Non-intrusive narrowband speech quality assessment method and system for low-rate communications

Abstract

The invention belongs to the technical field of speech signal processing, and provides a non-intrusive narrowband speech quality assessment method and system for low-rate communication. The technical solution is to train a constructed speech quality assessment model based on speech data to obtain a trained speech quality assessment model; wherein the construction process of the speech quality assessment model comprises: extracting Mel spectrum features of speech data; obtaining an asymmetric feature map of speech data based on the Mel spectrum feature extraction; dividing the asymmetric feature map of speech data into multiple frequency bands, enhancing the features of each frequency band, and obtaining an enhanced feature map; calculating the score of each frame of speech based on the enhanced feature map, and calculating the score of the entire speech data based on the score of each frame of speech; and performing quality assessment on speech data to be assessed based on the trained speech quality assessment model, without the need for an original reference signal, and effectively assessing the speech quality in an actual communication environment.

Description

Technical Field

[0001] The present invention belongs to the technical field of speech signal processing, and in particular relates to a non-intrusive narrowband speech quality assessment method and system for low-rate communication. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Compared with medium and high bit rate voice communication, low bit rate voice communication faces greater challenges in ensuring voice quality. This is because low bit rate speech coding usually leads to higher signal distortion and information loss, which directly affects the intelligibility and naturalness of speech signals. Traditional speech coding quality assessment methods, such as Perceptual Evaluation of Speech Quality (PESQ), Short-Time Objective Intelligibility (STOI), Perceptual Objective Listening Quality Assessment (POLQA) and ITU-T P.563, can effectively evaluate the quality of speech signals, but they require the original signal as a reference to give an accurate evaluation. However, in actual application scenarios, the original signal is not always available, which severely limits the use of these algorithms.

[0004] At present, researchers have designed non-intrusive PESQ speech quality assessment methods based on deep learning technology, but these studies mainly focus on wideband speech (16kHz or higher frequency), while PESQ was originally designed for narrowband speech signals. There is a relative lack of research on narrowband signals in low bit rate communications, especially in real communication environments, where the impact of common channel bit errors and other factors on speech quality has not been fully considered. Summary of the invention

[0005] In order to solve at least one technical problem existing in the above-mentioned background technology, the present invention provides a non-intrusive narrowband speech quality assessment method and system for low-rate communication, which can effectively assess the speech quality in an actual communication environment without the need for an original reference signal, and can not only cope with interference factors in the communication environment such as channel bit errors, but also provide a reliable basis for quality assurance and optimization of narrowband speech.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A first aspect of the present invention provides a non-intrusive narrowband speech quality assessment method for low-rate communication, comprising the following steps:

[0008] Acquire speech data, perform source channel coding on the speech data, and generate a speech quality assessment data set with a PESQ score;

[0009] The constructed speech quality assessment model is trained based on the speech quality assessment data set to obtain a trained speech quality assessment model; wherein the construction process of the speech quality assessment model includes:

[0010] Extract the Mel spectrum features of speech data;

[0011] An asymmetric feature map of speech data is obtained based on Mel spectrum feature extraction;

[0012] The asymmetric feature map of the speech data is divided into multiple frequency bands, and the features of each frequency band are enhanced to obtain an enhanced feature map;

[0013] Calculate the score of each frame of speech based on the enhanced feature map, and calculate the score of the entire speech data based on the score of each frame of speech;

[0014] The quality of the speech data to be evaluated is evaluated based on the trained speech quality evaluation model to obtain an evaluation result.

[0015] Further, the performing source channel coding on the speech data to generate a speech quality assessment data set with a PESQ score includes: selecting a plurality of low-rate source encoder and channel encoder combinations to simulate an actual communication system;

[0016] Input the speech sample file into the simulated communication system, and simulate the bit error conditions that may occur in the actual communication process by setting different bit error rates to generate a speech sample data set with the characteristics of a low-rate communication system;

[0017] Re-segment the voice file into voice segments, and store the voice segments as independent voice files;

[0018] The original speech signal without source channel coding is used as the reference signal, and the quality of the segmented independent speech files is evaluated to obtain a speech quality evaluation dataset with a PESQ score.

[0019] Furthermore, an asymmetric feature map of the speech data is obtained based on the Mel spectrum feature extraction, including:

[0020] The Mel spectrum feature is input into an asymmetric feature extraction network, which includes a densely connected convolution module and a transposed convolution module. The densely connected convolution module adopts a feature extraction network with three densely connected blocks as the backbone. The Mel spectrum feature is input into the first feature extraction network to obtain a first speech feature; the first speech feature is input into the first feature extraction network to obtain a second speech feature, and the second speech feature is input into the third feature extraction network to obtain a third speech feature; the third speech feature is input into the transposed convolution module to obtain a final asymmetric feature map of the speech data.

[0021] Furthermore, in the densely connected convolution module, each feature extraction network includes three identical common convolution blocks and a cascade structure, each cascade structure includes a cascade layer and a convolution layer, the input features and output features of the common convolution block are input into the cascade layer for concatenation, the concatenated features are input into the convolution layer, and the speech features of the corresponding feature extraction network are output;

[0022] The transposed convolution part includes a plurality of transposed convolution layers and ordinary convolution layers, and the transposed convolution layers and the ordinary convolution layers are cross-arranged.

[0023] Furthermore, the convolution part is transposed, and an activation function layer is applied after each convolution layer to activate the output data.

[0024] Furthermore, the asymmetric feature map of the speech data is divided into a plurality of frequency bands, and the features of each frequency band are enhanced to obtain an enhanced feature map, including:

[0025] Split the asymmetric feature map into multiple frequency bands from the frequency dimension;

[0026] The feature map corresponding to each frequency band is used to calculate the features of each channel contained in each frequency band through global average pooling;

[0027] Pass the features of each channel through a fully connected layer and use the Sigmoid activation function to obtain the attention weight matrix;

[0028] Adjust the attention weight matrix to the same matrix shape as the corresponding frequency band, and then multiply it with the attention weight matrix to obtain the weight-enhanced frequency band feature;

[0029] All weight-enhanced frequency band features are concatenated along the frequency dimension into the final feature map.

[0030] Furthermore, after acquiring the voice data, the voice data is preprocessed, including framing, windowing and pre-emphasis processing of the voice data.

[0031] Further, the score of each frame of speech is calculated based on the enhanced feature map, and the score of the entire speech data is calculated based on the score of each frame of speech, including:

[0032] The feature map after frequency band enhancement is rearranged into the format of (T, C, F), where T represents the time dimension, C represents the number of channels, and F represents the frequency dimension;

[0033] Flatten the channel and frequency dimensions to generate a two-dimensional feature map for each frame of speech;

[0034] The two-dimensional feature map of each frame of speech is input into the BiLSTM network to capture the bidirectional dependencies in the series. After the output features of the BiLSTM network are processed through multiple fully connected layers, the scores of each frame are averaged using global average pooling to obtain the score of the entire speech data.

[0035] A second aspect of the present invention provides a non-intrusive narrowband speech quality assessment system for low-rate communications, comprising:

[0036] A voice data acquisition module is used to acquire voice data, perform source channel coding on the voice data, and generate a voice quality assessment data set with a PESQ score;

[0037] The speech quality assessment model training module is used to train the constructed speech quality assessment model based on speech data to obtain a trained speech quality assessment model; wherein the construction process of the speech quality assessment model includes:

[0038] Extract the Mel spectrum features of speech data;

[0039] An asymmetric feature map of speech data is obtained based on Mel spectrum feature extraction;

[0040] The asymmetric feature map of the speech data is divided into multiple frequency bands, and the features of each frequency band are enhanced to obtain an enhanced feature map;

[0041] Calculate the score of each frame of speech based on the enhanced feature map, and calculate the score of the entire speech data based on the score of each frame of speech;

[0042] The speech quality assessment module is used to perform quality assessment on the speech data to be assessed based on the trained speech quality assessment model to obtain an assessment result.

[0043] Further, in the voice data acquisition module, the performing source channel coding on the voice data to generate a voice quality assessment data set with a PESQ score includes: selecting a plurality of low-rate source encoder and channel encoder combinations to simulate an actual communication system;

[0044] Input the speech sample file into the simulated communication system, and simulate the bit error conditions that may occur in the actual communication process by setting different bit error rates to generate a speech sample data set with the characteristics of a low-rate communication system;

[0045] Re-segment the voice file into voice segments, and store the voice segments as independent voice files;

[0046] The original speech signal without source channel coding is used as the reference signal, and the quality of the segmented independent speech files is evaluated to obtain a speech quality evaluation dataset with a PESQ score.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] 1. The present invention obtains an asymmetric feature map of speech data based on Mel spectrum feature extraction; divides the asymmetric feature map of speech data into multiple frequency bands, enhances the features of each frequency band, and obtains an enhanced feature map; calculates the score of each frame of speech based on the enhanced feature map, and calculates the score of the entire speech data based on the score of each frame of speech; can accurately evaluate low-bit stream speech signals in complex communication environments such as channel errors without relying on the original signal. At the same time, the model improves the expression ability of speech signal features in neural networks by effectively extracting and enhancing spectral features, and significantly enhances its adaptability and accuracy in actual communication scenarios.

[0049] 2. The present invention innovatively designs an asymmetric feature extraction module and proposes a feature extraction method combining densely connected convolutional networks and transposed convolution. The DCC module extracts abstract features of speech signals through multiple densely connected convolutional blocks with residuals, while the transposed convolution uses forward and back propagation mechanisms to adaptively learn features that are more important for speech quality assessment tasks.

[0050] 3. The present invention designs a frequency band attention mechanism, which divides the feature map into multiple frequency bands according to the frequency dimension, and weights each channel of each frequency band separately to enhance the importance representation ability of different frequency regions. This mechanism can effectively improve the utilization efficiency of frequency domain features, thereby improving the accuracy of overall speech quality assessment.

[0051] 4. The global average pooling scoring mechanism of the present invention calculates the overall system speech quality score based on frame-level features, ensuring the stability and accuracy of the evaluation results, and is particularly suitable for narrowband speech signal quality evaluation in low-rate communication scenarios.

[0052] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0054] Figure 1 It is a flow chart of a non-intrusive narrowband speech quality assessment method for low-rate communication provided by an embodiment of the present invention;

[0055] Figure 2 It is a diagram of the overall model architecture provided by an embodiment of the present invention;

[0056] Figure 3 It is an asymmetric feature extraction network provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0058] It should be noted that the following detailed descriptions are all illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0059] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0060] As mentioned in the background technology, there is a relative lack of research on narrowband signals in low bit rate communications, especially in real communication environments, where the impact of common channel bit errors and other factors on voice quality has not been fully considered.

[0061] The present invention provides a non-intrusive narrowband speech quality assessment method for low-rate communication, aiming to solve the quality assessment problem of narrowband speech signals in low-rate communication. The present invention realizes the effective extraction of multi-level features of speech signals by constructing a speech quality data set and designing an asymmetric feature extraction module. In order to enhance the degree of attention to important features, the present invention designs a frequency band attention mechanism, which can weight features in different frequency ranges, thereby highlighting features that have an important impact on speech quality. Subsequently, the features are input into a regression network, firstly BiLSTM is used to further capture the temporal context information, and each frame feature of the time dimension is input into a fully connected layer to obtain frame-level features, and finally the overall speech quality score is calculated through a global average pooling layer. The non-intrusive assessment method of the present invention does not need to refer to the original speech signal, can effectively deal with interference factors such as communication channel errors, and provides an accurate and adaptable solution for narrowband speech quality assessment in low-rate communication.

[0062] Embodiment 1

[0063] like Figure 1 and 2 As shown, this embodiment provides a non-intrusive narrowband voice quality assessment method for low-rate communication, comprising the following steps:

[0064] Step 1: Get voice data;

[0065] In this embodiment, it is necessary to first obtain a voice sample file, using 8k sampling and 16-bit quantization. By obtaining sufficient voice sample files, it is ensured that the samples have sufficient diversity and representativeness in voice content and quality, thereby providing reliable data support for subsequent model training.

[0066] Step 2: Construct a speech quality assessment dataset for low-rate communication scenarios based on the acquired speech data;

[0067] The specific steps include:

[0068] Step 201: Select a plurality of low-rate source encoders and channel encoders, and combine the source encoder and the channel encoder to simulate an actual communication system;

[0069] In this embodiment, a multi-rate Melpe encoder is used as a source coding scheme (2.4 kbit / s, 1.2 kbit / s and 0.6 kbit / s), and a hard decision convolutional code based on a Viterbi decoding algorithm is used as a channel coding and decoding scheme (1 / 2 efficiency, 1 / 4 efficiency). Among them, the 2.4 kbit / s Melpe encoder is not combined with channel coding, the 1.2 kbit / s Melpe is combined with a 1 / 2 efficiency convolutional code, and the 0.6 kbit / s Melpe is combined with a 1 / 4 efficiency convolutional code, which are called three coding modes.

[0070] Step 202: input the prepared speech sample file into the simulated communication system, and simulate the bit error conditions that may occur in the actual communication process by setting different bit error rates, so as to generate a speech sample data set with low-rate communication system characteristics;

[0071] The bit error rate is set based on the error correction capability of the channel coding. The bit error rate is a random bit error of the bit stream obtained after the source channel coding. The bit stream after the random bit error is input into the channel decoder for decoding and error correction. After that, the decoded speech signal obtained by the source decoding is a speech sample data set with the characteristics of the low-rate communication system.

[0072] Step 203, re-segmenting the voice file into voice segments to adapt to the voice sample length that can be evaluated by PESQ, and storing them as independent voice files;

[0073] Step 204: Using the original speech signal that has not been source-channel encoded as a reference signal, the PESQ tool is used to objectively evaluate the quality of the segmented independent speech files to obtain a speech quality evaluation data set for low-rate communication scenarios.

[0074] Step 3: Preprocess the speech data in the speech quality assessment dataset and extract Mel spectrum features;

[0075] In this embodiment, the voice file is framed, windowed, and pre-emphasized, and then the short-time Fourier transform is used to perform spectrum analysis on the voice signal of each frame, extract spectrum features, and map them to the Mel scale to obtain a Mel spectrum feature map.

[0076] Step 4: Input the Mel spectrum features into the asymmetric feature extraction network to extract the asymmetric feature map;

[0077] like Figure 3 As shown, the specific steps include:

[0078] Step 401: input the obtained Mel spectrum feature into the densely connected convolution part of the asymmetric feature extraction network. The densely connected convolution adopts a feature extraction network with three densely connected blocks (DCC) as the backbone. The Mel spectrum feature is input into the first feature extraction network to obtain the first speech feature; the first speech feature is input into the first feature extraction network to obtain the second speech feature, and the second speech feature is input into the third feature extraction network to obtain the third speech feature; the third speech feature is input into the transposed convolution module to obtain the final asymmetric feature map of the speech data;

[0079] Each DCC block includes three identical common convolution blocks and a cascade structure. Each cascade structure includes a cascade layer and a convolution layer. The input features and output features of the common convolution block are input into the cascade layer for concatenation. The concatenated features are input into the convolution layer, and the speech features of the corresponding feature extraction network are output.

[0080] In this embodiment, the step size of the convolution layer in the cascade structure is set to 2 (non-time dimension). In this way, the computational complexity of the model is reduced, while the receptive field is expanded, so that the deep features of the model can be extracted more effectively.

[0081] In this embodiment, an activation function layer is applied after each convolution layer to activate the output data.

[0082] Step 402: input the first speech feature into the transposed convolution part of the asymmetric feature extraction network, the transposed convolution part includes a plurality of transposed convolution layers and ordinary convolution layers, the transposed convolution layers and ordinary convolution layers are cross-arranged, and an activation function layer is applied after each convolution layer to activate the output data, and finally a second speech feature map is output;

[0083] More abstract speech features are extracted through densely connected convolution blocks, while reducing the feature loss in the time dimension. Then, transposed convolution is added for upsampling. Based on the forward and back propagation mechanism of the neural network, transposed convolution can not only restore the abstract low-dimensional features to high-dimensional features and capture more detailed information, but also adaptively learn features that are more suitable for the current task.

[0084] Step 5: Divide the asymmetric feature map into multiple frequency bands, enhance the features of each frequency band, and obtain an enhanced feature map;

[0085] The specific steps include:

[0086] Step 501: Split the asymmetric feature map into N In this embodiment, N = 8. The feature dimension is (frequency f ,time t ,aisle c ). No. i The characteristic diagram of the frequency band X i The definition of is shown in formula (1):

[0087] (1),

[0088] in, Indicates the size of each frequency band, i ∈{1,2,3,...,N}.

[0089] Step 502: Global average pooling: For the features of each frequency band , the features of each channel contained in each frequency band are calculated by global average pooling , the calculation formula is:

[0090] (2),

[0091] in, It is a feature after global pooling;

[0092] Step 503: Band attention weight matrix: Through a fully connected layer, and use the Sigmoid activation function to get the attention weight , as shown in formula (3):

[0093] (3),

[0094] in, ;

[0095] Step 504: broadcast the Adjust to the corresponding frequency band Same shape, get the adjusted attention weights , as shown in formula (4):

[0096] (4),

[0097] Step 505: Weight-enhanced frequency band features: Add attention weights Apply to the corresponding frequency band The enhanced frequency band characteristics are obtained , as shown in formula (5):

[0098] (5),

[0099] Step 506: Frequency band feature concatenation: Finally, all enhanced frequency band features are concatenated. Concatenate along the frequency dimension into the final feature map :

[0100] (6),

[0101] Through these steps, the model can better utilize local and global information in the frequency dimension, thereby improving the feature extraction effect of speech signals.

[0102] Step 6: Obtain a score for each frame of speech based on the enhanced feature map, and calculate a score for the entire speech data based on the score for each frame of speech;

[0103] The specific steps include:

[0104] Step 601: First, the feature map after frequency band enhancement is rearranged into the format of (T, C, F), where T represents the time dimension (number of frames), C represents the number of channels, and F represents the frequency dimension. In this way, the shape of the feature map becomes a format suitable for input into subsequent network processing:

[0105] (7),

[0106] Step 602: Flatten the channel (C) and frequency (F) dimensions to generate a two-dimensional feature map of each frame of speech, where the horizontal axis represents time (T) and the vertical axis represents the feature of each frame ( )). That is, the flattened feature map is:

[0107] (8),

[0108] Step 603: Input the two-dimensional feature map of each frame of speech into a bidirectional long short-term memory neural network (BiLSTM) to capture the bidirectional dependencies in the sequence and extract richer temporal context information. Assume that the input feature map is , then the output of the BiLSTM network for:

[0109] (9),

[0110] in, ∈R 1×2H , represents the features of the bidirectional LSTM output, and H is the number of hidden units of each LSTM unit.

[0111] Step 604: Apply multiple fully connected neural networks to each frame feature. Assume that each frame feature is processed by multiple FC layers to obtain the output , where the number of units in the last layer is 1, which is used to obtain the frame-level score.

[0112] (10),

[0113] in, is the output of BiLSTM at the tth frame, and FC represents the fully connected layer.

[0114] Step 605: Finally, a global average pooling (GAP) layer is used to calculate the overall system score based on these frame-level scores.

[0115] Global average pooling averages the scores of each frame to obtain a system-level score ,Right now:

[0116] (11),

[0117] in, Indicates the system rating of the speech, that is, the overall rating.

[0118] The speech samples that have been source channel coded and have quality labels are used as data sets to train the above speech quality assessment model. Furthermore, the data set is divided into training set, validation set and test set. We use the training set to optimize the model, select the mean square error (MSE) as the loss function, and update the parameters through the Adam optimizer. During the optimization process, in order to prevent overfitting, an early stopping mechanism is adopted. When the loss of the validation set does not show a significant decrease in 5 consecutive rounds, the training is automatically terminated. After the training is completed, multiple evaluation indicators (such as MSE, SRCC, LCC, etc.) are used to evaluate the performance of the model on the validation set and test set, so as to verify the robustness and accuracy of the model in a low-rate communication environment.

[0119] Embodiment 2

[0120] This embodiment provides a non-intrusive narrowband voice quality assessment system for low-rate communication, including:

[0121] A voice data acquisition module, used for acquiring voice data;

[0122] The speech quality assessment model training module is used to train the constructed speech quality assessment model based on speech data to obtain a trained speech quality assessment model; wherein the construction process of the speech quality assessment model includes:

[0123] Extract the Mel spectrum features of speech data;

[0124] An asymmetric feature map of speech data is obtained based on Mel spectrum feature extraction;

[0125] The asymmetric feature map of the speech data is divided into multiple frequency bands, and the features of each frequency band are enhanced to obtain an enhanced feature map;

[0126] A score for each frame of speech is obtained based on the enhanced feature map, and a score for the entire speech data is calculated based on the score for each frame of speech;

[0127] The speech quality assessment module is used to perform quality assessment on the speech data to be assessed based on the trained speech quality assessment model to obtain an assessment result.

[0128] Embodiment 3

[0129] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the non-intrusive narrowband speech quality assessment method for low-rate communication as described above are implemented.

[0130] Embodiment 4

[0131] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the non-intrusive narrowband speech quality assessment method for low-rate communication as described above are implemented.

[0132] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A non-intrusive narrowband speech quality assessment method for low-rate communications, characterized in that: The steps include: Acquire speech data, perform source channel coding on the speech data, and generate a speech quality assessment data set with a PESQ score; The constructed speech quality assessment model is trained based on the speech quality assessment data set to obtain a trained speech quality assessment model; wherein the construction process of the speech quality assessment model includes: Extract the Mel spectrum features of speech data; An asymmetric feature map of speech data is obtained based on Mel spectrum feature extraction; The asymmetric feature map of speech data is obtained based on Mel spectrum feature extraction, including: The Mel spectrum feature is input into an asymmetric feature extraction network, the asymmetric feature extraction network includes a densely connected convolution module and a transposed convolution module, the densely connected convolution module adopts a feature extraction network with three densely connected blocks as the backbone, the Mel spectrum feature is input into the first feature extraction network to obtain a first speech feature; the first speech feature is input into the second feature extraction network to obtain a second speech feature, the second speech feature is input into the third feature extraction network to obtain a third speech feature; the third speech feature is input into the transposed convolution module to obtain a final asymmetric feature map of the speech data; The asymmetric feature map of the speech data is divided into multiple frequency bands, and the features of each frequency band are enhanced to obtain an enhanced feature map; The asymmetric feature graph of the speech data is divided into a plurality of frequency bands, and the features of each frequency band are enhanced to obtain an enhanced feature graph, including: Split the asymmetric feature map into multiple frequency bands from the frequency dimension; The feature map corresponding to each frequency band is used to calculate the features of each channel contained in each frequency band through global average pooling; Pass the features of each channel through a fully connected layer and use the Sigmoid activation function to obtain the attention weight matrix; Adjust the attention weight matrix to the same matrix shape as the corresponding frequency band, and then multiply it with the attention weight matrix to obtain the weight-enhanced frequency band feature; Concatenate all weight-enhanced frequency band features along the frequency dimension into the final feature map; A score for each frame of speech is obtained based on the enhanced feature map, and a score for the entire speech data is calculated based on the score for each frame of speech; The quality of the speech data to be evaluated is evaluated based on the trained speech quality evaluation model to obtain an evaluation result.

2. The non-intrusive narrowband speech quality assessment method for low-rate communication according to claim 1, characterized in that: The method of performing source channel coding on the speech data to generate a speech quality assessment data set with a PESQ score includes: selecting a plurality of low-rate source encoder and channel encoder combinations to simulate an actual communication system; Input the speech sample file into the simulated communication system, and simulate the bit error conditions that may occur in the actual communication process by setting different bit error rates to generate a speech sample data set with the characteristics of a low-rate communication system; Re-segment the voice file into voice segments, and store the voice segments as independent voice files; The original speech signal without source channel coding is used as the reference signal, and the quality of the segmented independent speech files is evaluated to obtain a speech quality evaluation dataset with a PESQ score.

3. The non-intrusive narrowband speech quality assessment method for low-rate communication according to claim 1, characterized in that: In the densely connected convolution module, each feature extraction network includes three identical common convolution blocks and a cascade structure. Each cascade structure includes a cascade layer and a convolution layer. The input features and output features of the common convolution block are input into the cascade layer for concatenation. The concatenated features are input into the convolution layer, and the speech features of the corresponding feature extraction network are output. The transposed convolution part includes a plurality of transposed convolution layers and ordinary convolution layers, and the transposed convolution layers and the ordinary convolution layers are cross-arranged.

4. The non-intrusive narrowband speech quality assessment method for low-rate communication as claimed in claim 3, characterized in that: In the transposed convolution part, an activation function layer is applied after each convolution layer to activate the output data.

5. The non-intrusive narrowband speech quality assessment method for low-rate communication according to claim 1, characterized in that: After acquiring the voice data, the voice data is preprocessed, including framing, windowing and pre-emphasis processing of the voice data.

6. The non-intrusive narrowband speech quality assessment method for low-rate communication according to claim 1, characterized in that: The score of each frame of speech is obtained based on the enhanced feature map, and the score of the entire speech data is calculated based on the score of each frame of speech, including: The feature map after frequency band enhancement is rearranged into the format of (T, C, F), where T represents the time dimension, C represents the number of channels, and F represents the frequency dimension; Flatten the channel and frequency dimensions to generate a two-dimensional feature map for each frame of speech; The two-dimensional feature map of each frame of speech is input into the BiLSTM network to capture the bidirectional dependencies in the series. After the output features of the BiLSTM network are processed through multiple fully connected layers, the scores of each frame are averaged using global average pooling to obtain the score of the entire speech data.

7. A non-intrusive narrowband speech quality assessment system for low-rate communications, characterized in that: include: A voice data acquisition module is used to acquire voice data, perform source channel coding on the voice data, and generate a voice quality assessment data set with a PESQ score; The speech quality assessment model training module is used to train the constructed speech quality assessment model based on speech data to obtain a trained speech quality assessment model; wherein the construction process of the speech quality assessment model includes: Extract the Mel spectrum features of speech data; An asymmetric feature map of speech data is obtained based on Mel spectrum feature extraction; The asymmetric feature map of speech data is obtained based on Mel spectrum feature extraction, including: The Mel spectrum feature is input into an asymmetric feature extraction network, the asymmetric feature extraction network includes a densely connected convolution module and a transposed convolution module, the densely connected convolution module adopts a feature extraction network with three densely connected blocks as the backbone, the Mel spectrum feature is input into the first feature extraction network to obtain a first speech feature; the first speech feature is input into the second feature extraction network to obtain a second speech feature, the second speech feature is input into the third feature extraction network to obtain a third speech feature; the third speech feature is input into the transposed convolution module to obtain a final asymmetric feature map of the speech data; The asymmetric feature map of the speech data is divided into multiple frequency bands, and the features of each frequency band are enhanced to obtain an enhanced feature map; The asymmetric feature graph of the speech data is divided into a plurality of frequency bands, and the features of each frequency band are enhanced to obtain an enhanced feature graph, including: Split the asymmetric feature map into multiple frequency bands from the frequency dimension; The feature map corresponding to each frequency band is used to calculate the features of each channel contained in each frequency band through global average pooling; Pass the features of each channel through a fully connected layer and use the Sigmoid activation function to obtain the attention weight matrix; Adjust the attention weight matrix to the same matrix shape as the corresponding frequency band, and then multiply it with the attention weight matrix to obtain the weight-enhanced frequency band feature; Concatenate all weight-enhanced frequency band features along the frequency dimension into the final feature map; Calculate the score of each frame of speech based on the enhanced feature map, and calculate the score of the entire speech data based on the score of each frame of speech; The speech quality assessment module is used to perform quality assessment on the speech data to be assessed based on the trained speech quality assessment model to obtain an assessment result.

8. The non-intrusive narrowband speech quality assessment system for low-rate communication as claimed in claim 7, characterized in that: In the voice data acquisition module, the voice data is source-channel encoded to generate a voice quality assessment data set with a PESQ score, including: selecting a combination of multiple low-rate source encoders and channel encoders to simulate an actual communication system; Input the speech sample file into the simulated communication system, and simulate the bit error conditions that may occur in the actual communication process by setting different bit error rates to generate a speech sample data set with the characteristics of a low-rate communication system; Re-segment the voice file into voice segments, and store the voice segments as independent voice files; The original speech signal without source channel coding is used as the reference signal, and the quality of the segmented independent speech files is evaluated to obtain a speech quality evaluation dataset with a PESQ score.

Images