Voice reconstruction result evaluation method, device and equipment and storage medium thereof

Abstract

The embodiment of the invention belongs to the technical field of research and development design and audio processing, is applied to an audio reconstruction result evaluation scene, and relates to a voice reconstruction result evaluation method, device and equipment and a storage medium thereof. Therefore, the effect of converting the source voice into the target voice can be evaluated from the main information part or the pronunciation emphasizing part in the voice. According to the method, the voice reconstruction result is evaluated by only combining related accent features and neglecting slightly-affected genuine features, so that the feature data volume during result evaluation is reduced, it can be ensured that evaluation focuses on key features, reconstruction result evaluation is performed by only using relatively important voice features, and the reconstruction evaluation efficiency is improved. When the method is applied to an intelligent customer service voice reconstruction scene, the intelligent customer service reconstruction voice with the quality reaching the standard can be sent to a related target customer more quickly, and the service efficiency is improved.

Description

Technical Field

[0001] The present application relates to the field of R&D design and audio processing technology, and is applied to audio reconstruction result evaluation scenarios, and in particular to a speech reconstruction result evaluation method, device, equipment and storage medium thereof. Background Art

[0002] In recent years, existing neural vocoder technology has made significant progress in speech synthesis and voice conversion. However, current technology still has corresponding shortcomings in speech synthesis or speech reconstruction evaluation.

[0003] For example, some existing speech quality assessment standards, such as MOS (Mean Opinion Score), focus more on the sound quality and clarity of speech, which require the combination of a large number of acoustic features for assessment, resulting in complexity and low efficiency of the assessment. Therefore, the prior art still lacks a method for evaluating the speech reconstruction effect by combining only a small number of important acoustic features, so as to reduce the complexity of the assessment and improve the efficiency of speech reconstruction assessment. Summary of the invention

[0004] The purpose of the embodiments of the present application is to propose a method, device, equipment and storage medium for evaluating speech reconstruction results, so as to solve the problem that the existing speech quality evaluation method needs to combine a large number of acoustic features, resulting in complexity and low efficiency of the evaluation.

[0005] In order to solve the above technical problems, the embodiment of the present application provides a method for evaluating speech reconstruction results, which adopts the following technical solution:

[0006] A method for evaluating speech reconstruction results comprises the following steps:

[0007] Get the source voice;

[0008] Capturing the prosodic features of the source speech through a preset prosodic feature extraction network;

[0009] Inputting the prosodic features into a speech signal reconstruction network to obtain a reconstructed target speech;

[0010] Capturing the prosodic features of the target speech through the prosodic feature extraction network;

[0011] According to a preset screening and extraction strategy, respectively extracting stress features from the prosodic features of the source speech and the prosodic features of the target speech to obtain stress feature extraction results;

[0012] The accent feature extraction result is input into a discriminator based on a generative adversarial network to discriminate the transfer effect of the accent feature from the source speech to the target speech.

[0013] Furthermore, before executing the step of capturing the prosodic features of the source speech through a preset prosodic feature extraction network, the method further includes:

[0014] Obtaining a sound waveform change diagram of the source speech by an oscilloscope method;

[0015] Using a sound level meter, measuring a pressure change graph of the source speech;

[0016] According to a preset syllable-phoneme comparison table, identifying all syllables contained in the source speech and the phoneme sequence contained in each syllable;

[0017] The step of capturing the prosodic features of the source speech through a preset prosodic feature extraction network specifically includes:

[0018] The prosodic feature extraction network is used to capture the sound frequency values ​​and sound amplitude values ​​corresponding to all pronunciation periods of the source speech and the pronunciation duration of all syllables in the source speech;

[0019] The sound frequency values, sound amplitude values ​​and pronunciation durations of all syllables corresponding to the source speech in all pronunciation time periods are used as the prosodic features of the source speech.

[0020] Furthermore, the step of capturing the sound frequency values, sound amplitude values ​​and pronunciation durations of all syllables in the source speech corresponding to all pronunciation periods of the source speech through the prosodic feature extraction network specifically includes:

[0021] Inputting the sound waveform change graph and the pressure change graph into the rhythm feature extraction network;

[0022] According to the sampling time interval preset in the prosodic feature extraction network, the sound waveform change is sampled and processed to obtain the sound frequency values ​​corresponding to the source speech in all unit time periods;

[0023] According to the sampling time interval, the pressure change graph is sampled to obtain the sound amplitude values ​​corresponding to the source speech in all unit time periods;

[0024] Capturing the pronunciation durations corresponding to all phonemes in the source speech respectively through the prosodic feature extraction network;

[0025] The pronunciation durations of all the syllables are obtained according to the phoneme sequences respectively contained in each syllable and the pronunciation duration corresponding to each phoneme.

[0026] Furthermore, the step of extracting stress features from the prosodic features of the source speech and the prosodic features of the target speech respectively according to a preset screening and extraction strategy to obtain stress feature extraction results specifically includes:

[0027] Compare the sound frequency values ​​corresponding to the source voice in all unit time periods with the preset sound frequency thresholds one by one, and filter out all sound frequency values ​​exceeding the sound frequency threshold;

[0028] Compare the sound amplitude values ​​corresponding to the source voice in all unit time periods with the preset sound amplitude threshold value one by one, and filter out all sound amplitude values ​​exceeding the sound amplitude threshold value;

[0029] Compare the pronunciation durations of all syllables in the source speech with the preset pronunciation durations one by one, and filter out all syllables that exceed the preset pronunciation duration;

[0030] Taking all the sound frequency values ​​exceeding the sound frequency threshold, all the sound amplitude values ​​exceeding the sound amplitude threshold, and all the syllables exceeding the preset pronunciation duration as the stress feature extraction results of the source speech;

[0031] Similarly, the stress features of the target speech are screened out from the prosodic features of the target speech.

[0032] Furthermore, the step of inputting the stress feature extraction result into a discriminator based on a generative adversarial network to discriminate the transfer effect of the stress feature from the source voice to the target voice specifically includes:

[0033] Inputting the stress feature extraction result of the source speech and the stress feature extraction result of the target speech into the discriminator;

[0034] Performing multi-dimensional discrimination on the stress feature extraction result of the source speech and the stress feature extraction result of the target speech according to the multi-dimensional recognition component in the discriminator to obtain a multi-dimensional discrimination result;

[0035] Based on the multi-dimensional discrimination result, a comprehensive discrimination is performed to obtain the transfer effect of the accent feature from the source voice to the target voice.

[0036] Furthermore, the step of performing multi-dimensional discrimination on the stress feature extraction result of the source speech and the stress feature extraction result of the target speech according to the multi-dimensional recognition component in the discriminator to obtain the multi-dimensional discrimination result specifically includes:

[0037] Performing consistency comparison on all the sound frequency values ​​in the source speech exceeding the sound frequency threshold and all the sound frequency values ​​in the target speech exceeding the sound frequency threshold, and obtaining a consistency representation value of the target speech compared with the source speech as a first comparison result;

[0038] Performing consistency comparison on all the sound amplitude values ​​in the source speech exceeding the sound amplitude threshold and all the sound amplitude values ​​in the target speech exceeding the sound amplitude threshold, and obtaining a consistency representation value of the target speech compared with the source speech as a second comparison result;

[0039] Comparing all syllables in the source speech that exceed the preset pronunciation duration with all syllables in the target speech that exceed the preset pronunciation duration, and obtaining a consistency representation value of the target speech compared with the source speech as a third comparison result;

[0040] The first comparison result, the second comparison result and the third comparison result are used as the multi-dimensional discrimination result.

[0041] Furthermore, the step of performing comprehensive discrimination based on the multi-dimensional discrimination result to obtain the transfer effect of the accent feature from the source voice to the target voice specifically includes:

[0042] Obtaining comprehensive weight values ​​pre-set for the multi-dimensional discrimination results;

[0043] performing multiplication operations on the first comparison result, the second comparison result, and the third comparison result with corresponding comprehensive weight values ​​respectively to obtain multiplication results;

[0044] For the multiplication result, a comprehensive discrimination result is obtained by accumulating and averaging the result;

[0045] By comparison, determining whether the comprehensive discrimination result exceeds a preset comprehensive discrimination threshold;

[0046] If the comprehensive discrimination result exceeds the preset comprehensive discrimination threshold, the conversion from the source speech to the target speech meets the speech reconstruction requirement;

[0047] If the comprehensive judgment result does not exceed the preset comprehensive judgment threshold, the conversion from the source speech to the target speech does not meet the speech reconstruction requirements, and the network parameters of the speech signal reconstruction network are tuned until the conversion from the source speech to the target speech meets the speech reconstruction requirements.

[0048] In order to solve the above technical problems, the embodiment of the present application also provides a speech reconstruction result evaluation device, which adopts the following technical solution:

[0049] A speech reconstruction result evaluation device, comprising:

[0050] A source voice acquisition module, used to acquire source voice;

[0051] A first prosodic feature capturing module, used to capture the prosodic features of the source speech through a preset prosodic feature extraction network;

[0052] A target speech reconstruction module, used for inputting the prosodic features into a speech signal reconstruction network to obtain a reconstructed target speech;

[0053] A second prosodic feature capturing module, used for capturing the prosodic features of the target speech through the prosodic feature extraction network;

[0054] An accent feature screening and extraction module is used to extract accent features from the prosodic features of the source speech and the prosodic features of the target speech according to a preset screening and extraction strategy, and obtain an accent feature extraction result;

[0055] The stress feature transfer effect discrimination module is used to input the stress feature extraction result into a discriminator based on a generative adversarial network to discriminate the transfer effect of the stress feature from the source speech to the target speech.

[0056] In order to solve the above technical problems, the embodiment of the present application further provides a computer device, which adopts the following technical solution:

[0057] A computer device comprises a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the above-mentioned method for evaluating speech reconstruction results when executing the computer-readable instructions.

[0058] In order to solve the above technical problems, the embodiment of the present application further provides a computer-readable storage medium, which adopts the following technical solution:

[0059] A computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the method for evaluating speech reconstruction results as described above.

[0060] Compared with the prior art, the embodiments of the present application have the following beneficial effects:

[0061] The speech reconstruction result evaluation method described in the embodiment of the present application obtains the source speech; captures the prosodic features of the source speech through a preset prosodic feature extraction network; inputs the prosodic features into the speech signal reconstruction network to obtain the reconstructed target speech and capture the prosodic features of the target speech; extracts the stress features of the prosodic features of the source speech and the prosodic features of the target speech; inputs the stress feature extraction results into a discriminator based on a generative adversarial network to discriminate the transmission effect of the stress features from the source speech to the target speech. By extracting the stress features in the prosodic features of the source speech and the prosodic features of the target speech, it is convenient to evaluate the effect of converting the source speech to the target speech from the main information part or the pronunciation emphasis part of the speech. Only the relevant stress features are combined, and the unstressed features with little influence are ignored to evaluate the speech reconstruction results, which not only reduces the amount of feature data during the result evaluation, but also ensures that the evaluation focuses more on the key features, and realizes the evaluation of the reconstruction results using only relatively important speech features, thereby improving the efficiency of reconstruction evaluation. When applied to the scenario of intelligent customer service voice reconstruction, it can send intelligent customer service reconstructed voice that meets quality standards to relevant target customers more quickly, thereby improving business efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the scheme in the present application, a brief introduction is given below to the drawings required for use in the description of the embodiments of the present application. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0063] Figure 1 is an exemplary system architecture diagram to which the present application may be applied;

[0064] Figure 2 is a flow chart of an embodiment of a method for evaluating speech reconstruction results according to the present application;

[0065] Figure 3 It is a flowchart of a specific embodiment of preprocessing the source speech in the speech reconstruction result evaluation method described in the present application;

[0066] Figure 4 yes Figure 2 A flowchart of a specific embodiment of step 202 is shown;

[0067] Figure 5 yes Figure 4 A flowchart of a specific embodiment of step 401 is shown;

[0068] Figure 6 yes Figure 2 A flowchart of a specific embodiment of step 205 is shown;

[0069] Figure 7 yes Figure 2 A flowchart of a specific embodiment of step 206 is shown;

[0070] Figure 8 yes Figure 7 A flowchart of a specific embodiment of step 702 is shown;

[0071] Fig. 9 yes Figure 7 A flowchart of a specific embodiment of step 703 is shown;

[0072] Fig.10 is a structural schematic diagram of an embodiment of a speech reconstruction result evaluation device according to the present application;

[0073] Fig.11 It is a structural diagram of an embodiment of a computer device according to the present application. DETAILED DESCRIPTION

[0074] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by technicians in the technical field of the present application; the terms used in the specification of the application herein are only for the purpose of describing specific embodiments and are not intended to limit the present application; the terms "including" and "having" and any variations thereof in the specification and claims of the present application and the above-mentioned drawings are intended to cover non-exclusive inclusions. The terms "first", "second", etc. in the specification and claims of the present application or the above-mentioned drawings are used to distinguish different objects, not to describe a specific order.

[0075] Reference to "embodiments" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present application. The appearance of the phrase in various locations in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0076] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings.

[0077] like Figure 1As shown, the system architecture 100 may include a terminal device 101, a network 102 and a server 103. The terminal device 101 may be a laptop 1011, a tablet computer 1012 or a mobile phone 1013. The network 102 is used to provide a medium for a communication link between the terminal device 101 and the server 103. The network 102 may include various connection types, such as wired, wireless communication links or optical fiber cables.

[0078] The user can use the terminal device 101 to interact with the server 103 through the network 102 to receive or send messages, etc. Various communication client applications can be installed on the terminal device 101, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social platform software, etc.

[0079] The terminal device 101 can be any electronic device with a display screen and supporting web browsing. In addition to a laptop computer 1011, a tablet computer 1012 or a mobile phone 1013, the terminal device 101 can also be an e-book reader, an MP3 player (Moving Picture Experts Group Audio Layer III), an MP4 (Moving Picture Experts Group Audio Layer IV), a laptop computer, a desktop computer, etc.

[0080] The server 103 may be a server that provides various services, such as a background server that provides support for a web page displayed on the terminal device 101 .

[0081] It should be noted that the speech reconstruction result evaluation method provided in the embodiment of the present application is generally executed by a server, and accordingly, the speech reconstruction result evaluation device is generally arranged in the server.

[0082] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0083] Continue to refer Figure 2 , shows a flow chart of an embodiment of a method for evaluating speech reconstruction results according to the present application. The method for evaluating speech reconstruction results comprises the following steps:

[0084] Step 201, obtaining source speech.

[0085] In this embodiment, the source voice includes the original voice uttered by the speaker or the speaker, for example: a speech by the speaker, a song sung by the speaker, etc.; or, a product introduction voice when a manual customer service provides voice service.

[0086] Step 202: Capture the prosodic features of the source speech through a preset prosodic feature extraction network.

[0087] In this embodiment, the prosodic feature extraction network includes a speech feature extraction network based on RNN, which can at least extract corresponding prosodic features from the source speech, and the prosodic features include sound frequency values, sound amplitude values ​​and pronunciation durations of all syllables, and may further include word information, phrase information, pronunciation rhythm, timbre information, etc. in the source speech. More specifically, the prosodic features can be divided into three categories: stress features, unstress features or middle features, or only divided into two categories: stress features and unstress features.

[0088] By capturing the prosodic features of the source speech, the subsequent processing steps are facilitated and the relevant information in the prosodic features is used to evaluate the speech reconstruction results.

[0089] Step 203: input the prosodic features into a speech signal reconstruction network to obtain a reconstructed target speech.

[0090] In this embodiment, the speech signal reconstruction network can be a speech codec processing network based on a transformer neural network architecture. The prosodic features are input into the speech codec processing network based on a transformer neural network architecture as a type of feature when speech decoding is generated, so as to facilitate target speech reconstruction based on the prosodic features.

[0091] Step 204: capturing the prosodic features of the target speech through the prosodic feature extraction network.

[0092] In this embodiment, the prosodic features of the target speech may be captured by using the same prosodic feature extraction network as in step 202. By using the same prosodic feature extraction network to capture the prosodic features of the source speech and the target speech respectively, the reliability and persuasiveness of the speech reconstruction result evaluation can be ensured.

[0093] Step 205 , according to a preset screening and extraction strategy, respectively extracting accent features from the prosodic features of the source speech and the prosodic features of the target speech to obtain accent feature extraction results.

[0094] In this embodiment, the stress feature is often applied to the main information part and the pronunciation emphasis part of the speech according to the pronunciation habit. The stress feature includes a pronunciation segment with a higher pitch, i.e., a higher sound frequency value, a pronunciation segment with a higher sound intensity (decibel), i.e., a higher sound amplitude value, and a syllable with a longer pronunciation duration under equal phonemes. In general, the stress in pronunciation often corresponds to a syllable with a higher pitch, a larger sound intensity, and a longer pronunciation duration.

[0095] By extracting the prosodic features of the source speech and the stress features in the prosodic features of the target speech, the effect of converting the source speech into the target speech can be evaluated from the main information part or the pronunciation emphasis part in the speech.

[0096] Step 206: input the stress feature extraction result into a discriminator based on a generative adversarial network to discriminate the transfer effect of the stress feature from the source speech to the target speech.

[0097] A discriminator based on a generative adversarial network is used to discriminate the transmission effect of the accent feature from the source speech to the target speech. This can not only evaluate and discriminate the speech reconstruction result, but also reversely optimize and adjust the speech signal reconstruction network in step 203 based on the evaluation and discrimination result, thereby achieving the goal of not only evaluating the speech reconstruction effect by using the accent feature, but also adjusting and optimizing the speech signal reconstruction network.

[0098] The speech reconstruction result evaluation method described in this application is applied to the intelligent customer service speech reconstruction scenario, so that when evaluating the reconstruction result of the intelligent customer service speech, only the relevant stress features are combined, and the unstressed features with little influence are ignored, which reduces the amount of feature data during the result evaluation and ensures that the evaluation focuses more on key features, so that only relatively important speech features are used to evaluate the reconstruction results, thereby improving the efficiency of reconstruction evaluation. The intelligent customer service reconstructed speech that meets the quality standards can be sent to relevant target customers more quickly.

[0099] In this embodiment, the source speech is obtained; the prosodic features of the source speech are captured through a preset prosodic feature extraction network; the prosodic features are input into a speech signal reconstruction network to obtain a reconstructed target speech, and the prosodic features of the target speech are captured; the prosodic features of the source speech and the prosodic features of the target speech are extracted with accent features; the accent feature extraction result is input into a discriminator based on a generative adversarial network to discriminate the transmission effect of the accent feature from the source speech to the target speech. By extracting the accent features in the prosodic features of the source speech and the prosodic features of the target speech, the effect of converting the source speech to the target speech is evaluated from the main information part or the pronunciation emphasis part of the speech. Only the relevant accent features are combined, and the light tone features with little influence are ignored to evaluate the speech reconstruction results, which not only reduces the amount of feature data during the result evaluation, but also ensures that the evaluation focuses on the key features, and realizes the evaluation of the reconstruction results using only relatively important speech features, thereby improving the efficiency of the reconstruction evaluation. When applied to the intelligent customer service speech reconstruction scenario, the intelligent customer service reconstructed speech that meets the quality standards can be sent to the relevant target customers more quickly, thereby improving business efficiency.

[0100] Continue to refer Figure 3 In some optional implementations, a step of preprocessing the source speech is also included before step 202. Figure 3 This is a flowchart of a specific embodiment of preprocessing the source speech in the speech reconstruction result evaluation method described in the present application, comprising the following steps:

[0101] Step 301, obtaining a sound waveform change diagram of the source speech by an oscilloscope method;

[0102] Specifically, the sound waveform change diagram of the source speech is obtained through the oscilloscope method, so that the subsequent prosodic feature extraction network can sample and process the sound waveform change diagram to obtain the sound frequency value.

[0103] Step 302, using a sound level meter to measure a pressure change graph of the source speech;

[0104] Specifically, since the sound amplitude value generally represents the sound intensity, that is, the sound decibel, the pressure change graph of the source speech can be measured by a sound level meter, so that the subsequent prosodic feature extraction network can sample and process the pressure change graph to obtain the sound amplitude value.

[0105] Step 303: Identify all syllables contained in the source speech and the phoneme sequences contained in each syllable according to a preset syllable-phoneme comparison table.

[0106] Specifically, phonemes are the smallest units that constitute syllables, and speech is composed of several syllables. Through a preset syllable-phoneme comparison table, all syllables contained in the source speech and the phoneme sequences contained in each syllable are identified, so that the phoneme duration extracted by the rhythmic feature extraction network can be used to calculate the duration of all pronounced syllables in the source speech.

[0107] Continue to refer Figure 4 , Figure 4 yes Figure 2 The flowchart of a specific embodiment of step 202 shown includes the following steps:

[0108] Step 401, capturing the sound frequency values ​​and sound amplitude values ​​corresponding to all pronunciation periods of the source speech and capturing the pronunciation duration of all syllables in the source speech through the prosodic feature extraction network;

[0109] Step 402: The sound frequency values, sound amplitude values ​​and pronunciation durations of all syllables corresponding to the source speech in all pronunciation time periods are used as prosodic features of the source speech.

[0110] By capturing the prosodic features of the source speech, subsequent processing steps are facilitated, relevant information in the prosodic features is used to reconstruct the speech, and the speech reconstruction result is evaluated.

[0111] Continue to refer Figure 5 , Figure 5 yes Figure 4 The flowchart of a specific embodiment of step 401 shown includes the following steps:

[0112] Step 501, inputting the sound waveform change graph and the pressure change graph into the rhythm feature extraction network;

[0113] Step 502, sampling the sound waveform change according to the sampling time interval preset in the prosody feature extraction network, and obtaining the sound frequency values ​​corresponding to the source speech in all unit time periods;

[0114] Step 503, sampling the pressure change graph according to the sampling time interval, and obtaining the sound amplitude values ​​corresponding to the source speech in all unit time periods;

[0115] In this embodiment, the sampling time intervals in step 502 and step 503 remain consistent, and the sampling time interval can be set according to actual sampling requirements. Generally speaking, the smaller the sampling time interval is and the closer the sampling time interval is to the duration of each syllable, the more reliable the sampling result is.

[0116] Step 504, capturing the pronunciation durations corresponding to all the phonemes in the source speech respectively through the prosodic feature extraction network;

[0117] Step 505 , obtaining the pronunciation durations of all the syllables according to the phoneme sequences respectively included in each syllable and the pronunciation duration corresponding to each phoneme.

[0118] In this embodiment, after knowing the phoneme sequence contained in each syllable and the pronunciation duration corresponding to each phoneme, the pronunciation duration of all syllables can be obtained by accumulating the phoneme pronunciation duration.

[0119] By extracting the sound frequency values, sound amplitude values ​​and pronunciation durations of all syllables contained in the source speech corresponding to all unit time periods, the stress features in the source speech can be subsequently screened out according to the sound frequency values, sound amplitude values ​​and pronunciation durations of all syllables, so as to achieve the subsequent evaluation of speech reconstruction results using important features.

[0120] Continue to refer Figure 6 , Figure 6 yes Figure 2 The flowchart of a specific embodiment of step 205 shown includes the following steps:

[0121] Step 601, comparing the sound frequency values ​​corresponding to the source voice in all unit time periods with a preset sound frequency threshold one by one, and screening out all sound frequency values ​​exceeding the sound frequency threshold;

[0122] Step 602, comparing the sound amplitude values ​​corresponding to the source speech in all unit time periods with a preset sound amplitude threshold one by one, and screening out all sound amplitude values ​​exceeding the sound amplitude threshold;

[0123] Step 603, comparing the pronunciation durations of all the syllables in the source speech with the preset pronunciation durations one by one, and screening out all the syllables that exceed the preset pronunciation duration;

[0124] Step 604, taking all the sound frequency values ​​exceeding the sound frequency threshold, all the sound amplitude values ​​exceeding the sound amplitude threshold, and all the syllables exceeding the preset pronunciation duration as the stress feature extraction result of the source speech;

[0125] Step 605: Similarly, the stress features of the target speech are filtered out from the prosodic features of the target speech.

[0126] Specifically, the stress feature of the target speech is selected from the prosodic feature of the target speech, and combined with Figure 3 , Figure 4 and Figure 5 , replace the source speech with the target speech and perform the same processing steps, so that Figure 6 Steps 601 to 604 in the above step are used to obtain the stress feature extraction result of the target speech.

[0127] Since the larger the sound frequency value, the larger the sound amplitude value and the longer the pronunciation syllable of the corresponding syllable within a unit time period, the more important the corresponding speech pronunciation content is, in this embodiment, by comparison, all sound frequency values ​​exceeding the sound frequency threshold, all sound amplitude values ​​exceeding the sound amplitude threshold, and all syllables exceeding the preset pronunciation duration are screened out, and all sound frequency values ​​exceeding the sound frequency threshold, all sound amplitude values ​​exceeding the sound amplitude threshold and all syllables exceeding the preset pronunciation duration are used as the stress feature extraction results of the source speech, so as to facilitate the subsequent evaluation of the reconstruction effect of only the important part of the pronunciation content.

[0128] Continue to refer Figure 7 , Figure 7 yes Figure 2 The flowchart of a specific embodiment of step 206 shown includes the following steps:

[0129] Step 701, inputting the stress feature extraction result of the source speech and the stress feature extraction result of the target speech into the discriminator;

[0130] Step 702, performing multi-dimensional discrimination on the stress feature extraction result of the source speech and the stress feature extraction result of the target speech according to the multi-dimensional recognition component in the discriminator to obtain a multi-dimensional discrimination result;

[0131] Step 703: Perform comprehensive judgment based on the multi-dimensional judgment result to obtain the transfer effect of the accent feature from the source voice to the target voice.

[0132] By using the multi-dimensional recognition component in the discriminator to perform multi-dimensional discrimination on the stress feature extraction results of the source speech and the stress feature extraction results of the target speech, a multi-dimensional discrimination result is obtained, and then based on the multi-dimensional discrimination result, a comprehensive discrimination is performed to obtain the transmission effect of the stress feature from the source speech to the target speech, thereby avoiding the evaluation of the speech reconstruction effect from only a single dimension, which is more scientific and more convincing.

[0133] Continue to refer Figure 8 , Figure 8 yes Figure 7 The flowchart of a specific embodiment of step 702 shown includes the following steps:

[0134] Step 801, performing consistency comparison on all the sound frequency values ​​in the source speech exceeding the sound frequency threshold and all the sound frequency values ​​in the target speech exceeding the sound frequency threshold, and obtaining a consistency representation value of the target speech compared with the source speech as a first comparison result;

[0135] Step 802, performing consistency comparison on all the sound amplitude values ​​in the source speech exceeding the sound amplitude threshold and all the sound amplitude values ​​in the target speech exceeding the sound amplitude threshold, and obtaining a consistency representation value of the target speech compared with the source speech as a second comparison result;

[0136] Step 803, performing consistency comparison between all syllables in the source speech that exceed the preset pronunciation duration and all syllables in the target speech that exceed the preset pronunciation duration, and obtaining a consistency representation value of the target speech compared with the source speech as a third comparison result;

[0137] Step 804: Use the first comparison result, the second comparison result, and the third comparison result as the multi-dimensional discrimination result.

[0138] In this embodiment, multi-dimensional discrimination is performed on the stress feature extraction results of the source speech and the stress feature extraction results of the target speech. Specifically, the discrimination results are obtained through at least three dimensions of sound frequency value, sound amplitude value and syllable pronunciation duration. Then, based on the multi-dimensional discrimination results, a comprehensive discrimination is performed to obtain the transmission effect of the stress feature from the source speech to the target speech, thereby avoiding the evaluation of the speech reconstruction effect from only a single dimension, which is more scientific and convincing.

[0139] Continue to refer Fig. 9 , Fig. 9 yes Figure 7 The flowchart of a specific embodiment of step 703 shown includes the following steps:

[0140] Step 901, obtaining the comprehensive weight values ​​pre-set for the multi-dimensional discrimination results;

[0141] Step 902, multiplying the first comparison result, the second comparison result, and the third comparison result by corresponding comprehensive weight values ​​respectively to obtain a multiplication result;

[0142] Step 903, for the multiplication result, an average value is calculated by accumulating the result to obtain a comprehensive determination result;

[0143] Step 904, judging whether the comprehensive discrimination result exceeds a preset comprehensive discrimination threshold value by comparison;

[0144] Step 905: if the comprehensive discrimination result exceeds the preset comprehensive discrimination threshold, the conversion from the source speech to the target speech meets the speech reconstruction requirement;

[0145] Step 906, if the comprehensive judgment result does not exceed the preset comprehensive judgment threshold, the conversion from the source speech to the target speech does not meet the speech reconstruction requirements, and the network parameters of the speech signal reconstruction network are tuned until the conversion from the source speech to the target speech meets the speech reconstruction requirements.

[0146] In this embodiment, the discriminator based on the generative adversarial network can judge the transmission effect of the accent feature from the source speech to the target speech, and can not only evaluate and judge the speech reconstruction result, but also reversely optimize and adjust the speech signal reconstruction network in step 203 based on the evaluation and judgment result, thereby realizing that the accent feature can be used to evaluate the speech reconstruction effect and the speech signal reconstruction network can be adjusted and optimized.

[0147] The present application obtains source speech; captures the prosodic features of the source speech through a preset prosodic feature extraction network; inputs the prosodic features into a speech signal reconstruction network to obtain a reconstructed target speech and capture the prosodic features of the target speech; extracts stress features from the prosodic features of the source speech and the prosodic features of the target speech; inputs the stress feature extraction results into a discriminator based on a generative adversarial network to discriminate the transmission effect of the stress features from the source speech to the target speech. By extracting the stress features in the prosodic features of the source speech and the prosodic features of the target speech, it is convenient to evaluate the effect of converting the source speech to the target speech from the main information part or the pronunciation emphasis part of the speech. Only by combining the relevant stress features and ignoring the unstressed features that have little effect to evaluate the speech reconstruction results, the amount of feature data during the result evaluation is reduced, and the evaluation can be focused on the key features, so that only relatively important speech features are used to evaluate the reconstruction results, thereby improving the efficiency of the reconstruction evaluation. Applied to the intelligent customer service speech reconstruction scenario, the intelligent customer service reconstructed speech that meets the quality standards can be sent to relevant target customers more quickly, thereby improving business efficiency.

[0148] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Among them, artificial intelligence (AI) is the theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.

[0149] AI basic technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, mechatronics, etc. AI software technologies mainly include computer vision technology, robotics technology, biometrics technology, speech processing technology, natural language processing technology, and machine learning / deep learning.

[0150] In the embodiment of the present application, the source speech is obtained; the prosodic features of the source speech are captured through a preset prosodic feature extraction network; the prosodic features are input into a speech signal reconstruction network to obtain a reconstructed target speech, and the prosodic features of the target speech are captured; the prosodic features of the source speech and the prosodic features of the target speech are extracted with accent features; the accent feature extraction result is input into a discriminator based on a generative adversarial network to discriminate the transmission effect of the accent feature from the source speech to the target speech. By extracting the accent features in the prosodic features of the source speech and the prosodic features of the target speech, it is convenient to evaluate the effect of converting the source speech into the target speech from the main information part or the pronunciation emphasis part in the speech. Only the relevant accent features are combined, and the unaffected features are ignored to evaluate the speech reconstruction results, which not only reduces the amount of feature data during the result evaluation, but also ensures that the evaluation focuses more on the key features, and realizes the evaluation of the reconstruction results using only relatively important speech features, thereby improving the efficiency of the reconstruction evaluation. When applied to the scenario of intelligent customer service voice reconstruction, it can send intelligent customer service reconstructed voice that meets quality standards to relevant target customers more quickly, thereby improving business efficiency.

[0151] Further references Fig.10 , as a response to the above Figure 2 In order to realize the method shown in the figure, the present application provides an embodiment of a speech reconstruction result evaluation device, and the device embodiment is Figure 2 Corresponding to the method embodiment shown, the device can be specifically applied to various electronic devices.

[0152] like Fig.10 As shown, the speech reconstruction result evaluation device 10 described in this embodiment includes: a source speech acquisition module 10a, a first prosodic feature capture module 10b, a target speech reconstruction module 10c, a second prosodic feature capture module 10d, an accent feature screening and extraction module 10e, and an accent feature transfer effect determination module 10f. Among them:

[0153] A source speech acquisition module 10a, used to acquire source speech;

[0154] A first prosodic feature capturing module 10b, configured to capture the prosodic features of the source speech through a preset prosodic feature extraction network;

[0155] A target speech reconstruction module 10c is used to input the prosodic features into a speech signal reconstruction network to obtain a reconstructed target speech;

[0156] A second prosodic feature capturing module 10d, configured to capture the prosodic features of the target speech through the prosodic feature extraction network;

[0157] The stress feature screening and extraction module 10e is used to extract stress features from the prosodic features of the source speech and the prosodic features of the target speech according to a preset screening and extraction strategy, and obtain stress feature extraction results;

[0158] The accent feature transfer effect discrimination module 10f is used to input the accent feature extraction result into a discriminator based on a generative adversarial network to discriminate the transfer effect of the accent feature from the source speech to the target speech.

[0159] The present application obtains source speech; captures the prosodic features of the source speech through a preset prosodic feature extraction network; inputs the prosodic features into a speech signal reconstruction network to obtain a reconstructed target speech and capture the prosodic features of the target speech; extracts stress features from the prosodic features of the source speech and the prosodic features of the target speech; inputs the stress feature extraction results into a discriminator based on a generative adversarial network to discriminate the transmission effect of the stress features from the source speech to the target speech. By extracting the stress features in the prosodic features of the source speech and the prosodic features of the target speech, it is convenient to evaluate the effect of converting the source speech to the target speech from the main information part or the pronunciation emphasis part of the speech. Only by combining the relevant stress features and ignoring the unstressed features that have little effect to evaluate the speech reconstruction results, the amount of feature data during the result evaluation is reduced, and the evaluation can be focused on the key features, so that only relatively important speech features are used to evaluate the reconstruction results, thereby improving the efficiency of the reconstruction evaluation. Applied to the intelligent customer service speech reconstruction scenario, the intelligent customer service reconstructed speech that meets the quality standards can be sent to relevant target customers more quickly, thereby improving business efficiency.

[0160] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through computer-readable instructions, and the computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, the aforementioned storage medium can be a non-volatile storage medium such as a disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM).

[0161] It should be understood that, although the steps in the flowchart of the accompanying drawings are displayed in sequence as indicated by the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least a part of the steps in the flowchart of the accompanying drawings may include multiple sub-steps or multiple stages, and these sub-steps or stages are not necessarily executed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be executed in turn or alternately with other steps or at least a part of the sub-steps or stages of other steps.

[0162] To solve the above technical problems, the present application also provides a computer device. Fig.11 , Fig.11 This is a basic structural block diagram of the computer device in this embodiment.

[0163] The computer device 11 includes a memory 11a, a processor 11b, and a network interface 11c which are interconnected through a system bus. It should be noted that: Fig.11 Only a computer device 11 having a component memory 11a, a processor 11b, and a network interface 11c is shown, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead. Among them, those skilled in the art can understand that the computer device here is a device that can automatically perform numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes but is not limited to a microprocessor, an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), a digital processor (DSP), an embedded device, etc.

[0164] The computer device may be a computing device such as a desktop computer, a notebook, a PDA, a cloud server, etc. The computer device may interact with a user through a keyboard, a mouse, a remote controller, a touch pad, or a voice control device.

[0165] The memory 11a includes at least one type of readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, etc. In some embodiments, the memory 11a can be an internal storage unit of the computer device 11, such as a hard disk or memory of the computer device 11. In other embodiments, the memory 11a can also be an external storage device of the computer device 11, such as a plug-in hard disk equipped on the computer device 11, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. Of course, the memory 11a can also include both the internal storage unit of the computer device 11 and its external storage device. In this embodiment, the memory 11a is generally used to store the operating system and various application software installed on the computer device 11, such as a computer-readable instruction of a method for evaluating speech reconstruction results. In addition, the memory 11a can also be used to temporarily store various types of data that have been output or are to be output.

[0166] The processor 11b may be a central processing unit (CPU), a controller, a microcontroller, a microprocessor, or other data processing chips in some embodiments. The processor 11b is generally used to control the overall operation of the computer device 11. In this embodiment, the processor 11b is used to run the computer-readable instructions stored in the memory 11a or process data, such as running the computer-readable instructions of the speech reconstruction result evaluation method.

[0167] The network interface 11c may include a wireless network interface or a wired network interface. The network interface 11c is generally used to establish a communication connection between the computer device 11 and other electronic devices.

[0168] The computer device proposed in this embodiment belongs to the field of research and development design and audio processing technology, and is applied to the audio reconstruction result evaluation scenario. The present application obtains the source speech; captures the prosodic features of the source speech through a preset prosodic feature extraction network; inputs the prosodic features into the speech signal reconstruction network to obtain the reconstructed target speech, and captures the prosodic features of the target speech; extracts the stress features of the prosodic features of the source speech and the prosodic features of the target speech; inputs the stress feature extraction results into the discriminator based on the generative adversarial network to discriminate the transmission effect of the stress features from the source speech to the target speech. By extracting the stress features in the prosodic features of the source speech and the prosodic features of the target speech, it is convenient to evaluate the effect of converting the source speech into the target speech from the main information part or the pronunciation emphasis part of the speech. Only the relevant stress features are combined, and the light tone features with little influence are ignored to evaluate the speech reconstruction results, which not only reduces the amount of feature data during the result evaluation, but also ensures that the evaluation focuses more on the key features, and realizes the evaluation of the reconstruction results using only relatively important speech features, thereby improving the efficiency of reconstruction evaluation. When applied to the scenario of intelligent customer service voice reconstruction, it can send intelligent customer service reconstructed voice that meets quality standards to relevant target customers more quickly, thereby improving business efficiency.

[0169] The present application also provides another implementation, namely, providing a computer-readable storage medium, wherein the computer-readable storage medium stores computer-readable instructions, and the computer-readable instructions can be executed by a processor to enable the processor to perform the steps of the speech reconstruction result evaluation method as described above.

[0170] The computer-readable storage medium proposed in this embodiment belongs to the field of research and development design and audio processing technology, and is applied to the audio reconstruction result evaluation scenario. The present application obtains the source speech; captures the prosodic features of the source speech through a preset prosodic feature extraction network; inputs the prosodic features into the speech signal reconstruction network to obtain the reconstructed target speech, and captures the prosodic features of the target speech; extracts the stress features of the prosodic features of the source speech and the prosodic features of the target speech; inputs the stress feature extraction results into the discriminator based on the generative adversarial network to discriminate the transmission effect of the stress features from the source speech to the target speech. By extracting the stress features in the prosodic features of the source speech and the prosodic features of the target speech, it is convenient to evaluate the effect of converting the source speech into the target speech from the main information part or the pronunciation emphasis part of the speech. Only by combining the relevant stress features and ignoring the unstressed features that have little effect to evaluate the speech reconstruction results, the amount of feature data during the result evaluation is reduced, and it can also ensure that the evaluation focuses more on the key features, so that only relatively important speech features are used to evaluate the reconstruction results, and the efficiency of reconstruction evaluation is improved. When applied to the scenario of intelligent customer service voice reconstruction, it can send intelligent customer service reconstructed voice that meets quality standards to relevant target customers more quickly, thereby improving business efficiency.

[0171] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application.

[0172] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A method for evaluating speech reconstruction results, characterized in that: The steps include: Get the source voice; Capturing the prosodic features of the source speech through a preset prosodic feature extraction network; Inputting the prosodic features into a speech signal reconstruction network to obtain a reconstructed target speech; Capturing the prosodic features of the target speech through the prosodic feature extraction network; According to a preset screening and extraction strategy, respectively extracting stress features from the prosodic features of the source speech and the prosodic features of the target speech to obtain stress feature extraction results; The accent feature extraction result is input into a discriminator based on a generative adversarial network to discriminate the transfer effect of the accent feature from the source speech to the target speech.

2. The speech reconstruction result evaluation method according to claim 1, characterized in that: Before executing the step of capturing the prosodic features of the source speech through a preset prosodic feature extraction network, the method further includes: Obtaining a sound waveform change diagram of the source speech by an oscilloscope method; Using a sound level meter, measuring a pressure change graph of the source speech; According to a preset syllable-phoneme comparison table, identifying all syllables contained in the source speech and the phoneme sequence contained in each syllable; The step of capturing the prosodic features of the source speech through a preset prosodic feature extraction network specifically includes: The prosodic feature extraction network is used to capture the sound frequency values ​​and sound amplitude values ​​corresponding to all pronunciation periods of the source speech and the pronunciation duration of all syllables in the source speech; The sound frequency values, sound amplitude values ​​and pronunciation durations of all syllables corresponding to the source speech in all pronunciation time periods are used as the prosodic features of the source speech.

3. The speech reconstruction result evaluation method according to claim 2, characterized in that: The step of capturing the sound frequency values, sound amplitude values ​​and pronunciation durations of all syllables in the source speech corresponding to all pronunciation time periods of the source speech through the prosodic feature extraction network specifically includes: Inputting the sound waveform change graph and the pressure change graph into the rhythm feature extraction network; According to the sampling time interval preset in the prosodic feature extraction network, the sound waveform change is sampled and processed to obtain the sound frequency values ​​corresponding to the source speech in all unit time periods; According to the sampling time interval, the pressure change graph is sampled to obtain the sound amplitude values ​​corresponding to the source speech in all unit time periods; Capturing the pronunciation durations corresponding to all phonemes in the source speech respectively through the prosodic feature extraction network; The pronunciation durations of all the syllables are obtained according to the phoneme sequences respectively contained in each syllable and the pronunciation duration corresponding to each phoneme.

4. The speech reconstruction result evaluation method according to claim 2 or 3, characterized in that: The step of extracting stress features from the prosodic features of the source speech and the prosodic features of the target speech respectively according to a preset screening and extraction strategy to obtain stress feature extraction results specifically includes: Compare the sound frequency values ​​corresponding to the source voice in all unit time periods with the preset sound frequency thresholds one by one, and filter out all sound frequency values ​​exceeding the sound frequency threshold; Compare the sound amplitude values ​​corresponding to the source voice in all unit time periods with the preset sound amplitude threshold value one by one, and filter out all sound amplitude values ​​exceeding the sound amplitude threshold value; Compare the pronunciation durations of all syllables in the source speech with the preset pronunciation durations one by one, and filter out all syllables that exceed the preset pronunciation duration; Taking all the sound frequency values ​​exceeding the sound frequency threshold, all the sound amplitude values ​​exceeding the sound amplitude threshold, and all the syllables exceeding the preset pronunciation duration as the stress feature extraction results of the source speech; Similarly, the stress features of the target speech are screened out from the prosodic features of the target speech.

5. The method for evaluating speech reconstruction results according to claim 4, characterized in that: The step of inputting the stress feature extraction result into a discriminator based on a generative adversarial network to discriminate the transfer effect of the stress feature from the source voice to the target voice specifically includes: Inputting the stress feature extraction result of the source speech and the stress feature extraction result of the target speech into the discriminator; Performing multi-dimensional discrimination on the stress feature extraction result of the source speech and the stress feature extraction result of the target speech according to the multi-dimensional recognition component in the discriminator to obtain a multi-dimensional discrimination result; Based on the multi-dimensional discrimination result, a comprehensive discrimination is performed to obtain the transfer effect of the accent feature from the source voice to the target voice.

6. The method for evaluating speech reconstruction results according to claim 5, characterized in that: The step of performing multi-dimensional discrimination on the stress feature extraction result of the source speech and the stress feature extraction result of the target speech according to the multi-dimensional recognition component in the discriminator to obtain the multi-dimensional discrimination result specifically includes: Performing consistency comparison on all the sound frequency values ​​in the source speech exceeding the sound frequency threshold and all the sound frequency values ​​in the target speech exceeding the sound frequency threshold, and obtaining a consistency representation value of the target speech compared with the source speech as a first comparison result; Performing consistency comparison on all the sound amplitude values ​​in the source speech exceeding the sound amplitude threshold and all the sound amplitude values ​​in the target speech exceeding the sound amplitude threshold, and obtaining a consistency representation value of the target speech compared with the source speech as a second comparison result; Comparing all syllables in the source speech that exceed the preset pronunciation duration with all syllables in the target speech that exceed the preset pronunciation duration, and obtaining a consistency representation value of the target speech compared with the source speech as a third comparison result; The first comparison result, the second comparison result and the third comparison result are used as the multi-dimensional discrimination result.

7. The method for evaluating speech reconstruction results according to claim 6, characterized in that: The step of performing comprehensive discrimination based on the multi-dimensional discrimination result to obtain the transfer effect of the accent feature from the source voice to the target voice specifically includes: Obtaining comprehensive weight values ​​pre-set for the multi-dimensional discrimination results; performing multiplication operations on the first comparison result, the second comparison result, and the third comparison result with corresponding comprehensive weight values ​​respectively to obtain multiplication results; For the multiplication result, a comprehensive discrimination result is obtained by accumulating and averaging the result; By comparison, determining whether the comprehensive discrimination result exceeds a preset comprehensive discrimination threshold; If the comprehensive discrimination result exceeds the preset comprehensive discrimination threshold, the conversion from the source speech to the target speech meets the speech reconstruction requirement; If the comprehensive judgment result does not exceed the preset comprehensive judgment threshold, the conversion from the source speech to the target speech does not meet the speech reconstruction requirements, and the network parameters of the speech signal reconstruction network are tuned until the conversion from the source speech to the target speech meets the speech reconstruction requirements.

8. A speech reconstruction result evaluation device, characterized in that: include: A source voice acquisition module, used to acquire source voice; A first prosodic feature capturing module, used to capture the prosodic features of the source speech through a preset prosodic feature extraction network; A target speech reconstruction module, used for inputting the prosodic features into a speech signal reconstruction network to obtain a reconstructed target speech; A second prosodic feature capturing module, used for capturing the prosodic features of the target speech through the prosodic feature extraction network; An accent feature screening and extraction module is used to extract accent features from the prosodic features of the source speech and the prosodic features of the target speech according to a preset screening and extraction strategy, and obtain an accent feature extraction result; The stress feature transfer effect discrimination module is used to input the stress feature extraction result into a discriminator based on a generative adversarial network to discriminate the transfer effect of the stress feature from the source speech to the target speech.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores computer-readable instructions, and the processor implements the steps of the method for evaluating speech reconstruction results as claimed in any one of claims 1 to 7 when executing the computer-readable instructions.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer-readable instructions, and when the computer-readable instructions are executed by a processor, the steps of the method for evaluating speech reconstruction results according to any one of claims 1 to 7 are implemented.

Images