A remote speaker confirmation method and system based on semantic communication

Abstract

The present invention discloses a remote speaker confirmation method and system based on semantic communication, which specifically relates to the technical field of voice signal transmission and processing, and includes two parts: local transmission end encoding and remote receiving end decoding. The local transmission end encoding includes extracting semantic features from the spectrum of the original voice signal to obtain speaker semantic features, then adjusting the speaker semantic feature dimensions, and finally transmitting to the remote receiving end through a channel; the remote receiving end decoding includes: receiving the speaker semantic feature signal after dimension adjustment transmitted by the local transmission end; performing dimension recovery and normalization processing on the speaker semantic feature signal after dimension adjustment to obtain speaker embedding features; comparing the speaker embedding features with the target speaker embedding features to determine whether the speaker is the target speaker. The present invention can improve the reliability and accuracy of the speaker confirmation task at a low signal-to-noise ratio while saving communication resources.

Description

Technical Field

[0001] The invention relates to a remote speaker confirmation method and system based on semantic communication, belonging to the technical field of speech signal processing and transmission. Background Art

[0002] In recent years, with the development of artificial intelligence, the Internet of Things, and intelligent human-computer interaction, more and more devices have been deployed, and the amount of data required to be transmitted has increased exponentially. Efficient and reliable communications are urgently needed to cope with the increasingly tight communication resources.

[0003] In traditional speaker verification systems, the transmitter needs to send the entire speech to the receiver, which transmits a large amount of redundant information irrelevant to the task and wastes scarce communication resources. On the other hand, the voice signal is easily affected by the channel during transmission, especially in low signal-to-noise ratio conditions, making it almost impossible to complete the speaker verification task. At the same time, in traditional methods, once the signal is monitored, the transmitted voice channel can be easily restored by the monitor, resulting in user information leakage, posing a security risk.

[0004] Semantic communication is a technology that extracts, compresses and transmits features from original signals and uses semantic information for communication. In semantic communication technology, the sender uses a neural network to extract features related to the task and sends them to the receiver through a channel. The receiver completes the task based on the received semantic information. Due to the large number of parameters in the neural network, the feature values ​​after being interfered by the channel can be corrected by the neural network to a certain extent, so that the distance between the feature and the target feature is still smaller than the distance between the feature and other target features, ensuring that the task can still maintain a high accuracy rate at a low signal-to-noise ratio. At the same time, since in semantic communication, only the semantic information related to the task needs to be transmitted through the channel, the amount of data transmitted can be greatly reduced compared to traditional communication schemes. In addition, semantic communication only transmits features related to the target task to avoid direct theft of information during transmission, which can protect user privacy to a great extent. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies in the prior art and to provide a remote speaker verification method and system based on semantic communication, which can improve the reliability and accuracy of speaker verification tasks at low signal-to-noise ratios while saving communication resources.

[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0007] In a first aspect, the present invention provides a remote speaker confirmation method based on semantic communication, which is applied to a local sending end and specifically includes:

[0008] Obtaining the original speech signal;

[0009] The original speech signal is converted into Mel frequency cepstral coefficients to obtain the spectrum of the original speech signal;

[0010] Extracting semantic features based on the frequency spectrum of the original speech signal to obtain semantic features of the speaker;

[0011] According to the speaker semantic features, adjusting the dimensions of the speaker semantic features to obtain the speaker semantic features after the dimensions are adjusted;

[0012] The dimensionally adjusted speaker semantic features are transmitted to the remote receiving end through the channel, so that the remote receiving end performs the following steps:

[0013] receiving a speaker semantic feature signal after dimension adjustment transmitted by a local sending end;

[0014] Performing dimension restoration and normalization processing on the speaker semantic feature signal after dimension adjustment to obtain the speaker embedding feature;

[0015] Compare the speaker embedding features with the target speaker embedding features to determine whether the speaker is the target speaker.

[0016] Further, semantic features are extracted according to the spectrum of the original speech signal, and the semantic features of the speaker are obtained by a semantic encoder;

[0017] The semantic encoder includes an initialization layer, a plurality of squeeze excitation-residual modules with different expansion coefficients, a fully connected layer, and an attention statistics pooling layer module connected in sequence;

[0018] The initialization layer is used to expand the number of channels of the spectrum of the original speech signal, and includes a one-dimensional convolution layer and an activation function layer connected in sequence;

[0019] Each squeeze excitation-residual module is sequentially arranged to extract speaker embedding features of different scales and increase the receptive field, and includes a one-dimensional convolution layer, a one-dimensional residual convolution layer, a one-dimensional convolution layer and a squeeze excitation module connected sequentially;

[0020] The attention statistics pooling layer module is used to dynamically adjust the weights of different channels in the speaker embedding features to exploit global properties.

[0021] Furthermore, according to the speaker semantic features, the dimensions of the speaker semantic features are adjusted, and the speaker semantic features after the dimension adjustment are realized by a channel encoder. The channel encoder is used to change the dimensional shape of the speaker semantic features to make them suitable for channel transmission, which includes a reconstruction layer.

[0022] Furthermore, the dimension of the speaker semantic features after dimension adjustment is restored and normalized, and the speaker embedding features are obtained and implemented through a channel decoder and a semantic decoder;

[0023] The channel decoder is used to restore the dimension of semantic features, and includes a reconstruction layer; the semantic decoder is used to normalize the speaker's semantic features after the dimension is restored, and includes a batch normalization layer.

[0024] Furthermore, the speaker embedding feature is compared with the target speaker embedding feature to determine whether the speaker is the target speaker, including:

[0025] Calculate the cosine distance between the speaker embedding feature and the target speaker embedding feature, which is expressed as follows:

[0026]

[0027] in, Embedding features for the target speaker, is the speaker embedding feature, and n is the dimension size of the feature.

[0028] The cosine distance between the speaker embedding feature and the target speaker embedding feature is compared with a preset threshold. If the cosine distance between the speaker embedding feature and the target speaker embedding feature is greater than the preset threshold, the speaker is judged to be the target speaker. If the cosine distance between the speaker embedding feature and the target speaker embedding feature is less than the preset threshold, the speaker is judged not to be the target speaker.

[0029] In a second aspect, the present invention further provides a remote speaker confirmation method based on semantic communication, which is applied to a remote receiving end and specifically includes:

[0030] receiving a speaker semantic feature signal after dimension adjustment transmitted by a local sending end;

[0031] Performing dimension restoration and normalization processing on the speaker semantic feature signal after dimension adjustment to obtain the speaker embedding feature;

[0032] Compare the speaker embedding features with the target speaker embedding features to determine whether the speaker is the target speaker;

[0033] The dimension-adjusted speaker semantic feature signal is obtained by the local sending end performing the following steps:

[0034] Obtaining the original speech signal;

[0035] The original speech signal is converted into Mel frequency cepstral coefficients to obtain the spectrum of the original speech signal;

[0036] Extracting semantic features based on the frequency spectrum of the original speech signal to obtain semantic features of the speaker;

[0037] According to the speaker semantic features, the speaker semantic features are dimensionally adjusted to obtain the speaker semantic features after dimension adjustment.

[0038] In a third aspect, the present invention provides a remote speaker verification system based on semantic communication, which includes a local sending end and a remote receiving end, with a channel provided between the two;

[0039] The local transmitting end includes a speech signal acquisition module, a spectrum conversion module, a semantic encoder, and a channel encoder;

[0040] The remote receiving end includes a channel decoder, a semantic decoder and an analysis and judgment module;

[0041] The voice signal acquisition module is used to acquire the original voice signal;

[0042] The spectrum conversion module is used to convert the original speech signal into Mel-frequency cepstral coefficients to obtain the spectrum of the original speech signal;

[0043] The semantic encoder is used to extract semantic features according to the frequency spectrum of the original speech signal to obtain the semantic features of the speaker;

[0044] The channel encoder is used to adjust the dimension of the speaker semantic feature according to the speaker semantic feature to obtain the speaker semantic feature after the dimension adjustment;

[0045] The channel is used to transmit the dimension-adjusted speaker semantic features from the local sending end to the remote receiving end through the channel;

[0046] The channel decoder and the semantic decoder are used to perform dimension recovery and normalization processing on the speaker semantic feature signal after dimension adjustment, respectively, to obtain the speaker embedding feature;

[0047] The analysis and discrimination module is used to compare the speaker embedding features with the target speaker embedding features to determine whether the speaker is the target speaker.

[0048] Furthermore, the channel includes a Gaussian channel, a Rice channel or a Rayleigh channel.

[0049] Furthermore, the semantic encoder and the semantic decoder are pre-trained, and the pre-training method includes:

[0050] Obtain a training data set and a validation data set of the frequency spectrum of the speaker's speech signal;

[0051] The semantic encoder and the semantic decoder are placed in the remote speaker verification system, and the training data set is used as input and the speaker embedding features are used as output to train the remote speaker verification system. During the training process, AAM-softmax is used as the loss function, stochastic gradient descent is adopted, and the parameters are adjusted using the Adam optimizer to obtain the trained remote speaker verification system.

[0052] After each round of training is completed, the remote speaker verification system is tested with the verification data set as input and the speaker embedding feature as output, and the cosine distance between the speaker embedding feature and the current speaker embedding feature is calculated. The cosine distance between the speaker embedding feature and the current speaker embedding feature is compared with the preset threshold in turn to obtain the judgment result, and the system test error rate is obtained by combining all the judgment results. The hyperparameters are adjusted according to the system test error rate, and the next round of training is continued until the preset number of training times is reached to obtain the pre-trained semantic encoder and semantic decoder.

[0053] Furthermore, the signal-to-noise ratio of the channel is 0-20 dB.

[0054] Beneficial Effects

[0055] The present invention completes the extraction of speaker embedding features required for speaker confirmation at the local sending end, and then transmits the speaker embedding features, which can reduce the amount of transmitted data, reduce the time required for transmission, save communication resources, and due to the generalization of neural networks, can reduce the impact of channels on features to a certain extent, and improve the accuracy of speaker confirmation tasks at low signal-to-noise ratios; at the same time, due to the principle that semantic communication only transmits the target task features, direct theft is avoided during the transmission process, which can effectively protect user privacy. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flow chart of the remote speaker confirmation method based on semantic communication of the present invention. DETAILED DESCRIPTION

[0057] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0058] Example 1

[0059] like Figure 1 As shown, the remote speaker confirmation method based on semantic communication provided by this embodiment includes two parts: local sending end encoding and remote receiving end decoding.

[0060] Local sender encoding includes the following steps:

[0061] The original speech signal is obtained, and then the original speech signal is converted into Mel-frequency cepstral coefficients (MFCC) to obtain the spectrum of the original speech signal.

[0062] According to the frequency spectrum of the original speech signal, semantic features are extracted to obtain the semantic features of the speaker, which is specifically achieved through a semantic encoder.

[0063] The semantic encoder consists of a sequentially connected initialization layer, multiple squeeze-excitation-residual (SE-Res2Blocks) modules with different expansion coefficients, a fully connected layer, and an attention statistical pooling layer (ASP) module.

[0064] In this embodiment, the initialization layer includes, in sequence, a one-dimensional convolution layer with a stride of 1, a zero padding of 2 bits, and a convolution kernel size of 5, a ReLU activation function layer, and a normalization layer with a channel number of 1024.

[0065] In this embodiment, the number of squeeze-excitation-residual modules is 3, and their expansion coefficients are 2, 3, and 4 respectively. Each squeeze-excitation-residual module includes a one-dimensional convolution layer with a convolution kernel size of 1, a ReLU activation layer, a normalization layer, a Res2 one-dimensional residual convolution layer with an expansion coefficient of 2 / 3 / 4, a convolution kernel size of 3, and a scaling factor of 8, a ReLU activation layer, a normalization layer, a one-dimensional convolution layer with a convolution kernel size of 1, a ReLU activation layer, a normalization layer, and a squeeze-excitation (SE) module.

[0066] The attention statistics pooling layer module includes a linear activation layer, a Tanh activation function layer, a linear activation layer, a Softmax activation function layer, and a fully connected layer connected in sequence.

[0067] The processing of the semantic encoder includes:

[0068] The 80 channels of the original speech signal spectrum are expanded to 1024 through the initialization layer, and then the features of the 1024 channels are sequentially input into the squeeze-excitation-residual modules with expansion coefficients of 2, 3, and 4 to extract features of different scales. The output structures of the three squeeze-excitation-residual modules are cascaded and multi-layer feature aggregation (MFA) is performed through the fully connected layer to achieve residual connection. Finally, the attention statistical pooling layer module is used to perform attention statistical pooling to effectively utilize global attributes, thereby completing the semantic feature extraction task and obtaining the speaker's semantic features.

[0069] The dimension of the speaker's semantic features is adjusted to make them suitable for channel transmission, which is specifically achieved through a channel encoder. The channel encoder includes a reconstruction (Reshape) layer.

[0070] The processing expressions of the semantic encoder and channel encoder are as follows:

[0071]

[0072] in, is the spectrum of the original speech signal, The parameters are The semantic encoder is the channel encoder, It is the speaker semantic feature after dimension adjustment.

[0073] Finally, the dimensionally adjusted speaker semantic feature signal is transmitted to the remote receiving end through the channel.

[0074] The remote receiving end decoding includes the following steps:

[0075] The speaker semantic feature signal after dimension adjustment transmitted by the local sending end is received, and the received speaker semantic feature signal can be expressed as:

[0076]

[0077] in, is the received speaker semantic feature signal, is the channel parameter, is the speaker semantic feature after dimension adjustment, represents Gaussian noise, where represents the variance of Gaussian noise in each channel, Represents the identity matrix.

[0078] The dimension of the speaker semantic features after dimension adjustment is restored through a channel decoder, which includes a reconstruction layer. The restored speaker semantic features are received by the semantic decoder for normalization to obtain speaker embedding features. The semantic decoder includes a normalization layer, the purpose of which is to normalize the feature vector after passing through the channel. Its generalization can reduce the influence of the channel on the transmitted feature vector to a certain extent, and improve the accuracy of the speaker verification task.

[0079] The processing expressions of the semantic decoder and channel decoder are as follows:

[0080]

[0081] in, Embedding features for speakers, represents the channel decoder, Indicates that the parameter is The semantic decoder of is the received speaker semantic feature signal.

[0082] Compare the speaker embedding features with the target speaker embedding features to determine whether the speaker is the target speaker. Specifically:

[0083] Calculate the cosine distance between the speaker embedding feature and the target speaker embedding feature, which is expressed as follows:

[0084]

[0085] in, Embedding features for the target speaker, is the speaker embedding feature, and n is the dimension size of the feature.

[0086] The cosine distance between the speaker embedding feature and the target speaker embedding feature is compared with the preset threshold. If the cosine distance between the speaker embedding feature and the target speaker embedding feature is greater than the preset threshold, the speaker is judged to be the target speaker. If the cosine distance between the speaker embedding feature and the target speaker embedding feature is less than the preset threshold, the speaker is judged not to be the target speaker, thereby completing the speaker confirmation task.

[0087] The following is a description of a remote speaker confirmation method based on semantic communication in conjunction with a specific embodiment:

[0088] The bit rate of the spectrum of the original speech signal used in the test is 256 kb / s. The traditional method is to transmit the speech signal first and then recognize it.

[0089] The comparison between the data volume of the method of the present invention and the data volume of the traditional method is shown in Table 1:

[0090] Table 1: Comparison of the data volume of the method of the present invention and the data volume of the traditional method

[0091]

[0092] From the data in Table 1, it can be seen that in the traditional scheme, the data volume of the 25ms speech segment is 6400 bits; while the semantic features obtained by the method of the present invention for the 25ms speech segment are 192 bits, and compared with 16-bit binary quantization, the amount of data transmitted is 384 bits. The present invention reduces the amount of data by 94%, proving that the present invention can reduce the amount of data transmitted, reduce the time required for transmission, and save communication resources.

[0093] In the traditional scheme, 16-bit pulse code modulation (PCM) is used to encode the test spectrum, the modulation mode is binary phase shift keying (BPSK), the channel encoding and decoding mode is long code low density parity check code (LDPC), and the channel type is the same as the present invention.

[0094] Under different channel signal-to-noise ratios, the comparison results of the speaker confirmation error rate obtained by the method of the present invention and the speaker confirmation error rate obtained by the traditional solution are shown in Table 2:

[0095] Table 2: Comparison of speaker confirmation error rates obtained by the method of the present invention and those obtained by traditional solutions

[0096]

[0097] As can be seen from Table 2, under low signal-to-noise ratio conditions, under the three channel conditions, the accuracy of the speaker confirmation task of the present invention is much higher than the accuracy of the task of first transmitting and then identifying under the traditional scheme. In particular, under 0dB channel conditions, the Gaussian channel, Rayleigh channel and Rice channel are improved by 95.2%, 95.9% and 96.6% respectively compared with the traditional scheme, with significant effects. In addition, under high signal-to-noise ratio, except for the Gaussian channel, the results of this scheme are better than the traditional scheme. Therefore, the method of the present invention can effectively improve the reliability and accuracy of the speaker confirmation task under low signal-to-noise ratio.

[0098] Example 2

[0099] On the basis of Example 1, this embodiment provides a remote speaker verification system based on semantic communication, which includes a local sending end and a remote receiving end, and a channel is provided between the two.

[0100] The local sending end includes a speech signal acquisition module, a spectrum conversion module, a semantic encoder, and a channel encoder. The remote receiving end includes a channel decoder, a semantic decoder, and an analysis and discrimination module.

[0101] The voice signal acquisition module is used to acquire the original voice signal.

[0102] The spectrum conversion module is used to convert the original speech signal into Mel-frequency cepstral coefficients to obtain the spectrum of the original speech signal.

[0103] The semantic encoder is used to extract semantic features according to the frequency spectrum of the original speech signal to obtain the semantic features of the speaker.

[0104] The channel encoder is used to adjust the dimension of the speaker semantic feature according to the speaker semantic feature to obtain the speaker semantic feature after the dimension is adjusted.

[0105] The channel is used to transmit the dimensionally adjusted speaker semantic features from the local transmitter to the remote receiver through the channel. The channel is one of multiple types such as Gaussian channel, Rice channel and Rayleigh channel to adapt to multiple channel transmission modes. The signal-to-noise ratio of the channel is 0~20dB.

[0106] The channel decoder and the semantic decoder are respectively used to restore the dimension and normalize the speaker semantic feature signal after dimension adjustment to obtain the speaker embedding feature.

[0107] The analysis and discrimination module is used to compare the speaker embedding features with the target speaker embedding features to determine whether the speaker is the target speaker.

[0108] Among them, the semantic encoder and semantic decoder are implemented through deep neural networks, which need to be pre-trained and trained in an end-to-end manner with joint channels. AAM-softmax is used as the loss function, stochastic gradient descent is adopted, and the Adam optimizer is used to adjust the parameters.

[0109] Each type of channel needs to be trained separately. During the training process, the signal-to-noise ratio of the channel is randomly selected between 0 and 20 dB to improve the generalization of the model.

[0110] The main task of the semantic encoder is to extract semantic information from the input raw speech. During training, it focuses on extracting the speaker embedding features required for the target task, removing redundant information irrelevant to the task, and significantly reducing the amount of data transmitted.

[0111] The pre-training methods include:

[0112] Get the training dataset and validation dataset of the frequency spectrum of the speaker's speech signal.

[0113] The semantic encoder and semantic decoder are placed in the remote speaker verification system, and the training dataset is used as input and the speaker embedding features are used as output to train the remote speaker verification system. During the training process, AAM-softmax is used as the loss function, stochastic gradient descent is adopted, and the Adam optimizer is used to adjust the parameters to obtain the trained remote speaker verification system.

[0114] After each round of training is completed, the system is tested with the verification data set as input and the speaker embedding feature as output, and the cosine distance between the speaker embedding feature and the current speaker embedding feature is calculated. The cosine distance between the speaker embedding feature and the current speaker embedding feature is compared with the preset threshold in turn to obtain the judgment result, and the system test error rate is obtained by combining all the judgment results. The hyperparameters are adjusted according to the system test error rate, and the next round of training is continued, that is, the training process is repeated until the preset number of training times is reached, and the pre-trained semantic encoder and semantic decoder are obtained.

[0115] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A remote speaker verification method based on semantic communication, characterized in that: Applied to the local sender, including: Obtaining the original speech signal; The original speech signal is converted into Mel frequency cepstral coefficients to obtain the spectrum of the original speech signal; Extracting semantic features based on the frequency spectrum of the original speech signal to obtain semantic features of the speaker; According to the speaker semantic features, adjusting the dimensions of the speaker semantic features to obtain the speaker semantic features after the dimensions are adjusted; The dimensionally adjusted speaker semantic feature signal is transmitted to a remote receiving end through a channel, so that the remote receiving end performs the following steps: receiving a speaker semantic feature signal after dimension adjustment transmitted by a local sending end; Performing dimension restoration and normalization processing on the speaker semantic feature signal after dimension adjustment to obtain the speaker embedding feature; Compare the speaker embedding features with the target speaker embedding features to determine whether the speaker is the target speaker; According to the spectrum of the original speech signal, semantic features are extracted to obtain the semantic features of the speaker through a semantic encoder, wherein the semantic encoder includes an initialization layer, a plurality of squeeze-excitation-residual modules with different expansion coefficients, a fully connected layer, and an attention statistical pooling layer module connected in sequence; The initialization layer is used to expand the number of channels of the spectrum of the original speech signal, and includes a one-dimensional convolution layer and an activation function layer connected in sequence; Each squeeze excitation-residual module is sequentially arranged to extract speaker embedding features of different scales and increase the receptive field, and includes a one-dimensional convolution layer, a one-dimensional residual convolution layer, a one-dimensional convolution layer and a squeeze excitation module connected sequentially; The attention statistics pooling layer module is used to dynamically adjust the weights of different channels in the speaker embedding features to exploit global properties.

2. The remote speaker verification method based on semantic communication according to claim 1, characterized in that: According to the speaker semantic features, the speaker semantic features are dimensionally adjusted, and the dimensionally adjusted speaker semantic features are implemented through a channel encoder. The channel encoder is used to change the dimensional shape of the speaker semantic features to make them suitable for channel transmission, and includes a reconstruction layer.

3. The remote speaker verification method based on semantic communication according to claim 1, characterized in that: Perform dimension restoration and normalization on the speaker semantic features after dimension adjustment, and obtain the speaker embedding features through the channel decoder and semantic decoder; The channel decoder is used to restore the dimension of semantic features, and includes a reconstruction layer; the semantic decoder is used to normalize the speaker's semantic features after the dimension is restored, and includes a batch normalization layer.

4. The remote speaker verification method based on semantic communication according to claim 1, characterized in that: Compare the speaker embedding features with the target speaker embedding features to determine whether the speaker is the target speaker, including: Calculate the cosine distance between the speaker embedding feature and the target speaker embedding feature, which is expressed as follows: ； in, Embedding features for the target speaker, is the speaker embedding feature, n is the feature dimension size; The cosine distance between the speaker embedding feature and the target speaker embedding feature is compared with a preset threshold. If the cosine distance between the speaker embedding feature and the target speaker embedding feature is greater than the preset threshold, the speaker is judged to be the target speaker. If the cosine distance between the speaker embedding feature and the target speaker embedding feature is less than the preset threshold, the speaker is judged not to be the target speaker.

5. A remote speaker verification method based on semantic communication, characterized in that: Applied to the remote receiving end, including: receiving a speaker semantic feature signal after dimension adjustment transmitted by a local sending end; Performing dimension restoration and normalization processing on the speaker semantic feature signal after dimension adjustment to obtain the speaker embedding feature; Compare the speaker embedding features with the target speaker embedding features to determine whether the speaker is the target speaker; The dimension-adjusted speaker semantic feature signal is obtained by the local sending end performing the following steps: Obtaining the original speech signal; The original speech signal is converted into Mel frequency cepstral coefficients to obtain the spectrum of the original speech signal; Extracting semantic features based on the frequency spectrum of the original speech signal to obtain semantic features of the speaker; According to the speaker semantic features, the speaker semantic features are dimensionally adjusted to obtain the speaker semantic features after dimension adjustment.

6. A remote speaker verification system based on semantic communication, characterized in that: It includes a local sending end and a remote receiving end, with a channel between them; The local transmitting end includes a speech signal acquisition module, a spectrum conversion module, a semantic encoder, and a channel encoder; The remote receiving end includes a channel decoder, a semantic decoder and an analysis and judgment module; The voice signal acquisition module is used to acquire the original voice signal; The spectrum conversion module is used to convert the original speech signal into Mel-frequency cepstral coefficients to obtain the spectrum of the original speech signal; The semantic encoder is used to extract semantic features according to the frequency spectrum of the original speech signal to obtain the semantic features of the speaker; The channel encoder is used to adjust the dimension of the speaker semantic feature according to the speaker semantic feature to obtain the speaker semantic feature after the dimension adjustment; The channel is used to transmit the dimension-adjusted speaker semantic features from the local sending end to the remote receiving end through the channel; The channel decoder and the semantic decoder are used to perform dimension recovery and normalization processing on the speaker semantic feature signal after dimension adjustment, respectively, to obtain the speaker embedding feature; The analysis and discrimination module is used to compare the speaker embedding features with the target speaker embedding features to determine whether the speaker is the target speaker.

7. The remote speaker verification system based on semantic communication according to claim 6, characterized in that: The channel includes a Gaussian channel, a Rice channel or a Rayleigh channel.

8. The remote speaker verification system based on semantic communication according to claim 6, characterized in that: The semantic encoder and the semantic decoder are pre-trained, and the pre-training method includes: Obtain a training data set and a validation data set of the frequency spectrum of the speaker's speech signal; The remote speaker verification system is trained using the training data set as input and the speaker embedding feature as output. During the training process, AAM-softmax is used as the loss function, stochastic gradient descent is adopted, and the parameters are adjusted using the Adam optimizer to obtain a trained remote speaker verification system; After each round of training is completed, the remote speaker verification system is tested with the verification dataset as input and the speaker embedding feature as output, and the cosine distance between the speaker embedding feature and the target speaker embedding feature is calculated. The cosine distance between the speaker embedding feature and the target speaker embedding feature is compared with the preset threshold in turn to obtain the judgment result. The system test error rate is obtained by combining all the judgment results. The hyperparameters are adjusted according to the system test error rate, and the next round of training is continued until the preset number of training times is reached to obtain the pre-trained semantic encoder and semantic decoder.

9. The remote speaker verification system based on semantic communication according to claim 6, characterized in that: The signal-to-noise ratio of the channel is 0-20dB.

Images