chip 5001, a first voice conversion chip 5002, a second voice conversion chip 5003, a first voice audio decoding chip 5004, a second voice audio decoding chip 5005 and a power chip 5006; the main control chip 5001 is respectively connected to the first voice conversion chip 5002, the second voice conversion chip 5003, the first voice audio decoding chip 5004, the second voice audio decoding chip 5005 and the power chip 5006, the input end of the first voice conversion chip 5002 and the second voice conversion chip 5003 are connected to the output end of the voice processor 100, the output end of the first voice conversion chip 5002 and the second voice conversion chip 5003 is correspondingly connected to the input end of the first voice audio decoding chip 5004 and the second voice audio decoding chip 5005, and the output end of the first voice audio decoding chip 5004 and the second voice audio decoding chip 5005 is correspondingly connected to the artificial mouth 600 and the surround sound box 700.

[0119] Specifically, the main control chip 5001 can be implemented using an MSP430FG461x series microcontroller, the first voice conversion chip 5002 and the second voice conversion chip 5003 can be implemented using an SSM2404S series voice conversion chip, and the first voice audio decoding chip 5004 and the second voice audio decoding chip 5005 can be implemented using a TPA3116D2DADR series voice audio decoding chip.

[0120] Among them, the first voice conversion chip 5002 and the second voice conversion chip 5003 are multi-channel inputs, which are used to receive voice files of voiceprint voice registration sets and test sets of multiple speakers and environmental noise files in multiple specific scenarios, that is, each input of the first voice conversion chip 5002 corresponds to receiving voice files of voiceprint voice registration sets and test sets of one speaker. Each input of the second voice conversion chip 5003 corresponds to receiving an environmental noise file in a specific scenario. The first voice conversion chip 5002 is used to convert the received voice files of voiceprint voice registration sets and test sets into corresponding digital audio; the second voice conversion chip 5003 is used to convert the environmental noise file into corresponding digital audio.

[0121] The first voice and audio decoding chip 5004 and the second voice and audio decoding chip are multi-channel outputs, used to convert the corresponding digital audio into analog audio, and send it to the corresponding artificial mouth 600 and surround sound box 700.

[0122] The main control chip 5001 is mainly used to control the operation of the first voice conversion chip 5002, the second voice conversion chip 5003, the first voice audio decoding chip 5004 and the second voice audio decoding chip 5005 and set corresponding parameters, and also provides power for the first voice conversion chip 5002, the second voice conversion chip 5003, the first voice audio decoding chip 5004 and the second voice audio decoding chip 5005. Its circuit structure is as follows Figure 7 shown.

[0123] like Figure 8 As shown, the artificial mouth 600 includes a cavity 6001, in which a driver 6002, a speaker 6003, and a sound-conducting microphone 6004 are arranged. A data interface 6005 for receiving analog audio sent by the multi-channel speech conversion controller 500 is arranged on the side of the outside of the cavity 6001 close to the driver 6002, a lip ring 6006 is arranged on the outside of the front mask of the cavity 6001, and a fixed bracket interface 6008 is arranged at the bottom of the outside of the cavity 6001. The speaker 6003 used in the artificial mouth 600 is a standard sound source (speaker) with stable, wide-band response and low distortion, so as to accurately simulate the characteristics of the human voice sound field. The driver 6002 is installed in a special sound cavity that is approximately closed, and can simulate the average acoustic characteristics of the simulated human mouth, including acoustic damping coefficient, acoustic radiation characteristics and transmission characteristics. The shape of the front mask 6007 and the sound outlet hole should take into account the influence of the reflection phenomenon between the nasal cavity, the human face and the human lips when the human body makes sounds. Therefore, the front mask 6007 is designed to be a convex reflection surface, and the sound outlet hole simulates the human mouth and nasal cavity. The lip circle 6006 plane of the sound outlet hole is a reference plane to ensure that the attenuation of sound pressure with distance starting from this plane is close to the attenuation of sound pressure with distance starting from the human lip plane when the human body makes sounds.

[0124] In the present invention, the voiceprint recognition device to be tested can be connected to the voiceprint recognition speech evaluation device in two ways, remote and local, to complete the evaluation and verification of the technical indicators of the voiceprint recognition function. The remote evaluation method is implemented in the application interface mode; the local evaluation method is implemented in the dynamic library mode. The application interface mode requires the IP address and port number of the voiceprint recognition device to be tested, and is tested after connecting to the voiceprint recognition speech evaluation device through the Internet. The dynamic library mode directly copies the dynamic library of the voiceprint recognition device to be tested to the specific directory of the voiceprint recognition speech evaluation device for testing.

[0125] In addition, the present invention can also be applied to the detection of related intelligent voice equipment, such as intelligent voiceprint locks, voiceprint sign-in machines, security products with voiceprint recognition and verification functions, voice recognition performance detection, and semantic device performance detection.

[0126] In addition, an embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, and when the computer-readable storage medium is run on the computer, the computer executes the above-mentioned Figure 1 and Figure 2 The specific implementation of the voiceprint recognition speech evaluation method will not be described here.

[0127] In addition, an embodiment of the present invention further provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute the above-mentioned Figure 1 and Figure 2 The specific implementation of the voiceprint recognition speech evaluation method will not be described here.

[0128] The voiceprint recognition speech evaluation method and device provided by the present invention provide an operational, scientific and fair evaluation technology method for the current different voiceprint recognition methods and devices. The present invention can be applied to the evaluation of technical indicators of voiceprint recognition devices in different application scenarios and different implementation methods based on file data streams and voice data streams, thus expanding its application scope. In addition, the present invention uses an intelligent method to automatically perform statistical analysis and calculation on the voiceprint test results of the voiceprint recognition device to be tested, which greatly improves the test efficiency and test accuracy.

[0129] The above is a detailed description of the voiceprint recognition speech evaluation method and device provided by the present invention. For those skilled in the art, any obvious changes made to it without departing from the essence of the present invention will fall within the scope of protection of the patent right of the present invention.

Claims

1. A voiceprint recognition speech evaluation method, characterized in that The steps include: Step S1, establishing a voiceprint speech test database; Step S2: according to the user's voiceprint voice test requirements and test database, establish a voiceprint voice registration set and a test set corresponding to each speaker; Step S3, sending the voiceprint speech registration sets to at least one voiceprint recognition device to be tested one by one to generate a corresponding voiceprint recognition model; Step S4, sending the voiceprint speech test set to at least one voiceprint recognition device to be tested one by one, completing the voiceprint recognition test, and receiving the voiceprint recognition results fed back by each voiceprint recognition device to be tested; Step S5: determine whether each of the voiceprint recognition devices to be tested has completed the voiceprint recognition test on all the voice test sets. If completed, calculate the performance index of voiceprint recognition according to the voiceprint recognition result. Otherwise, continue to wait for receiving the voiceprint recognition result fed back by each of the voiceprint recognition devices to be tested. In step S5, the optimal recognition number among the performance indicators of voiceprint recognition is It is an indicator for simultaneously evaluating multiple voiceprint recognition devices to be tested from the two dimensions of timeliness and accuracy, and is calculated according to the following formula; In the above formula, is the normalized voiceprint registration time, is the voiceprint registration completion rate, is the normalized voiceprint recognition time, is the voiceprint recognition accuracy; Voiceprint recognition accuracy in voiceprint recognition performance indicators , according to the following formula: is the false rejection rate of each voiceprint recognition device to be tested, ,in, is the total number of all real speaker test examples; The total number of real speaker test cases judged as impersonation test cases; is the false acceptance rate of each voiceprint recognition device to be tested, ,in, is the total number of all impersonation test cases, is the total number of impersonation test cases that are judged as real speaker test cases.

2. A voiceprint recognition speech evaluation method, characterized in that The steps include: Step S10, establishing a voiceprint speech test database; Step S20: establishing voiceprint voice registration sets and test sets corresponding to each speaker according to the user's voiceprint voice test requirements and test database; Step S30, playing the voiceprint speech registration set one by one, so that at least one voiceprint recognition device to be tested can pick up the voiceprint speech and generate a corresponding voiceprint recognition model; Step S40, playing the voiceprint speech test set one by one, so that at least one of the voiceprint recognition devices to be tested can pick up the voiceprint recognition test, complete the voiceprint recognition test, and feed back the voiceprint recognition result; Step S50: determine whether each of the voiceprint recognition devices to be tested has completed the voiceprint recognition test on all the voice test sets. If completed, calculate the performance index of voiceprint recognition according to the voiceprint recognition result. Otherwise, continue to wait for receiving the voiceprint recognition result fed back by each of the voiceprint recognition devices to be tested. In step S50, the optimal recognition number among the performance indicators of voiceprint recognition is It is an indicator for simultaneously evaluating multiple voiceprint recognition devices to be tested from the two dimensions of timeliness and accuracy, and is calculated according to the following formula; In the above formula, is the normalized voiceprint registration time, is the voiceprint registration completion rate, is the normalized voiceprint recognition time, is the voiceprint recognition accuracy; Voiceprint recognition accuracy in voiceprint recognition performance indicators , according to the following formula: is the false rejection rate of each voiceprint recognition device to be tested, ,in, is the total number of all real speaker test examples; The total number of real speaker test cases judged as impersonation test cases; is the false acceptance rate of each voiceprint recognition device to be tested, ,in, is the total number of all impersonation test cases, is the total number of impersonation test cases that are judged as real speaker test cases.

3. The voiceprint recognition speech evaluation method according to claim 1 or 2, characterized in that In steps S1 and S10, The voiceprint speech test database includes speech files of several speakers and annotation files corresponding to each speech file, and each speech file of each speaker corresponds to a piece of speech content collected by a certain data collection channel.

4. The voiceprint recognition speech evaluation method according to claim 1 or 2, characterized in that In steps S2 and S20, When the user's channel test requirement is a same-channel test, after selecting the voice files of each speaker that meet the user's voiceprint voice test requirements and belong to the same data acquisition channel according to the annotated files in the voiceprint voice test database, a preset number of voice files are selected from the voice files of each speaker to form the voiceprint voice registration set of the speaker, and a preset number of voice files are selected from the remaining voice files of the speaker to form the voiceprint voice test set of the speaker.

5. The voiceprint recognition speech evaluation method according to claim 1 or 2, characterized in that In steps S2 and S20, When the user's channel test requirement is a cross-channel test, after selecting the voice files of each speaker that meet the user's voiceprint voice test requirements and belong to different data acquisition channels according to the annotated files in the voiceprint voice test database, a preset number of voice files are selected from the voice files of one data acquisition channel of each speaker to form the voiceprint voice registration set of the speaker, and a preset number of voice files are selected from the voice files of one of the remaining data acquisition channels of the speaker to form the voiceprint voice test set of the speaker.

6. The voiceprint recognition speech evaluation method according to claim 1 or 2, characterized in that: The performance indicators of voiceprint recognition also include equal error rate, Under the threshold corresponding to the equal error rate, the false rejection rate, the false acceptance rate and the equal error rate satisfy: .

7. The voiceprint recognition speech evaluation method according to claim 1 or 2, characterized in that: The performance index of the voiceprint recognition also includes a minimum error loss, which is the minimum value of the error losses calculated by adjusting and setting different judgment thresholds; The error loss is the overall error loss cost of the voiceprint recognition device to be tested after setting the error cost weighting coefficients of the false rejection rate and the false acceptance rate. The calculation formula of the error loss is: In the formula, is the error loss, is the error cost weighting coefficient of the set false rejection rate, is the error cost weighting coefficient of the set false acceptance rate, is the ratio of real speaker test cases to the total test cases, It is the ratio of fake test cases to the total test cases.

8. A voiceprint recognition speech evaluation device, characterized in that The invention comprises a speech processor and a speech memory, wherein the speech processor reads a computer program or instruction in the speech memory to perform the following operations: Establish a voiceprint speech test database; According to the user's voiceprint voice test requirements and test database, establish voiceprint voice registration sets and test sets corresponding to each speaker; Sending the voiceprint speech registration sets one by one to at least one voiceprint recognition device to be tested to generate a corresponding voiceprint recognition model; Sending the voiceprint speech test set to at least one voiceprint recognition device to be tested one by one, completing the voiceprint recognition test, and receiving the voiceprint recognition results fed back by each voiceprint recognition device to be tested; Determine whether each of the voiceprint recognition devices to be tested has completed the voiceprint recognition test of all the voice test sets. If completed, calculate the performance index of voiceprint recognition according to the voiceprint recognition result. Otherwise, continue to wait for receiving the voiceprint recognition result fed back by each of the voiceprint recognition devices to be tested. The optimal recognition number in the performance index of the voiceprint recognition It is an indicator for simultaneously evaluating multiple voiceprint recognition devices to be tested from the two dimensions of timeliness and accuracy, and is calculated according to the following formula; In the above formula, is the normalized voiceprint registration time, is the voiceprint registration completion rate, is the normalized voiceprint recognition time, is the voiceprint recognition accuracy; Voiceprint recognition accuracy in voiceprint recognition performance indicators , according to the following formula: is the false rejection rate of each voiceprint recognition device to be tested, ,in, is the total number of all real speaker test examples; The total number of real speaker test cases judged as impersonation test cases; is the false acceptance rate of each voiceprint recognition device to be tested, ,in, is the total number of all impersonation test cases, is the total number of impersonation test cases that are judged as real speaker test cases.

9. A voiceprint recognition speech evaluation device, characterized in that It includes a speech processor, a speech memory, a multi-channel speech conversion controller, an artificial mouth and a surround sound box; the speech processor is connected to the speech memory on one hand, and is connected to the multi-channel speech conversion controller on the other hand, and the multi-channel speech conversion controller is respectively connected to the artificial mouth and the surround sound box; The voice processor is used to read the computer program or instruction in the voice memory to execute the establishment of a voiceprint voice test database, establish a voiceprint voice registration set and a test set corresponding to each speaker according to the user's voiceprint voice test requirements and the test database, send the voiceprint voice registration set, the test set and the environmental noise file to the multi-channel voice conversion controller according to the test requirements, and receive the voiceprint recognition result fed back by the voiceprint recognition device to be tested, and when it is determined that each of the voiceprint recognition devices to be tested has completed the voiceprint recognition test of all voice test sets, calculate the performance index of voiceprint recognition according to the voiceprint recognition result; The multi-channel speech conversion controller is used to convert the voiceprint speech registration set, the test set and the environmental noise file into corresponding simulated audio and send them to at least one of the artificial mouths and the surround speakers respectively; The artificial mouth is used to simulate the human voice production mode to output the audio corresponding to the received voiceprint speech registration set and test set in a full anechoic chamber environment; The surround sound box is used to simulate the environmental noise in a specific scene according to the received environmental noise file in a full anechoic room environment. The front mask of the artificial mouth is a convex reflective surface, the sound outlet simulates the human mouth and nasal cavity, and the lip circle plane of the sound outlet is a reference plane to ensure that the attenuation of sound pressure with distance from this plane is similar to the attenuation of sound pressure with distance from the human lip plane when the human body makes a sound. The optimal recognition number in the performance index of the voiceprint recognition It is an indicator for simultaneously evaluating multiple voiceprint recognition devices to be tested from the two dimensions of timeliness and accuracy, and is calculated according to the following formula; In the above formula, is the normalized voiceprint registration time, is the voiceprint registration completion rate, is the normalized voiceprint recognition time, is the voiceprint recognition accuracy; Voiceprint recognition accuracy in voiceprint recognition performance indicators , according to the following formula: is the false rejection rate of each voiceprint recognition device to be tested, ,in, is the total number of all real speaker test examples; The total number of real speaker test cases judged as impersonation test cases; is the false acceptance rate of each voiceprint recognition device to be tested, ,in, is the total number of all impersonation test cases, is the total number of impersonation test cases that are judged as real speaker test cases.

10. The voiceprint recognition speech evaluation device according to claim 9, characterized in that: The multi-channel voice conversion controller includes a main control chip, a first voice conversion chip, a second voice conversion chip, a first voice audio decoding chip, a second voice audio decoding chip and a power chip; the main control chip is respectively connected to the first voice conversion chip, the second voice conversion chip, the first voice audio decoding chip, the second voice audio decoding chip and the power chip, the input ends of the first voice conversion chip and the second voice conversion chip are connected to the output end of the voice processor, the output ends of the first voice conversion chip and the second voice conversion chip are correspondingly connected to the input ends of the first voice audio decoding chip and the second voice audio decoding chip, and the output ends of the first voice audio decoding chip and the second voice audio decoding chip are correspondingly connected to the artificial mouth and the surround sound box.

Images