Beidou short message voice communication method and system based on semantic coding of initial and final consonants

Abstract

This invention discloses a BeiDou short message voice communication method and system based on initial and final semantic coding. The method includes: preprocessing the input speech signal and extracting frame-level initial and final features; constructing syllable-level semantic coding triples; compressing and coding the syllable-level semantic units and encapsulating them into structured data conforming to the BeiDou short message format; decoding the semantic coding data and reconstructing the speech signal based on the pinyin sequence and prosody control parameters to achieve BeiDou short message voice communication. This method models at the language structure level, significantly improving speech compression efficiency and reducing communication load, and is suitable for ultra-low bit rate communication links such as BeiDou short message. It achieves the restoration of speech content, rhythm, and emotional features. Compared with existing speech compression schemes that only reconstruct speech outlines, this invention can reconstruct more natural and intelligible speech content.

Description

Technical Field

[0001] This invention belongs to the field of satellite communication and speech coding technology, and particularly relates to a BeiDou short message speech communication method and system based on initial and final semantic coding. Background Technology

[0002] The BeiDou short message communication system is widely used in emergency rescue, ocean navigation, and field exploration, boasting advantages such as wide coverage and strong anti-interference capabilities. This system achieves short message transmission without relying on ground base stations via the BeiDou RDSS link. However, limited by communication bandwidth and protocol design, the maximum data transmission per message is typically no more than 14,000 bits, making it difficult to meet the continuity and clarity requirements of voice services. Traditional voice coding methods, such as waveform coding and parametric coding, while possessing some compression capabilities, struggle to balance compression efficiency and semantic restoration in the BeiDou short message environment. Speech semantic coding technology, through deep networks to extract semantic and acoustic features, can achieve efficient speech reconstruction at low bit rates, showing potential for application in the BeiDou system. Especially in Chinese speech, the structure of initials and finals is naturally decomposable, suitable for semantic modeling and coding compression. However, currently, there is a lack of solutions that can efficiently integrate speech semantic coding with the BeiDou short message system. Ensuring voice quality, transmission efficiency, and communication robustness at extremely low bit rates is a key technical challenge that urgently needs to be overcome. Summary of the Invention

[0003] Purpose of the Invention: The purpose of this invention is to provide a BeiDou short message voice communication method and system based on initial consonant-final vowel semantic coding. It aims to solve the problem of achieving efficient, clear, and low-latency voice information transmission under the condition of limited bandwidth in BeiDou short message communication. This invention achieves ultra-low bit encoding and high-fidelity synthesis of voice content by constructing an initial consonant-final vowel-prosody ternary semantic model and a lightweight deep neural network. This method significantly improves the voice compression rate and reconstruction quality, and is suitable for the voice communication needs of BeiDou short message communication in emergency and field scenarios, ensuring communication robustness and low-latency performance.

[0004] Technical solution: The present invention provides a BeiDou short message voice communication method based on initial and final semantic coding, comprising the following steps:

[0005] S1: Preprocess the input speech signal and extract frame-level initial consonant and final vowel features;

[0006] S2: Based on the extracted frame-level initial consonant and final vowel features, construct syllable-level semantic coding triples;

[0007] S3: Based on syllable-level semantic coding triples, syllable-level semantic units are compressed and encoded, and encapsulated into structured data conforming to the BeiDou short message format to obtain semantic coding data;

[0008] S4: Decodes the semantically encoded data and reconstructs the speech signal based on the pinyin sequence and prosody control parameters to realize BeiDou short message voice communication.

[0009] Furthermore, in step S1, the preprocessing of the input speech signal specifically involves performing a first-order high-pass filter on the original speech sampling signal x[n] to obtain the pre-emphasized speech signal y[n], the expression of which is as follows:

[0010] y[n] = x[n] - αx[n-1]

[0011] Wherein, α is the pre-emphasis coefficient, with a value ranging from 0.95 to 0.97, used to enhance the short-term high-frequency energy variation characteristics of the speech signal;

[0012] The continuous speech signal is divided into several short time frames, and the nth sampling point of the t-th frame is defined as:

[0013] y t (n)=y[t·H+n],n=0,1,...,N-1

[0014] Where t is the frame index, H is the frame shift, and N is the frame length;

[0015] Windowing is applied to reduce discontinuities between frames. The windowed frame signal is represented as follows:

[0016]

[0017] Where w(n) is the Hamming window function, its expression is as follows:

[0018]

[0019] The resulting structured short-time frame sequences are used for frame-level initial consonant and final vowel feature extraction.

[0020] Furthermore, in step S1, the extraction of frame-level initial consonant features and final vowel features specifically includes the following steps:

[0021] S1.1, Frame-level initial consonant feature extraction: First, the preprocessed speech frame sequence... The 13-dimensional Mel-frequency cepstral coefficients (MFCCs) and their first and second-order difference features are extracted to form a 39-dimensional frame-level vector. The current frame is then concatenated with the context of the two frames before and after it, resulting in a 195-dimensional input vector. This 195-dimensional input vector is then fed into a lightweight DNN model for consonant classification. The DNN structure contains two hidden layers, each with 64 nodes, using ReLU activation. The output layer employs the softmax function, outputting a 22-dimensional probability vector. The probabilities correspond to 21 common Chinese initials and cases without initials, respectively; the total number of DNN model parameters is controlled within 100,000 to adapt to the computing power constraints of Beidou portable terminals.

[0022] S1.2 Frame-level Final Feature Extraction: First, the same 195-dimensional MFCC enhanced features as those used for initial consonant extraction are used as input. Since finals are usually distributed in the main vowel region of a syllable, a main vowel region recognition window is introduced to avoid relying on explicit syllable boundaries. Multiple consecutive frames are selected within the sliding window as a set of candidate final frames, and a steady-state final feature representation for this segment is constructed using a weighted average method, thereby improving the robustness of the modeling. Next, this average feature sequence is input into a single-layer Long Short-Term Memory (LSTM) network model to capture inter-frame semantic evolution features. The number of hidden units in the LSTM is set to 64. Subsequently, a softmax output layer is connected, corresponding to 35 commonly used Chinese finals and the case without finals in linguistics. The output is a 36-dimensional normalized probability distribution vector. The LSTM model introduces a gated sparse connection structure and supports a dynamic pruning mechanism.

[0023] Furthermore, step S2 specifically includes the following steps:

[0024] S2.1 Sliding window smoothing: Constructing a joint probability vector of initials and finals for the t-th frame. The smoothing vector is calculated using a moving average with a window length of ω. for:

[0025]

[0026] The syllable boundary determination criteria are defined as follows: when the KL divergence shows a significant change or the Top-1 predicted class changes abruptly, the current position t is considered the boundary of the k-th syllable, a new syllable segment is initiated, and the termination frame of the previous syllable segment is recorded as t. The starting frame of the new syllable segment is

[0027] The judgment criteria are expressed as follows:

[0028]

[0029] Where δ is the empirical threshold. This represents the probability vectors for smoothing consonants in frame t. Smooth probability vector of vowels Select the index corresponding to the maximum value from the list, and combine them to obtain the initial and final index pairs. These are the index values ​​of the initials and finals, respectively;

[0030] The initial consonant index is represented as follows:

[0031] 0→null, 1→b, 2→p, 3→m, …, 21→s

[0032] Where null represents the case without an initial consonant;

[0033] The vowel index is represented as follows:

[0034] 0→null, 1→a, 2→o, 3→e, 4→i, …35→üan

[0035] Where null indicates the case without a vowel;

[0036] S2.2 Syllable Pinyin Unit Generation: In each syllable segment, the smooth probability vectors of the initials and finals are aggregated at the frame level. That is, the initial and final indices of the k-th syllable segment are obtained by the following formulas:

[0037]

[0038] in, Let t be the consonant smoothing probability vector of the t-th frame. The probability of the corresponding initial consonant category index i, i∈{0,1,...,21}. Let t be the vowel smoothing probability vector. The probability of the corresponding vowel category index j, j∈{0,1,...,35};

[0039] S2.3, Prosodic Feature Extraction: In the k-th syllable segment Within the frame feature sequence, three prosodic features are extracted: fundamental frequency, sound energy, and duration.

[0040] Fundamental frequency extraction: For frame segments within syllable boundaries, the YIN algorithm is used to estimate the fundamental frequency value for each frame. It also filters out unvoiced frames and averages the fundamental frequencies of all valid frames to obtain the fundamental frequency characteristics of the syllables.

[0041]

[0042] Where 1[·] is the indicator function, which is only included in the average for frames with a fundamental frequency, and N f It is the number of valid frames;

[0043] Energy calculation: Take the 0th dimension component E(t) from the MFCC features of each frame in the k-th syllable segment, and then take the average value to obtain the average energy e of that syllable segment. (k) :

[0044]

[0045] Pitch length calculation: based on the start and end points of the syllable boundary frame. and Combining the frame shift H, the duration of the k-th syllable segment is calculated as follows:

[0046]

[0047] Finally, the prosodic features of each syllable are represented as a three-dimensional vector.

[0048] S2.4 Construct the smallest semantic unit at the syllable level. The semantic content of the k-th syllable segment can be combined into a triplet form:

[0049]

[0050] Each U (k) As the smallest transmission unit of the BeiDou short message coding system, it provides dual information support of phonemes and prosody for subsequent speech reconstruction.

[0051] Furthermore, step S3 specifically involves the following: the encoding unit of each speech syllable consists of three parts: initial consonant category ID, final vowel category ID, and prosodic feature parameters; the initial consonant category ID represents the initial consonant category of the syllable and corresponds to the index of the predefined initial consonant set; the final vowel category ID represents the final vowel category of the syllable and corresponds to the index of the predefined final vowel set; the prosodic feature parameters include three elements: fundamental frequency, energy, and syllable duration.

[0052] Therefore, the encoding unit for each syllable is represented as:

[0053]

[0054] in, These are the initial consonant category ID and the final vowel category ID, respectively, indexed by the initial consonant and final vowel category of the k-th syllable. and The binary encoding structure, E (k) D (k) They represent their corresponding fundamental frequencies. Energy e (k) With duration d (k) The quantized value;

[0055] All syllable coding units will be combined sequentially and encapsulated into the payload of a BeiDou short message. To support protocol identification, version compatibility, and data parsing, the following application layer protocol data frame structure is defined:

[0056]

[0057] The application layer header contains information such as protocol identifier (1 byte), version number (1 byte), and timestamp (3 bytes). The number of syllables M is a 2-byte integer representing the number of syllables encapsulated in this frame. The syllable encoding sequence consists of M smallest encoding units arranged in order. The CRC checksum (2 bytes) is used for integrity verification at the receiving end.

[0058] Furthermore, the initial consonant category ID and final vowel category ID are encoded using 5 bits and 6 bits respectively to cover the standard pinyin category set; to adapt to the maximum data payload limit of BeiDou short messages, the three prosodic feature parameters of each syllable are encoded using discretization and fixed-point quantization methods respectively; among which, To normalize the logarithmic compression of the syllable average fundamental frequency value, 6-bit non-uniform quantization encoding is used; E (k) The mean energy of syllable segments is calculated, and after taking the logarithm, non-uniform coding is performed based on the corpus distribution, represented by 6 bits; D (k) Using 3-bit uniform quantization encoding as the unit of frame count; the encoding bit width is adjusted according to the communication link and application requirements; when and When both values ​​are 0, the system determines that the syllable is a silent syllable. E (k) The fixed codes are 000000 (6 bits) and 000000 (6 bits) respectively.

[0059] Furthermore, step S4 specifically includes the following steps:

[0060] S4.1 The receiving end first parses the format of the BeiDou short message data frame, and according to the defined encoding structure, parses each 32-bit fixed-length minimum unit into... Triples; among which This represents the initial consonant index corresponding to the decoded syllable encoding unit of the i-th syllable. This represents the vowel index corresponding to the decoded syllable encoding unit of the i-th syllable. This represents the corresponding prosodic feature encoding vector, which includes the quantized values ​​of the fundamental frequency, energy, and duration parameters;

[0061] S4.2. Based on the predefined initial consonant index table and final vowel index table, decode each initial consonant index obtained from the initial consonant index table. Index of finals Each initial and final character is mapped to a standard Chinese Pinyin character; the initials and finals are concatenated according to pronunciation rules to form a syllable-level Pinyin text sequence, which drives the subsequent speech synthesis process; to improve the legality of Pinyin combinations and the accuracy of speech synthesis, a Pinyin dictionary comparison mechanism and syllable legality verification rules are introduced to verify and correct syllable combinations as necessary, ensuring that the generated Pinyin sequence conforms to language norms; simultaneously, for each syllable, a prosodic vector is generated. The quantized integer values ​​are restored to continuous physical quantities according to a predefined quantization inverse mapping table or linear mapping function; by accumulating the duration of each syllable, a syllable-level prosodic tag sequence is constructed on the time axis to provide temporal information guidance for the speech synthesis model;

[0062] S4.3 Construct a speech synthesis model based on dual input control of pinyin and prosody, using an improved FastSpeech2 network as the core generation module; the speech synthesis model uses the pinyin sequence as the main input and the prosody feature parameters as the auxiliary control input, and after passing through the embedding encoding and temporal modeling modules, it outputs the corresponding mel spectrogram.

[0063] Furthermore, in step S4.3, the construction of the speech synthesis model based on dual input control of pinyin and prosody specifically includes the following steps:

[0064] S4.3.1 Embedding layer: Maps Pinyin characters to high-dimensional vectors, and embeds prosodic feature parameters into control vectors. Prosodic feature parameters include fundamental frequency, energy, and duration.

[0065] S4.3.2 Transformer encoder-decoder structure: fusing semantic and prosodic information to model contextual relationships;

[0066] S4.3.3, Spectrum Output Layer: Generates a sequence of mel spectrograms corresponding to the speech;

[0067] S4.3.4, Vocoder Module: Inputs the mel spectrum into the HiFi-GAN neural vocoder to finally generate a high-fidelity speech waveform signal.

[0068] This invention also discloses a BeiDou short message voice communication system based on initial and final semantic coding, comprising:

[0069] The speech front-end processing module completes the extraction of initials, finals, and prosodic semantic features;

[0070] The semantic coding module, deployed at the transmitter, enables real-time encoding and compression.

[0071] The BeiDou short message communication module completes the link encapsulation and transmission of coded data;

[0072] The decoding and reconstruction module is deployed at the receiving end to restore the voice.

[0073] The end-to-end system supports low-latency, high-compression, and highly robust transmission of voice information in the BeiDou RDSS environment through hardware and software integration.

[0074] The present invention also discloses a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method of the present invention.

[0075] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0076] 1. This invention proposes a semantic coding communication method based on a ternary structure of initial consonant-final vowel-prosody. It parses Chinese speech into the smallest semantic units at the syllable level and performs efficient compression using a fixed-length structure. Compared to existing frame-level parameter compression methods (such as linear predictive coding and code-excited linear predictive coding), this method models at the language structure level, significantly improving speech compression efficiency, reducing communication load, and is suitable for ultra-low bit-rate communication links such as BeiDou short message service.

[0077] 2. This invention addresses the challenge of jointly modeling speech content and prosodic expression by pronouncing a speech synthesis method driven by the fusion of Pinyin and prosodic feature parameters. By inputting the decoded Pinyin sequence along with fundamental frequency, energy, and duration features into an improved FastSpeech2 speech synthesis model and combining it with a HiFi-GAN vocoder, the method achieves the restoration of speech content, rhythm, and emotional features. Compared to existing speech compression schemes that only reconstruct speech outlines, this invention can reconstruct more natural and intelligible speech content. Attached Figure Description

[0078] Figure 1 This is a schematic diagram of the process described in Embodiment 1 of the present invention.

[0079] Figure 2 This is a schematic diagram of the implementation process of the speech reconstruction technology solution described in Embodiment 1 of the present invention.

[0080] Figure 3 This is a simulation diagram comparing the compression performance of the initial and final semantic coding scheme described in Embodiment 1 of the present invention with that of AMR and SPEEX coding methods under different sampling rate conditions on the original speech data.

[0081] Figure 4 This is a schematic diagram of the process described in Embodiment 2 of the present invention. Detailed Implementation

[0082] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0083] like Figure 1 As shown, the present invention provides a BeiDou short message voice communication method based on initial and final semantic coding, comprising the following steps:

[0084] S1: Preprocess the input speech signal and extract frame-level semantic features of initials and finals;

[0085] S2: Based on the extracted initial and final categories and prosodic features, construct syllable-level semantic coding triples;

[0086] S3: Syllable-level semantic unit compression encoding, and encapsulation into structured data conforming to the BeiDou short message format;

[0087] S4: Decode the received semantically encoded data and reconstruct the speech signal based on the pinyin sequence and prosody control parameters.

[0088] Furthermore, step S1 includes:

[0089] S11: Speech Preprocessing

[0090] This step performs preprocessing operations such as pre-emphasis, framing, and windowing on the original speech sampling signal to enhance the high-frequency components of the speech signal and improve the expressive power of short-time spectral energy changes, thereby improving the accuracy of subsequent feature extraction and semantic coding.

[0091] First, a first-order high-pass filter is applied to the original speech sampling signal x[n] to obtain the pre-emphasized speech signal y[n], whose expression is as follows:

[0092] y[n] = x[n] - αx[n-1]

[0093] Wherein, α is the pre-emphasis coefficient, with a value ranging from 0.95 to 0.97, used to enhance the short-term high-frequency energy variation characteristics of the speech signal.

[0094] Then, the continuous speech signal is divided into several short time frames to approximate the quasi-stationary characteristics of the speech signal, facilitating subsequent frequency domain feature extraction and encoding operations. The nth sampling point of the t-th frame is defined as:

[0095] y t (n)=y[t·H+n],n=0,1,...,N-1

[0096] Where t is the frame index, H is the frame shift, and N is the frame length.

[0097] Finally, to reduce discontinuities between frames and avoid spectral leakage, windowing is applied. The windowed frame signal is represented as follows:

[0098]

[0099] Where w(n) is the Hamming window function, its expression is as follows:

[0100]

[0101] After the above steps, a structured short-time frame sequence is obtained for subsequent extraction of initials, finals, and prosodic semantic features.

[0102] S12: Frame-level initial consonant feature extraction, specifically including the following steps: First, the preprocessed speech frame sequence... A 13-dimensional MFCC (Mel-Frequency Cepstral Coefficients) and its first and second-order difference features are extracted to form a 39-dimensional frame-level vector. The current frame is then concatenated with the context of the two frames before and after it, resulting in a 195-dimensional input vector. This 195-dimensional input vector is then fed into a lightweight DNN model for consonant classification. The DNN structure contains two hidden layers, each with 64 nodes, using ReLU activation. The output layer employs the softmax function, outputting a 22-dimensional probability vector. The probabilities correspond to 21 commonly used Chinese initials (such as b, p, m, f, d, etc.) and the probability of no initials. The total number of parameters in the model is controlled within 100,000 to adapt to the computing power constraints of Beidou portable terminals.

[0103] S13: Frame-level final feature extraction, specifically including the following steps: First, the same 195-dimensional MFCC enhancement feature as that used for initial consonant extraction is used as input. However, since finals usually cover the entire main vowel region of a syllable, a main vowel region recognition window is introduced. Multiple consecutive frames within the sliding window are selected as candidate regions representing final features. These frames are then weighted and averaged to construct a steady-state input, enhancing the stability of the expression. Next, this average feature sequence is input into a single-layer LSTM (Long Short-Term Memory) model to capture inter-frame semantic evolution features. The number of hidden units in the LSTM is set to 64. Subsequently, a softmax output layer is connected, corresponding to 35 commonly used Chinese finals (such as a, o, ai, e, ian, ü, etc.) and the case without finals. The output is a 36-dimensional normalized probability distribution vector. This LSTM model introduces a gated sparse connection structure and supports a dynamic pruning mechanism, which significantly compresses the number of model parameters while maintaining performance, in order to meet the computing power constraints of Beidou portable terminals.

[0104] Furthermore, step S2 includes:

[0105] This step, based on the speech frame sequence obtained from S1 preprocessing and the extracted frame-level initial and final features, further extracts stable syllable-level phonetic structures and corresponding prosodic features to form syllable-level semantic units. Specifically, it includes the following sub-steps:

[0106] S21: Sliding window smoothing process, constructing a joint probability vector of initials and finals for frame t. The smoothing vector is calculated using a moving average with a window length of w (e.g., 5 frames). for:

[0107]

[0108] This smoothing helps to enhance sensitivity to real semantic fluctuations while suppressing short-term model errors.

[0109] The criteria for determining "syllable boundaries" are defined as follows: when the KL divergence shows a significant change or the Top-1 predicted class changes abruptly, the current position t is considered the boundary of the k-th syllable, a new syllable segment is initiated, and the termination frame of the previous syllable segment is recorded as t. The starting frame of the new syllable segment is

[0110] The judgment criterion can be expressed as:

[0111]

[0112] Where δ is an empirical threshold (e.g., 0.5–1.0), This represents the probability vectors for smoothing consonants in frame t. Smooth probability vector of vowels Select the index corresponding to the maximum value from the list, and combine them to obtain the initial and final index pairs. These are the initial and final index values, respectively.

[0113] The initial consonant index is represented as follows:

[0114] 0→null, 1→b, 2→p, 3→m, …, 21→s

[0115] "null" indicates the case where there is no initial consonant.

[0116] The vowel index is represented as follows:

[0117] 0→null, 1→a, 2→o, 3→e, 4→i,…, 35→üan

[0118] "null" indicates the case where there is no vowel.

[0119] S22: Syllable Pinyin Unit Generation. In each syllable segment, the smooth probability vectors of the initials and finals are aggregated at the frame level (extracting the most likely Pinyin combination corresponding to the syllable segment). That is, the initial and final indices of the k-th syllable segment can be obtained by the following formulas:

[0120]

[0121] in, Let t be the consonant smoothing probability vector of the t-th frame. The probability of the corresponding initial consonant category index i, i∈{0,1,...,21}. Let t be the vowel smoothing probability vector. The probability of the corresponding vowel category index j, j∈{0,1,...,35}.

[0122] S23: Prosodic feature extraction, in the k-th syllable segment Within the frame feature sequence, three prosodic features are extracted: fundamental frequency, sound energy, and duration.

[0123] Fundamental frequency extraction: For frame segments within syllable boundaries, the YIN algorithm is used to estimate the fundamental frequency value for each frame. Non-voice frames are filtered out. The fundamental frequency of all valid frames is averaged to obtain the fundamental frequency characteristics of the syllable.

[0124]

[0125] Where 1[·] is the indicator function, which is only included in the average for frames with a fundamental frequency, and N f This is the number of valid frames.

[0126] Energy calculation: Take the 0th dimension component E(t) (representing the overall energy) from the MFCC features of each frame in the k-th syllable segment, and then take the average value to obtain the average energy e of that syllable segment. (k) :

[0127]

[0128] Pitch length calculation: based on the start and end points of the syllable boundary frame. and Combining the frame shift H, the duration of the k-th syllable segment is calculated as follows:

[0129]

[0130] Finally, the prosodic features of each syllable are represented as a three-dimensional vector.

[0131] S24: Construct the smallest semantic unit at the syllable level. The semantic content of the k-th syllable segment can be combined into a triplet form:

[0132]

[0133] Each U (k) As the smallest transmission unit of the BeiDou short message coding system, it provides dual information support of phonemes and prosody for subsequent speech reconstruction.

[0134] Furthermore, step S3 includes:

[0135] To address the syllable-level transmission structure proposed in this invention, a minimum coding unit compression scheme and message encapsulation format with high compression ratio, well-defined structure, and effective transmission over the BeiDou short message channel are designed. This scheme balances transmission reliability and decoding recoverability, constructing an effective compressed representation of speech semantic information.

[0136] Each syllable's encoding unit consists of three parts: initial consonant category ID, final vowel category ID, and prosodic feature parameters. The initial consonant category ID indicates the initial consonant category of the syllable and corresponds to the index of the predefined initial consonant set; the final vowel category ID indicates the final vowel category of the syllable and corresponds to the index of the predefined final vowel set; the prosodic feature parameters include three elements: the fundamental frequency, energy, and syllable duration.

[0137] Therefore, the encoding unit for each syllable can be represented as:

[0138]

[0139] in, These are the initial consonant category ID and the final vowel category ID, respectively, indexed by the initial consonant and final vowel category of the k-th syllable. and The binary encoding structure, E (k) D (k) They represent their corresponding fundamental frequencies. Energy e (k) With duration d (k) The value after quantization.

[0140] Specifically, the initial consonant and final vowel category indices are encoded using 5 bits and 6 bits respectively to cover the standard Pinyin category set. To adapt to the maximum data payload limit of BeiDou short messages, the three prosodic feature parameters of each syllable are encoded using discretization and fixed-point quantization methods respectively. To normalize the logarithmic compression of the syllable average fundamental frequency value, 6-bit non-uniform quantization encoding is used. (k) By calculating the mean energy of syllable segments, taking the logarithm, and then performing non-uniform coding based on the corpus distribution, it is represented by 6 bits. D (k) 3-bit uniform quantization encoding is used, with the number of frames as the unit. The above encoding bit width can be flexibly adjusted according to the communication link and application requirements. Specifically, when... and When both values ​​are 0, the system determines that the syllable is a silent syllable. E (k) The fixed codes are 000000 (6 bits) and 000000 (6 bits) respectively.

[0141] All syllable coding units will be combined sequentially and encapsulated into the payload of a single BeiDou short message. To support protocol identification, version compatibility, and data parsing, the following application layer protocol data frame structure is defined:

[0142]

[0143] The application layer header contains information such as protocol identifier (1 byte), version number (1 byte), and timestamp (3 bytes). The number of syllables M is a 2-byte integer representing the number of syllables encapsulated in this frame. The syllable encoding sequence consists of M smallest encoding units arranged in order. The CRC checksum (2 bytes) is used for integrity verification at the receiving end.

[0144] According to the BeiDou-3 short message communication protocol, under specific frequency bands and authorized application conditions, the system supports a maximum transmission payload of 14,000 bits (1,750 bytes). In this invention, the syllable-level coding unit is 26 bits in size, which can be uniformly encapsulated into a fixed-length structure of 4 bytes (32 bits) for easy boundary alignment and decoding. The remaining 6 bits can be reserved as a field for subsequent expansion (such as voiceprint encoding, tone identification, etc.). Under the 1,750-byte upper limit, after deducting the application layer protocol header (protocol identifier, version number, timestamp, totaling 5 bytes), the syllable number segment (2 bytes), and the CRC checksum (2 bytes), totaling 9 bytes, a maximum of 435 syllables can be encapsulated. Calculated at an average speech rate of 4-6 syllables / second, a single short message can carry approximately 70-100 seconds of effective voice content, which is sufficient to meet the needs of routine task reporting, emergency dispatch, and brief interactive communication.

[0145] Furthermore, step S4 includes:

[0146] To achieve complete closed-loop decoding and speech reconstruction of the BeiDou short message voice communication system based on initial and final semantic coding, this invention designs an integrated reconstruction process of pinyin recombination, prosody restoration and speech synthesis at the receiving end.

[0147] The receiving end first parses the format of the BeiDou short message data frame, and according to the encoding structure defined in step S3, parses each 32-bit fixed-length minimum unit into... Triples. Among them, This represents the initial consonant index corresponding to the decoded syllable encoding unit of the i-th syllable. This represents the vowel index corresponding to the decoded syllable encoding unit of the i-th syllable. This represents the corresponding prosodic feature encoding vector, which includes the quantized values ​​of the fundamental frequency, energy, and duration parameters.

[0148] The decoding process supports batch processing, which can quickly restore consecutive short messages into complete syllable sequences and their prosodic feature sets.

[0149] Then, based on the predefined initial consonant index table and final vowel index table, each initial consonant index obtained from decoding is... Index of finals These are mapped to standard Chinese Pinyin characters (e.g., initials "b", "d", "zh", finals "a", "ai", "uan", etc.). Then, the initials and finals are concatenated according to pronunciation rules to form a syllable-level Pinyin text sequence, which drives the subsequent speech synthesis process. To improve the legality of Pinyin combinations and the accuracy of speech synthesis, a Pinyin dictionary comparison mechanism and syllable legality verification rules are introduced to verify and correct syllable combinations, ensuring that the generated Pinyin sequence conforms to language norms. Simultaneously, for each syllable, a prosodic vector is generated... The quantized integer values ​​are restored to continuous physical quantities based on a predefined quantization inverse mapping table or linear mapping function. By accumulating the duration of each syllable, a syllable-level prosodic tag sequence can be constructed on the time axis, providing temporal information guidance for the speech synthesis model.

[0150] A speech synthesis model based on dual input control of pinyin and prosody is constructed, using an improved FastSpeech2 network as the core generation module. This model takes the pinyin sequence as the main input and prosodic feature parameters as auxiliary control input. After passing through embedding encoding and temporal modeling modules, it outputs the corresponding mel spectrogram. The specific steps are as follows:

[0151] 1) Embedding layer: Maps Pinyin characters to high-dimensional vectors, and embeds prosodic feature parameters (fundamental frequency, energy, duration) into control vectors;

[0152] 2) Transformer encoder-decoder structure: fuses semantic and prosodic information to model contextual relationships;

[0153] 3) Spectrum output layer: Generates a sequence of mel spectrograms corresponding to the speech;

[0154] 4) Vocoder module: Inputs the mel spectrum into the HiFi-GAN neural vocoder to finally generate a high-fidelity speech waveform signal.

[0155] A BeiDou short message voice communication system based on initial and final semantic coding, the system comprising:

[0156] The speech front-end processing module completes the extraction of initials, finals, and prosodic semantic features;

[0157] The semantic coding module, deployed at the transmitter, enables real-time encoding and compression.

[0158] The BeiDou short message communication module completes the link encapsulation and transmission of coded data;

[0159] The decoding and reconstruction module is deployed at the receiving end to restore the voice.

[0160] The end-to-end system supports low-latency, high-compression, and highly robust transmission of voice information in the BeiDou RDSS environment through hardware and software integration.

[0161] Example:

[0162] In field operation scenarios, in order to achieve efficient voice communication in remote environments, the system deploys a semantic coding communication terminal based on the present invention, which has functions such as voice acquisition, encoding compression, Beidou short message transmission and remote voice restoration, and supports real-time voice transmission under conditions of no public network coverage and limited link bandwidth.

[0163] The system consists of: an embedded voice acquisition module, a semantic coding module (DNN+LSTM network), a Beidou RDSS communication module, a decoding and reconstruction module (improved FastSpeech2+HiFi-GAN model), and a terminal playback module.

[0164] The operation process is as follows:

[0165] 1. Voice Acquisition and Preprocessing

[0166] Operators input voice signals via microphones, and the terminal acquires the audio signals in real time at a sampling rate of 16kHz. The system uses a 25ms frame length and a 10ms frame shift parameter to pre-emphasize, frame, and window the original speech, forming a short-time frame sequence.

[0167] 2. Extraction of initials, finals, and prosodic features

[0168] For each frame, MFCC features and their first and second differences are calculated, forming a 39-dimensional feature vector. After context concatenation, this vector is input into a lightweight DNN model to extract the probability distributions of 22 initial consonants. Simultaneously, multi-frame MFCC sequences within syllable segments are extracted and input into an LSTM network to obtain the probability distributions of 36 finals. Syllable phonetic units are generated, and the YIN algorithm and energy analysis method are used to extract the fundamental frequency, energy, and duration of each syllable segment as prosodic features.

[0169] 3. Semantic encoding and encapsulation

[0170] The system constructs triples from the initial and final vowel indices and quantized prosodic feature parameters, and encapsulates these triples into 32-bit fixed-length coding units, syllable by syllable. All coding units are sequentially arranged to form the payload of a short message. The message structure includes a protocol header, syllable count, coding sequence, and CRC checksum to ensure communication integrity.

[0171] 4. Short message transmission

[0172] After encoding, the short message is sent to the command center through the BeiDou-3 RDSS link. The typical encoded frame size does not exceed 14,000 bits, and the transmission delay is less than 1 second, which meets the timeliness requirements of emergency response.

[0173] 5. Receiver decoding and speech synthesis

[0174] After receiving the short message, the command center parses the initial and final characters and prosodic vectors based on a predefined pinyin / prosodic inverse mapping table, and inputs them into the improved FastSpeech2 speech synthesis model to generate a mel spectrogram. The speech waveform is then reconstructed in real time using a HiFi-GAN neural vocoder, and finally, the reconstructed speech is played back through a speaker.

Claims

1. A BeiDou short message voice communication method based on initial and final semantic encoding, characterized in that, Includes the following steps: S1: Preprocess the input speech signal and extract frame-level initial consonant and final vowel features; S2: Based on the extracted frame-level initial consonant and final vowel features, construct syllable-level semantic coding triples; S3: Based on syllable-level semantic coding triples, syllable-level semantic units are compressed and encoded, and encapsulated into structured data conforming to the BeiDou short message format to obtain semantic coding data; Step S3 specifically involves: Each speech syllable coding unit consists of three parts: initial consonant category ID, final vowel category ID, and prosodic feature parameters; the initial consonant category ID represents the initial consonant category of the syllable and corresponds to the index of the predefined initial consonant set; the final vowel category ID represents the final vowel category of the syllable and corresponds to the index of the predefined final vowel set; the prosodic feature parameters include three elements: the fundamental frequency, energy, and syllable duration corresponding to the syllable; Therefore, the encoding unit for each syllable is represented as: ； in, , These are the initial consonant category ID and the final vowel category ID, respectively, derived from the first... Index of initials and finals for each syllable and The binary encoding structure, , , They represent their corresponding fundamental frequencies. ,energy With duration The quantized value; All syllable coding units will be combined sequentially and encapsulated into the payload of a BeiDou short message. To support protocol identification, version compatibility, and data parsing, the following application layer protocol data frame structure is defined: ； Among them, the application layer It contains information including a 1-byte protocol identifier, a 1-byte version number, and a 3-byte timestamp, along with the number of syllables. This is a 2-byte integer representing the number of encapsulated syllables in this frame. The syllable encoding sequence is arranged sequentially. The smallest coding unit is 2 bytes. The checksum is used for integrity verification at the receiving end; The initial consonant category ID and final vowel category ID are encoded using 5 bits and 6 bits respectively to cover the standard pinyin category set; to adapt to the maximum data payload limit of BeiDou short messages, the three prosodic feature parameters of each syllable are encoded using discretization and fixed-point quantization methods respectively; among which, To normalize the logarithmic compression of the syllable average fundamental frequency value, 6-bit non-uniform quantization encoding is used; The mean energy of syllable segments is calculated, and after taking the logarithm, non-uniform coding is performed based on the corpus distribution, represented by 6 bits. Using 3-bit uniform quantization encoding as the unit of frame count; the encoding bit width is adjusted according to the communication link and application requirements; when and When both values ​​are 0, the system determines that the syllable is a silent syllable. , The fixed codes are 000000 (6 bits) and 000000 (6 bits) respectively; S4: Decodes the semantically encoded data and reconstructs the speech signal based on the pinyin sequence and prosody control parameters to realize BeiDou short message voice communication.

2. The BeiDou short message voice communication method based on initial and final semantic coding according to claim 1, characterized in that, In step S1, the preprocessing of the input speech signal specifically involves: processing the original speech sampling signal... A first-order high-pass filter is performed to obtain the pre-emphasized speech signal. Its expression is as follows: ； in, This is the pre-emphasis factor, with a value ranging from 0.95 to 0.97, used to enhance the short-term high-frequency energy variation characteristics of the speech signal; The continuous speech signal is divided into several short time frames, the first... The first frame Each sampling point is defined as: ； in, For frame index, For frame shift, Frame length; Windowing is applied to reduce discontinuities between frames. The windowed frame signal is represented as follows: ； in, The Hamming window function has the following expression: ； The resulting structured short-time frame sequences are used for frame-level initial consonant and final vowel feature extraction.

3. The BeiDou short message voice communication method based on initial and final semantic encoding according to claim 2, characterized in that, In step S1, the extraction of frame-level initial consonant features and final vowel features specifically includes the following steps: S1.1, Frame-level initial consonant feature extraction: First, the preprocessed speech frame sequence... The 13-dimensional Mel-frequency cepstral coefficients (MFCCs) and their first and second-order difference features are extracted to form a 39-dimensional frame-level vector. The current frame is then concatenated with the context of the two frames before and after it, resulting in a 195-dimensional input vector. This 195-dimensional input vector is then fed into a lightweight DNN model for consonant classification. The DNN structure contains two hidden layers, each with 64 nodes, using ReLU activation. The output layer employs the softmax function, outputting a 22-dimensional probability vector. These correspond to the probabilities of 21 commonly used Chinese initials and cases without initials, respectively; the total number of DNN model parameters is controlled within 100,000 to adapt to the computing power constraints of Beidou portable terminals. S1.2 Frame-level Final Feature Extraction: First, the same 195-dimensional MFCC enhanced features as those used for initial consonant extraction are used as input. Since finals are usually distributed in the main vowel region of a syllable, a main vowel region recognition window is introduced to avoid relying on explicit syllable boundaries. Multiple consecutive frames are selected within the sliding window as a candidate frame set for finals, and a steady-state final feature representation for this segment is constructed using a weighted average method, thereby improving modeling robustness. Then, this average feature sequence is input into a single-layer Long Short-Term Memory (LSTM) network model to capture inter-frame semantic evolution features. The LSTM has 64 hidden units; it is then connected to a softmax output layer, corresponding to 35 common Chinese vowels and the case without vowels in linguistics, and the output is a 36-dimensional normalized probability distribution vector. The LSTM model introduces a gated sparse connection structure and supports a dynamic pruning mechanism.

4. The BeiDou short message voice communication method based on initial and final semantic encoding according to claim 1, characterized in that, Step S2 specifically includes the following steps: S2.1 Smoothing of sliding windows: For the first Frame construction of joint probability vector of initials and finals Using a window length of The moving average is used to calculate the smoothing vector. for: ； The syllable boundary determination criteria are defined as follows: when the KL divergence shows a significant change or the Top-1 prediction category changes abruptly, the current position is considered to be... For the first At the boundary of a syllable, initiate a new syllable segment and record the ending frame of the previous syllable segment as... The starting frame of the new syllable segment is ; The judgment criteria are expressed as follows: ； in, As an empirical threshold, Indicates the first In the frame, from the initial consonant smoothing probability vector respectively Smooth probability vector of vowels Select the index corresponding to the maximum value from the list, and combine them to obtain the initial and final index pairs. , These are the index values ​​of the initials and finals, respectively; The initial consonant index is represented as follows: ； in Indicates the case where there is no initial consonant; The vowel index is represented as follows: ； in Indicates the case where there is no vowel; S2.2, Syllable Pinyin Unit Generation: In each syllable segment, the smooth probability vectors of the initials and finals are aggregated at the frame level, i.e., the... The initial and final indexes of each syllable segment are obtained using the following formulas: ； ； in, For the first Frame consonant smoothing probability vector Chinese corresponding initial consonant category index The probability, , For the first Frame vowel smoothing probability vector Chinese corresponding vowel category index The probability, ; S2.3, Prosodic Feature Extraction: In the first... syllable segment Within the frame feature sequence, three prosodic features are extracted: fundamental frequency, sound energy, and duration. Fundamental frequency extraction: For frame segments within syllable boundaries, the YIN algorithm is used to estimate the fundamental frequency value for each frame. It filters out unvoiced frames and averages the fundamental frequencies of all valid frames to obtain the fundamental frequency characteristics of the syllables. : ； in, As an indicator function, the average is only calculated for frames with a fundamental frequency. It is the number of valid frames; Energy calculation: From the first Take the 0th dimension component from the MFCC features of each frame in each syllable segment. Then, the average energy of the syllable segment is obtained by taking the average value. : ； Pitch length calculation: based on the start and end points of the syllable boundary frame. and Combined with frame shift Calculate the first The duration of each syllable segment is: ； Finally, the prosodic features of each syllable are represented as a three-dimensional vector. ; S2.4 Constructing the smallest semantic unit at the syllable level, the first... The semantic content of a syllable segment can be combined into a triplet form: ； Each As the smallest transmission unit of the BeiDou short message coding system, it provides dual information support of phonemes and prosody for subsequent speech reconstruction.

5. The BeiDou short message voice communication method based on initial and final semantic encoding according to claim 1, characterized in that, Step S4 specifically includes the following steps: S4.1 The receiving end first parses the format of the BeiDou short message data frame, and according to the defined encoding structure, parses each 32-bit fixed-length minimum unit into... Triples; among which Indicates the first The initial consonant index corresponding to each syllable encoding unit after decoding. Indicates the first The final index corresponding to each syllable encoding unit after decoding. This represents the corresponding prosodic feature encoding vector, which includes the quantized values ​​of the fundamental frequency, energy, and duration parameters; S4.

2. Based on the predefined initial consonant index table and final vowel index table, decode each initial consonant index obtained from the initial consonant index table. Index of finals Each initial and final character is mapped to a standard Chinese Pinyin character; the initials and finals are concatenated according to pronunciation rules to form a syllable-level Pinyin text sequence, which drives the subsequent speech synthesis process; to improve the legality of Pinyin combinations and the accuracy of speech synthesis, a Pinyin dictionary comparison mechanism and syllable legality verification rules are introduced to verify and correct syllable combinations as necessary, ensuring that the generated Pinyin sequence conforms to language norms; simultaneously, for each syllable, a prosodic vector is generated. The quantized integer values ​​are restored to continuous physical quantities according to a predefined quantization inverse mapping table or linear mapping function; by accumulating the duration of each syllable, a syllable-level prosodic tag sequence is constructed on the time axis to provide temporal information guidance for the speech synthesis model; S4.3 Construct a speech synthesis model based on dual input control of pinyin and prosody, using an improved FastSpeech2 network as the core generation module; the speech synthesis model uses the pinyin sequence as the main input and the prosody feature parameters as the auxiliary control input, and after passing through the embedding encoding and temporal modeling modules, it outputs the corresponding mel spectrogram.

6. The BeiDou short message voice communication method based on initial and final semantic coding according to claim 5, characterized in that, In step S4.3, the construction of the speech synthesis model based on dual input control of pinyin and prosody specifically includes the following steps: S4.3.1 Embedding layer: Maps Pinyin characters to high-dimensional vectors, and embeds prosodic feature parameters into control vectors. Prosodic feature parameters include fundamental frequency, energy, and duration. S4.3.2 Transformer encoder-decoder structure: fusing semantic and prosodic information to model contextual relationships; S4.3.3, Spectrum Output Layer: Generates a sequence of mel spectrograms corresponding to the speech; S4.3.4, Vocoder Module: Inputs the mel spectrum into the HiFi-GAN neural vocoder to finally generate a high-fidelity speech waveform signal.

7. A BeiDou short message voice communication system based on initial and final semantic encoding, used to implement the method as described in claim 1, characterized in that, include: The speech front-end processing module completes the extraction of initials, finals, and prosodic semantic features; The semantic coding module, deployed at the transmitter, enables real-time encoding and compression. The BeiDou short message communication module completes the link encapsulation and transmission of coded data; The decoding and reconstruction module is deployed at the receiving end to restore the voice. The end-to-end system supports low-latency, high-compression, and highly robust transmission of voice information in the BeiDou RDSS environment through hardware and software integration.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method of claim 1.

Images