A voice entropy encoding method and system for multi-level emergency broadcasting and communication

Abstract

The application discloses a voice entropy coding method and system for multi-level emergency broadcasting and communication, which comprises the following steps: input voice is coded into a discrete voice Token sequence by a limited scalar quantization tokenizer; autoregressive modeling is performed based on the structured representation of the multidimensional scalar quantization coordinates of historical Tokens; the joint conditional probability of a current Token is calculated by a product category distribution mixture model to predict the conditional probability distribution of the current Token; the conditional probability distribution is quantized into an integer frequency, a first integer cumulative frequency table is constructed to perform entropy coding on the current Token; after the entropy coding of the voice Token sequence is sequentially completed, a compressed Token load is output and sent to a receiving end through a transmission link, so that the receiving end restores the original voice Token sequence and reconstructs the voice through a detokenizer. The application can be connected to an existing voice communication link as an independent conditional entropy coding and decoding module, and is suitable for disaster voice reporting, emergency dispatch instruction issuing and voice communication between multi-level emergency nodes under a bandwidth limited condition.

Description

Technical Field

[0001] This invention relates to the field of emergency management technology, and in particular to a voice entropy coding method and system for multi-level emergency broadcasting and communication. Background Technology

[0002] In extreme disaster scenarios such as earthquakes, floods, typhoons, and major accidents, ground base stations, fiber optic links, and power supply facilities may be damaged, and emergency communication links typically suffer from bandwidth limitations, network congestion, and unstable communication resources. Simultaneously, on-site voice communication needs to share limited communication resources with data such as personnel location, telemetry, images, text, and equipment status. Therefore, reducing the transmission bandwidth consumed by voice services is crucial for ensuring the timely transmission of disaster reports and dispatch instructions.

[0003] Existing low-bit-rate speech coding technologies mainly include two categories: traditional parametric speech coding and neural speech coding. Traditional methods typically extract acoustic parameters such as fundamental frequency, spectral envelope, line spectrum frequency, and excitation parameters for encoding. However, under extremely low bit-rate conditions, these methods are prone to problems such as decreased naturalness, increased mechanical sound, and weakened speaker characteristics.

[0004] In recent years, neural speech coding methods have achieved good speech reconstruction quality at lower bit rates by converting speech waveforms into discrete tokens and then using neural networks to reconstruct speech based on the tokens. For example, some neural audio encoders use residual vector quantization (RVQ), which quantizes the speech representation step by step through multiple codebooks. However, each time frame usually needs to transmit multiple codebook indices, and there are dependencies between different quantization levels.

[0005] Another approach uses Finite Scalar Quantization (FSQ) to generate low frame rate discrete speech tokens. Taking existing FSQ speech coding as an example, each token can be determined by multiple finite scalar quantization dimensions, forming a discrete token vocabulary. For an FSQ token with a vocabulary size of 6561, if fixed-length encoding is used directly, each token requires at least 13 bits for representation. At a token rate of approximately 25 tokens / s, the corresponding token payload code rate is approximately 325 bits / s. Existing FSQ tokens typically allocate a fixed number of bits according to the vocabulary size during direct transmission, using the same code length regardless of the probability of a token's occurrence, thus still exhibiting inefficiency. Furthermore, since FSQ tokens are actually formed by combining multiple scalar quantization coordinates, if a token is directly used as an unstructured category number as an ordinary discrete number input into the probability model, it is difficult for different tokens sharing some quantization coordinates to fully share statistical information.

[0006] Furthermore, a single marginal probability distribution can be established based on the frequency of occurrence of each token in the training data, and entropy encoding can be performed based on this probability distribution. While this method can take advantage of the uneven probability of different tokens, it uses the same probability distribution for different time positions, fails to utilize the temporal correlation between adjacent speech tokens, and does not explicitly utilize the combination structure of multiple scalar quantization coordinates within the FSQ token. Speech is a signal with a continuous temporal structure, and the current speech token usually has a strong correlation with previously occurring tokens. Simply calculating a fixed token probability distribution based on the entire training set cannot dynamically adjust the encoded token probability according to the current context.

[0007] Furthermore, if tokenizers and detokenizers have already been deployed, retraining or even replacing the entire speech codec system would increase the costs of model training, validation, deployment, and terminal upgrades, which would be detrimental to incremental upgrades of existing emergency broadcasting and communication dispatch systems.

[0008] Therefore, it is necessary to fully utilize the internal structure and temporal correlation of the FSQ Token to perform low-bitrate lossless compression of the voice Token without retraining the original tokenizer and detokenizer. Summary of the Invention

[0009] The purpose of this invention is to provide a voice entropy coding method and system for multi-level emergency broadcasting and communication.

[0010] To achieve the above objectives, in a first aspect, the present invention provides a speech entropy coding method for multi-level emergency broadcasting and communication, the method comprising:

[0011] S1. Obtain the input speech to be transmitted, and encode the input speech into a discrete speech token sequence through a finite scalar quantization tokenizer. Each token in the speech token sequence is determined by a combination of quantization coordinates of multiple scalar quantization dimensions, and each scalar quantization dimension has multiple finite quantization levels.

[0012] S2. Autoregressive modeling is performed based on the structured representation of the multidimensional scalar quantized coordinates of historical tokens. The joint conditional probability of the current token is calculated through a product category distribution mixture model, and the conditional probability distribution of the current token is predicted.

[0013] S3. Quantize the conditional probability distribution into integer frequencies and construct a first integer cumulative frequency table; perform entropy encoding on the current Token according to the first integer cumulative frequency table;

[0014] S4. After performing entropy encoding on the speech token sequence in sequence, output the compressed token payload; send the compressed token payload to the receiving end through the transmission link so that the receiving end can recover the original speech token sequence and reconstruct the speech through the detokenizer.

[0015] Specifically, a one-way long short-term memory network is used to establish the time dependency relationship between historical tokens. The structured input representation, hidden state and cell state corresponding to the previous token are used as inputs, and the hidden state and cell state of the current token are output.

[0016] The product category distribution hybrid model outputs multiple hybrid components based on the current context representation obtained by the autoregressive modeling. Each quantized coordinate in each hybrid component is modeled using a category distribution. The joint probability of the hybrid component is calculated by the product of the category probabilities of each quantized coordinate. Then, the joint probability of each hybrid component is weighted and summed to obtain the joint conditional probability of the current token.

[0017] The multidimensional scalar quantized coordinates of the previous token are converted into low-dimensional feature vectors after fully connected projection and layer normalization, which serve as the input representation of the temporal probability model.

[0018] In the first integer cumulative frequency table, the integer frequency of each Token is not less than 1, so that each Token corresponds to a valid entropy coding interval.

[0019] The encoder recursively updates the entropy encoding interval according to the cumulative frequency interval corresponding to the Token, and after completing the entropy encoding of the entire discrete speech Token sequence and encoding the sequence terminator, the compressed Token payload is formed.

[0020] Before each speech encoding begins, the encoder is restored to a preset initial interval, the hidden state and cell state of the temporal probability model are initialized to zero, and the sequence start symbol is used as the initial input of the temporal probability model; the encoder state and temporal probability model state are not inherited between different speech.

[0021] The conditional entropy model is trained using a teacher-mandated method, with the training objective being to minimize the negative log-likelihood of the speech token sequence; the parameters of the finite scalar quantizer are kept frozen during training.

[0022] Secondly, this invention provides a voice entropy decoding method for multi-level emergency broadcasting and communication, the method comprising:

[0023] R1 receives the compressed token payload output by the sender;

[0024] R2. Based on the structured representation of the multidimensional scalar quantized coordinates of the recovered historical tokens, the joint conditional probability of the current token is regenerated using the conditional entropy model consistent with the sending end, and the joint conditional probability is quantized into integer frequencies. The sequence terminator is included in the second integer cumulative frequency table according to the rules consistent with the sending end, and the second integer cumulative frequency table is constructed.

[0025] R3. Perform entropy decoding on the compressed Token payload according to the second integer cumulative frequency table, gradually recover the current Token, and add the recovered current Token to the historical Token until the end of the recovery sequence is recovered;

[0026] R4. Input the recovered original speech token sequence into the detokenizer to obtain the reconstructed speech.

[0027] Thirdly, the present invention provides a voice entropy coding system for multi-level emergency broadcasting and communication, comprising: a transmitting side processing module and a receiving side processing module, wherein the transmitting side processing module and the receiving side processing module are connected via a transmission link for exchanging compressed token payloads;

[0028] The transmitting side processing module includes: a token generation module, used to convert the input speech into a discrete speech token sequence through a finite scalar quantization tokenizer, wherein each token in the speech token sequence is determined by a combination of quantization coordinates of multiple scalar quantization dimensions, and each scalar quantization dimension has multiple finite quantization levels;

[0029] The conditional probability modeling module, connected to the token generation module, is used to perform autoregressive modeling based on the structured representation of the multidimensional scalar quantized coordinates of historical tokens, calculate the joint conditional probability of the current token through a product category distribution mixture model, and predict the conditional probability distribution of the current token.

[0030] An entropy encoding module, connected to the conditional probability modeling module, is used to quantize the conditional probability distribution into integer frequencies and construct a first integer cumulative frequency table; and to perform entropy encoding on the current Token according to the first integer cumulative frequency table to generate the compressed Token payload.

[0031] The receiving-side processing module includes: a conditional probability reconstruction module, used to regenerate the joint conditional probability and second integer cumulative frequency table of the current token based on the recovered historical tokens;

[0032] The entropy decoding module, connected to the conditional probability reconstruction module, is used to perform entropy decoding on the compressed token payload according to the second integer cumulative frequency table, and gradually recover the original speech token sequence until the end of the sequence is recovered.

[0033] The speech reconstruction module, connected to the entropy decoding module, is used to input the original token sequence into the detokenizer to obtain the reconstructed speech;

[0034] The transmitting-side processing module and the receiving-side processing module use the same conditional entropy model parameters, model initialization rules, probability quantization rules, integer cumulative frequency table construction rules, sequence termination rules, and entropy encoding and decoding rules.

[0035] Compared with existing technologies, the speech entropy coding method and system for multi-level emergency broadcasting and communication according to the present invention reduces the transmission load of speech tokens in emergency communication links by explicitly utilizing the multi-dimensional scalar quantization structure inside the FSQ Token, establishing the conditional probability distribution of the current token using historical tokens, and then combining entropy coding to perform post-processing lossless compression on the speech token sequence. Attached Figure Description

[0036] Figure 1This is a schematic flowchart of a voice entropy coding method for multi-level emergency broadcasting and communication according to an embodiment of the present invention;

[0037] Figure 2 This is a schematic flowchart of a voice entropy decoding method for multi-level emergency broadcasting and communication according to an embodiment of the present invention;

[0038] Figure 3 This is a schematic diagram of a voice entropy coding system for multi-level emergency broadcasting and communication according to an embodiment of the present invention. Detailed Implementation

[0039] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0040] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0041] This invention discloses a speech entropy coding method for multi-level emergency broadcasting and communication, such as... Figure 1 As shown, the method includes:

[0042] S1. Acquire the input speech to be transmitted and encode it into a discrete speech token sequence using a finite scalar quantization (FSQ) tokenizer. Each token in the speech token sequence is determined by a combination of quantization coordinates of multiple scalar quantization dimensions, and each scalar quantization dimension has multiple finite quantization levels. Specifically, the transmitting end acquires the speech waveform to be transmitted and inputs it into the finite scalar quantization (FSQ) tokenizer to generate a discrete speech token sequence of length T. Each time step t in the speech token sequence corresponds to a token, which is determined by a combination of quantization coordinates of D scalar quantization dimensions.

[0043] In one specific implementation, each token of the FSQ tokenizer consists of 8 scalar quantization dimensions, each dimension having 3 quantization levels. Therefore, it can form Different types of tokens. This invention does not treat the token as a simple integer number, but restores the token to multiple corresponding scalar quantized coordinates to explicitly preserve the internal combination structure of the token.

[0044] Preferably, the quantization vector corresponding to a token is represented as follows:

[0045] , i = 1, 2, ..., D

[0046] This represents the FSQ quantization vector corresponding to time step t; This represents the quantization coordinate of the FSQ Token at time step t on the Dth scalar quantization dimension. The quantization coordinate of the token at time step t on the i-th scalar quantization dimension is represented; t represents the time step index in the token sequence; i represents the scalar quantization dimension index, i=1,2,…,D; D represents the total number of scalar quantization dimensions. This represents the set of finite quantization levels corresponding to the i-th scalar quantization dimension.

[0047] S2. Autoregressive modeling is performed based on the structured representation of the multidimensional scalar quantized coordinates of historical tokens. The joint conditional probability of the current token is calculated using a product-class distribution hybrid model to predict the conditional probability distribution of the current token. Specifically, for the token at the current time step t, the sender first converts the multidimensional scalar quantized coordinates of the previous token into a low-dimensional feature vector after fully connected projection and layer normalization, which serves as the structured input representation of the temporal probability model. Fully connected projection maps discrete quantized coordinates to a continuous low-dimensional feature space, and layer normalization facilitates stable training and inference of the temporal probability model. Subsequently, the temporal probability model receives the low-dimensional feature vector corresponding to the previous token, as well as the hidden state and cell state of the previous time step, as input, and outputs the hidden state and cell state of the current time step. The product-class distribution hybrid model uses the hidden state of the current time step as the current context representation to generate the weights of each hybrid component and the class probabilities of each FSQ quantization dimension.

[0048] In a preferred embodiment, modeling is performed based on the conditional probabilities of historical tokens to predict the conditional probability of the current token in subsequent iterations. The joint probability of the voice token sequence can be decomposed chronologically as follows:

[0049]

[0050] in: This represents the joint probability given by the conditional probability model, determined by parameter θ, for the entire speech token sequence z; θ represents the conditional probability model parameter; z represents the complete speech token sequence; T represents the length of the speech token sequence; and t represents the time step index in the token sequence. This represents the current Token corresponding to time step t; This represents the historical token sequence before time step t. .

[0051] S3. Quantize the conditional probability distribution into integer frequencies and construct a first integer cumulative frequency table; perform entropy encoding on the current Token according to the first integer cumulative frequency table. Specifically, the sending end deterministically quantizes the conditional probability distribution of the current Token output by the product category distribution hybrid model into integer frequencies, and includes the sequence terminator EOS as an additional encoded symbol in the first integer cumulative frequency table according to preset rules. Specifically, the probabilities corresponding to all Tokens at time step t are deterministically quantized into integer frequencies. Construct the first integer cumulative frequency table:

[0052]

[0053] At the same time, it stipulates , ,in This represents the integer frequency assigned to candidate Token j at time step t; This represents the cumulative frequency of the k-th candidate token up to time step t; t represents the time step index; j represents the index of the candidate token; J represents the upper limit of the token index for cumulative summation; V represents the token vocabulary size; and N represents the total cumulative frequency.

[0054] After the first integer cumulative frequency table is constructed, entropy encoding is performed on the current token using arithmetic encoding. For the current token The corresponding integer cumulative frequency range is The arithmetic encoder recursively updates the encoding interval based on this interval, gradually encoding the information of the token sequence into the compressed bitstream. Arithmetic coding can perform lossless compression of the token sequence with an average code length close to the lower bound of the information entropy, and directly interface with the first integer cumulative frequency table output by the conditional probability model.

[0055] S4. After sequentially performing entropy encoding on the speech token sequence, a compressed token payload is output. The compressed token payload is then transmitted to the receiving end via a transmission link, enabling the receiving end to recover the original speech token sequence and reconstruct the speech using a detokenizer. Specifically, the sending end sequentially performs steps S2 and S3 on the entire discrete speech token sequence. After the last speech token is encoded, the sequence end symbol EOS is encoded, and the current arithmetic encoding process ends. The remaining bits in the encoder are refreshed and output to form a complete compressed token payload, which is then transmitted to the receiving end via a transmission link. In one embodiment, EOS is used as an additional entropy encoding symbol, and the conditional entropy model simultaneously provides the termination probability of EOS. When encoding EOS, the cumulative frequency table is expanded to include V+1 symbols containing V FSQ tokens and EOS. The sending end and the receiving end construct this table according to the same rules.

[0056] In a preferred implementation, the temporal probabilistic model uses a one-way long short-term memory (LSTM) network to establish the temporal dependencies between historical tokens. It takes the structured input representation, hidden state, and cell state corresponding to the previous token as input and outputs the hidden state and cell state of the current token. The recursive relationship is as follows: .in, This represents the LSTM hidden state at time step t; This indicates the LSTM cell state at time step t; This represents the FSQ structured input representation of the Token corresponding to the previous time step t−1; and Let represent the hidden state and cell state of the previous time step, respectively; t represent the time step index of the current token; and LSTM represent Long Short-Term Memory network. That is, taking the FSQ structured input representation, hidden state, and cell state corresponding to the previous token as input, the output is the hidden state and cell state of the current time step. In this way, the probability prediction of the current token can simultaneously utilize the content of all previous tokens and the temporal correlation in the token sequence, dynamically adjusting the encoding probability and overcoming the limitation of fixed marginal probability distributions that cannot utilize contextual information.

[0057] In a preferred embodiment, the product-category distribution mixture model outputs K mixture components. For each mixture component, the joint probability of that mixture component is obtained by multiplying the class probabilities of each quantized coordinate. The joint probability of the current token is obtained by weighting the joint probability of each mixture component with its respective weight. Specifically, the product-category distribution mixture model outputs K mixture components. For the k-th mixture component, each scalar quantized coordinate is independently modeled using a category distribution, and the joint probability of the mixture component is calculated by multiplying the class probabilities of each quantized coordinate. The joint probability of the mixture component is then calculated by weighting the joint probability of each mixture component with its respective weight. The joint conditional probability of the current token is obtained by weighted summation of the joint probabilities of each component. This modeling method utilizes the combined structure of multiple scalar quantized coordinates within the FSQ token, which, compared to directly using a large-scale TokenSoftmax output, covers the entire token vocabulary with a smaller parameter scale. This allows different tokens sharing some quantized coordinates to fully share statistical information, improving the accuracy of conditional probability prediction. In one specific implementation, the number of mixed components K is 128, but the specific embodiments of this invention are not limited to this value and can be flexibly adjusted according to computational resources and compression ratio requirements.

[0058] Specifically, the joint conditional probability of the current token can be expressed as:

[0059]

[0060] in: This represents the quantization vector in the historical FSQ. Current quantization vector under the condition The conditional probability; θ represents the parameters of the conditional probability model; This represents the current FSQ quantization vector corresponding to time step t; This represents the historical FSQ quantization vector sequence before time step t; t represents the time step index; K represents the total number of mixed components; k represents the mixed component index. The weight of the k-th hybrid component is represented at time step t; D represents the total number of FSQ quantization dimensions; i represents the FSQ quantization dimension index. This indicates that at time step t, the k-th mixture component is on the i-th FSQ quantization dimension. The given class probabilities; This represents the quantization coordinate of the current token at time step t in the i-th FSQ quantization dimension.

[0061] Using the above calculations, we can obtain the conditional probability distribution of all tokens at the current time step.

[0062] In a preferred embodiment, a first integer cumulative frequency table is constructed, wherein the integer frequency of each token in the first integer cumulative frequency table is not less than 1, so that each token corresponds to a valid entropy coding interval. Specifically, for each candidate token, its integer frequency is guaranteed to be not less than 1, thereby ensuring that each token corresponds to a valid entropy coding interval and preventing entropy coding failure due to token frequencies with extremely low probability being quantized to zero. In one specific embodiment, a 16-bit cumulative frequency precision is used, i.e., N=65536. In another specific embodiment, one integer frequency is reserved for each candidate coded symbol, and then the remaining frequencies are allocated deterministically according to conditional probability; if there is still a remainder after rounding, it is allocated from largest to smallest according to the decimal remainder, and if the remainders are the same, they are determined according to the token index order, until the total frequency is N. The sending end and the receiving end use the same rules.

[0063] In a preferred embodiment, the encoder recursively updates the entropy encoding interval according to the cumulative frequency interval corresponding to the token, and after encoding the entire discrete speech token sequence and the sequence terminator, forms the compressed token payload.

[0064] In a preferred embodiment, before each speech encoding begins, the encoder is restored to a preset initial interval, the hidden state and cell state of the temporal probability model are initialized to zero, and the sequence start symbol (BOS) is used as the initial input of the temporal probability model; different speech segments do not inherit the encoder state or the temporal probability model state. Specifically, to ensure that the transmitter and receiver generate completely consistent conditional probability distributions, each speech segment undergoes independent entropy encoding and entropy decoding, employing a deterministic model initialization rule. When encoding a speech segment, the arithmetic encoder is restored to a preset initial interval, the hidden state and cell state of the LSTM are initialized to zero, and the sequence start symbol (BOS) is used as the initial input of the temporal probability model. Different speech segments do not inherit the encoder state or the LSTM state. This deterministic initialization rule ensures that the transmitter and receiver are in the exact same initial state at the start of each speech encoding, allowing the receiver to synchronously reconstruct the same conditional probability without additional transmission of probability distributions or cumulative frequency tables through the transmission link, thereby achieving lossless decoding per token. The BOS uses a fixed starting feature vector pre-consistent between the transmitter and receiver and is not treated as a regular FSQ token.

[0065] In a preferred embodiment, the conditional entropy model is trained using a teacher-forced method, with the training objective being to minimize the negative log-likelihood of the speech token sequence; the parameters of the finite scalar quantizer are frozen during training. Specifically, during teacher-forced training, real historical token sequences are used as model input, and the training objective is to minimize the negative log-likelihood of the token sequence.

[0066]

[0067] in, The average negative log-likelihood loss of the speech token sequence is represented by T; the length of the speech token sequence is represented by T; and the time step index in the token sequence is represented by t. This represents the conditional probability model determined by θ in the historical token sequence. Under the condition of the current token The given conditional probabilities; θ represents the parameters of the conditional probability model; This represents the current token at time step t; This represents the historical token sequence before time step t.

[0068] For each time step t in the sequence, the log probability of the current token under the given historical token conditions is calculated, and the negative value of the average over the entire sequence is used as the loss function. During the training of the conditional entropy model, the parameters of the original FSQ tokenizer remain frozen, eliminating the need for retraining. Teacher-forced training ensures that the conditional probability model uses the real historical token sequence as input, accelerating convergence and improving model stability; freezing the tokenizer parameters ensures that there is no need to retrain or replace the original speech codec, reducing system modification costs. In one specific implementation, the conditional entropy model uses a two-layer unidirectional LSTM with 512 hidden units per layer; the FSQ input is converted to a 64-dimensional representation after passing through a fully connected layer and layer normalization; the output uses a product-class distribution hybrid model containing 128 hybrid components. This invention does not limit the conditional entropy model to the specific network size described above; other temporal probability models can also be used if the conditional probability of the current token can be obtained from historical tokens.

[0069] The decoding process at the receiving end includes the following four steps.

[0070] Specific embodiments of the present invention also provide a voice entropy decoding method for multi-level emergency broadcasting and communication, such as... Figure 2 As shown, the method includes:

[0071] R1 receives the compressed token payload output by the transmitter. Specifically, the receiver obtains the compressed token payload generated by the transmitter from the restricted / multi-standard transmission link, which serves as the input data for subsequent entropy decoding.

[0072] R2. Based on the structured representation of the multidimensional scalar quantized coordinates of the recovered historical tokens, the joint conditional probability of the current token is regenerated using a conditional entropy model consistent with that of the sender. This joint conditional probability is then quantized into integer frequencies. The sequence terminator is included in the second integer cumulative frequency table according to the same rules as the sender, thus constructing the second integer cumulative frequency table. Specifically, the receiver, based on the recovered token history, uses a conditional entropy model completely consistent with that of the sender (including the same LSTM structure, the same product-class distribution mixture model parameters, and the same deterministic initialization rules) to regenerate the joint conditional probability and the second integer cumulative frequency table for the current time step. The specific process is as follows: the multidimensional scalar quantized coordinates of the recovered tokens are converted into low-dimensional feature vectors through fully connected projection and layer normalization. These vectors are input into an LSTM to update the hidden state. The product-class distribution mixture model then outputs the conditional probability distribution for the current time step. Finally, the conditional probability is deterministically quantized into the second integer cumulative frequency table, and EOS is included in the second integer cumulative frequency table according to the same rules as the sender. Since the transmitter and receiver use the same model parameters, the same initialization rules (LSTM state reset to zero, BOS initial input, encoder preset initial interval), and the same probability quantization rules, the second integer cumulative frequency table reconstructed by the receiver at each time step is completely consistent with that of the transmitter, and there is no need to transmit the probability distribution or the first integer cumulative frequency table itself through the transmission link.

[0073] R3. The receiver performs entropy decoding on the compressed token payload according to the second integer cumulative frequency table, gradually recovering the current token and adding the recovered current token to the historical tokens, until the end-of-sequence marker is reached. Specifically, the receiver performs arithmetic decoding on the compressed token payload according to the current second integer cumulative frequency table, gradually recovering the original tokens until the end-of-sequence marker EOS is reached. Each recovered token is added to the recovered token history for use in conditional probability reconstruction at the next time step, forming a token-by-total autoregressive decoding loop. Under error-free transmission conditions, the token sequence recovered by the receiver is completely identical to the token sequence before compression at the sender.

[0074] R4. Input the recovered original speech token sequence into the detokenizer (speech token decoder) to obtain the reconstructed speech. Specifically, this step enables speech reconstruction. The receiver inputs the recovered original token sequence into the detokenizer, which converts the token sequence into a reconstructed speech waveform and outputs it. Since entropy compression is a lossless operation, the token sequence input into the detokenizer is exactly the same as the token sequence before compression at the sending end. Therefore, no additional loss in speech reconstruction quality is introduced by token compression itself.

[0075] In another implementation, the arithmetic coding described above is replaced with interval coding. Interval coding also performs lossless entropy coding on the current token based on the first integer cumulative frequency table. The update rules for the coding interval differ from those of arithmetic coding, but the conditional probability modeling process, the calculation method of the product category distribution hybrid model, and the construction process of the first integer cumulative frequency table are exactly the same as in the above implementation, achieving the same lossless compression function. Interval coding exhibits different encoding / decoding throughput characteristics on some hardware platforms, and the appropriate entropy coding method can be selected based on the computing resources and real-time requirements of the specific emergency communication terminal.

[0076] Specific embodiments of the present invention also provide a voice entropy coding system for multi-level emergency broadcasting and communication, such as... Figure 3 As shown, it includes: a sending-side processing module and a receiving-side processing module. The sending-side processing module and the receiving-side processing module are connected via a transmission link for exchanging compressed token payloads.

[0077] The transmitting side processing module 30 includes: a token generation module 301, used to convert the input speech into a discrete speech token sequence through a finite scalar quantization tokenizer, wherein each token in the speech token sequence is determined by a combination of quantization coordinates of multiple scalar quantization dimensions, and each scalar quantization dimension has multiple finite quantization levels;

[0078] The conditional probability modeling module 302 is connected to the Token generation module and is used to perform autoregressive modeling on historical Tokens, calculate the joint conditional probability of the current Token through a product category distribution mixture model, and predict the conditional probability distribution of the current Token.

[0079] Entropy encoding module 303, connected to the conditional probability modeling module, is used to quantize the conditional probability distribution into integer frequencies and include the sequence terminator as an additional encoding symbol in the first integer cumulative frequency table; and to perform entropy encoding on the current Token according to the first integer cumulative frequency table to generate the compressed Token payload;

[0080] The receiving side processing module 31 includes: a conditional probability reconstruction module 311, used to regenerate the joint conditional probability of the current token based on the recovered historical tokens, and construct a second integer cumulative frequency table containing the sequence terminator according to the same rules as the sending side.

[0081] Entropy decoding module 312, connected to the conditional probability reconstruction module, is used to perform entropy decoding on the compressed token payload according to the second integer cumulative frequency table, and gradually recover the original speech token sequence until the end of the sequence is recovered;

[0082] The speech reconstruction module 313 is connected to the entropy decoding module and is used to input the original token sequence into the detokenizer to obtain the reconstructed speech.

[0083] The transmitting-side processing module and the receiving-side processing module use the same conditional entropy model parameters, model initialization rules, probability quantization rules, integer cumulative frequency table construction rules, sequence termination rules, and entropy encoding and decoding rules.

[0084] Specifically, the overall architecture of the voice entropy coding system for multi-level emergency broadcasting and communication consists of three parts: a transmitter, a restricted / multi-standard transmission link, and a receiver. The transmitter includes an FSQ tokenizer and an autoregressive conditional entropy coding module, while the receiver includes an autoregressive conditional entropy decoding module and a detokenizer. The transmitter and receiver are connected via a restricted / multi-standard transmission link for exchanging compressed token payloads. The transmitter and receiver represent logical roles in the communication process and are not limited to specific physical devices. The transmitter can be deployed on an emergency broadcasting and communication dispatch platform, command terminal, mobile terminal, or other devices with voice communication capabilities; the receiver can also be a dispatch platform, mobile terminal, portable emergency communication terminal, or other voice receiving device. In different communication directions, the same device can act as both a transmitter and a receiver. The compressed token payload is output by the entropy coding module 303 and sent to the entropy decoding module 312 via the transmission link; the conditional probability is not transmitted via the link but is reconstructed by the receiver based on the recovered token.

[0085] Specifically, the ultra-low bitrate voice entropy coding system is integrated into multi-level emergency broadcasting and communication dispatch scenarios. In multi-level emergency broadcasting and communication dispatch applications, when the upper-level emergency broadcasting and communication dispatch platform issues voice dispatch instructions to the lower-level mobile terminals or emergency communication terminals, the upper-level platform acts as the sender and executes steps S1 to S4, while the lower-level terminal acts as the receiver and executes steps R1 to R4. When on-site rescue personnel report disaster information via mobile terminals to county-level, city-level, provincial-level, or other emergency broadcasting and communication dispatch platforms, the mobile terminal acts as the sender, and the dispatch platform acts as the receiver. This invention can be applied to voice communication between different levels, such as ministries, provinces, cities, counties, townships, and villages, and the encoding and decoding method itself does not depend on specific administrative levels. The same device can serve as both a sender and a receiver in different communication directions. This system can share communication resources with other services such as positioning, telemetry, image, text, and device status, reducing the bandwidth occupation of limited communication links by voice services by reducing the voice token transmission load.

[0086] In one embodiment of an electronic device, the electronic device includes a processor, a memory, and a communication interface, wherein the memory stores a computer program. The processor is communicatively connected to the memory, and when the processor executes the computer program stored in the memory, the electronic device performs the method described in any of the preceding embodiments.

[0087] This invention reduces the transmission load of voice tokens in emergency communication links by explicitly utilizing the multidimensional scalar quantization structure inside the FSQ Token, establishing the conditional probability distribution of the current Token using historical Tokens, and then combining entropy coding to perform post-processing lossless compression on the voice token sequence.

[0088] This invention eliminates the need to retrain the existing tokenizer and detokenizer, and can be used as an independent conditional entropy encoding / decoding module to connect to existing voice communication links. It is suitable for disaster voice reporting, emergency dispatch command issuance, and voice communication between multi-level emergency nodes under bandwidth-constrained conditions.

[0089] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A speech entropy coding method for multi-level emergency broadcasting and communication, characterized in that, The method includes: S1. Obtain the input speech to be transmitted, and encode the input speech into a discrete speech token sequence through a finite scalar quantization tokenizer. Each token in the speech token sequence is determined by a combination of quantization coordinates of multiple scalar quantization dimensions, and each scalar quantization dimension has multiple finite quantization levels. S2. Autoregressive modeling is performed based on the structured representation of the multidimensional scalar quantized coordinates of historical tokens. The joint conditional probability of the current token is calculated through a product category distribution mixture model, and the conditional probability distribution of the current token is predicted. S3. Quantize the conditional probability distribution into integer frequencies and construct a first integer cumulative frequency table; perform entropy encoding on the current Token according to the first integer cumulative frequency table; S4. After performing entropy encoding on the speech token sequence in sequence, output the compressed token payload; The compressed token payload is sent to the receiving end via a transmission link so that the receiving end can recover the original voice token sequence and reconstruct the voice using a detokenizer.

2. The method according to claim 1, characterized in that, A one-way long short-term memory network is used to establish the time dependency relationship between historical tokens. The structured input representation, hidden state and cell state corresponding to the previous token are used as inputs, and the hidden state and cell state of the current token are output.

3. The method according to claim 1, characterized in that, The product category distribution hybrid model outputs multiple hybrid components based on the current context representation obtained by the autoregressive modeling. Each quantized coordinate in each hybrid component is modeled using a category distribution. The joint probability of the hybrid component is calculated by the product of the category probabilities of each quantized coordinate. Then, the joint probability of each hybrid component is weighted and summed to obtain the joint conditional probability of the current token.

4. The method according to claim 1, characterized in that, The multidimensional scalar quantized coordinates of the previous token are converted into low-dimensional feature vectors after fully connected projection and layer normalization, which serve as the input representation of the temporal probability model.

5. The method according to claim 1, characterized in that, In the first integer cumulative frequency table, the integer frequency of each Token is not less than 1, so that each Token corresponds to a valid entropy coding interval.

6. The method according to claim 1, characterized in that, The encoder recursively updates the entropy encoding interval according to the cumulative frequency interval corresponding to the Token, completes entropy encoding for the entire discrete speech Token sequence, and encodes the sequence end symbol as an additional entropy encoding symbol into the first integer cumulative frequency table to form the compressed Token payload.

7. The method according to claim 2, characterized in that, Before each speech encoding begins, the encoder is restored to the preset initial interval, the hidden state and cell state of the temporal probability model are initialized to zero, and the sequence start symbol is used as the initial input of the temporal probability model; the encoder state and temporal probability model state are not inherited between different speech.

8. The method according to claim 1, characterized in that, The conditional entropy model is trained using a teacher-forced method, with the training objective being to minimize the negative log-likelihood of the speech token sequence; the parameters of the finite scalar quantizer are kept frozen during training.

9. A voice entropy decoding method for multi-level emergency broadcasting and communication, characterized in that, The method includes: R1 receives the compressed token payload output by the sender; R2. Based on the structured representation of the multidimensional scalar quantized coordinates of the recovered historical tokens, the joint conditional probability of the current token is regenerated using the conditional entropy model consistent with the sending end, and the joint conditional probability is quantized into integer frequencies. The sequence terminator is included in the second integer cumulative frequency table according to the rules consistent with the sending end, and the second integer cumulative frequency table is constructed. R3. Perform entropy decoding on the compressed Token payload according to the second integer cumulative frequency table, gradually recover the current Token, and add the recovered current Token to the historical Token until the end of the recovery sequence is recovered; R4. Input the recovered original speech token sequence into the detokenizer to obtain the reconstructed speech.

10. A voice entropy coding system for multi-level emergency broadcasting and communication, characterized in that, include: The sending-side processing module and the receiving-side processing module are connected via a transmission link to exchange compressed token payloads. The transmitting side processing module includes: a token generation module, used to convert the input speech into a discrete speech token sequence through a finite scalar quantization tokenizer, wherein each token in the speech token sequence is determined by a combination of quantization coordinates of multiple scalar quantization dimensions, and each scalar quantization dimension has multiple finite quantization levels; The conditional probability modeling module, connected to the token generation module, is used to perform autoregressive modeling based on the structured representation of the multidimensional scalar quantized coordinates of historical tokens, calculate the joint conditional probability of the current token through a product category distribution mixture model, and predict the conditional probability distribution of the current token. An entropy encoding module, connected to the conditional probability modeling module, is used to quantize the conditional probability distribution into integer frequencies and include the sequence terminator as an additional encoding symbol in a first integer cumulative frequency table; and to perform entropy encoding on the current Token according to the first integer cumulative frequency table to generate the compressed Token payload; The receiving-side processing module includes: a conditional probability reconstruction module, used to regenerate the joint conditional probability of the current token based on the recovered historical tokens, and construct a second integer cumulative frequency table containing sequence terminators according to the same rules as the sending side; The entropy decoding module, connected to the conditional probability reconstruction module, is used to perform entropy decoding on the compressed token payload according to the second integer cumulative frequency table, and gradually recover the original speech token sequence until the end of the sequence is recovered. The speech reconstruction module, connected to the entropy decoding module, is used to input the original token sequence into the detokenizer to obtain the reconstructed speech; The transmitting-side processing module and the receiving-side processing module use the same conditional entropy model parameters, model initialization rules, probability quantization rules, integer cumulative frequency table construction rules, sequence termination rules, and entropy encoding and decoding rules.

Images