Streaming voice content binding anti-downgrade verification method and system

Abstract

This invention relates to the field of audio signal processing technology and discloses a method and system for preventing degradation verification of streaming speech content binding. The method divides the streaming speech to be verified into multiple current audio segments, recovers the source token containing record identifiers and verification policy commitment information through in-band audio watermark detection, and after obtaining candidate external content records, first performs record security verification and compatibility verification. If security verification fails, the degradation verification path is directly blocked. If both verifications pass, the main verification is entered, generating segment-level content differences and cross-segment sequence integrity features, and outputting a structured verification result after comprehensive judgment. If no record is obtained and there is no historical security failure state, degradation permission determination is completed based on policy commitment information, and controlled execution of self-contained degradation verification is performed, outputting a result with a degradation mark. This invention can effectively identify content tampering and policy replacement attacks, prevent security failures from being overwritten by degradation, and balance limited degradation verification capabilities with verification security.

Description

Technical Field

[0001] This invention relates to the field of audio signal processing technology, and more specifically, to a method and system for preventing degradation of streaming voice content binding. Background Technology

[0002] With the development of speech synthesis, speech cloning, speech editing, and streaming media distribution technologies, licensed speech may undergo processing such as trimming, transcoding, loudness normalization, noise addition, segment replacement, silence insertion, splicing, segment rearrangement, or segment repetition during dissemination. Verification requires not only determining whether a speech segment carries a specific source identifier, but also whether the local acoustic content of the speech remains consistent with its original release content.

[0003] Existing audio watermarking schemes typically embed identifiers or short messages within audio for copyright marking, source identification, or local detection. However, in streaming speech scenarios, the in-band watermark payload that a single audio segment can carry is usually small. If only source tokens or object identifiers are embedded in the audio, even if the source token is correctly recovered, it cannot definitively prove that the speech content has not been partially replaced, spliced, or rearranged.

[0004] Existing content verification schemes based on audio fingerprinting, perceptual hashing, external lists, digital signatures, or sequence commitments can compare the current speech with reference features. However, if the reference features, quantization rules, thresholds, model versions, or state priorities are not securely bound to the source token and external content records, they may be affected by record replacement, policy replacement, interpretation rule confusion, or false passage with low evidence.

[0005] In some attack scenarios, attackers do not directly forge fragment-level fingerprints, but instead induce the verification system to use different feature versions, quantization rules, matching thresholds, discrimination parameter versions, or downgraded verification paths. If the system still outputs a state of consistency between source and content when recording security failures, incompatible records, or insufficient evidence, or uses weaker self-contained downgraded verification to cover external content recording security failures, it will produce erroneous deterministic outputs. Therefore, there is an urgent need for a streaming voice content binding anti-downgrade verification method and system to solve the above problems. Summary of the Invention

[0006] To address the aforementioned technical issues, this invention provides a streaming voice content binding anti-downgrade verification method and system. Through multi-feature binding, full-parameter signature protection, and state priority mechanism, it identifies content tampering and policy replacement attacks, preventing security failures from being overwritten by downgrades. Controlled downgrades and delayed micro-fingerprints provide limited verification, reducing the risk of misjudgments and attacks.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] A method for preventing degradation of streaming voice content binding includes:

[0009] Obtain the streaming audio to be verified and divide it into multiple current audio segments;

[0010] Perform in-band audio watermark detection on each current audio segment, determine the target synchronization offset and recover the source token, wherein the source token includes a record identifier or equivalent index, token binding information and verification strategy commitment information;

[0011] Based on the record identifier or equivalent index, attempt to obtain candidate external content records corresponding to the multiple current audio segments; if obtained, perform record security verification and record compatibility verification on the candidate external content records; if the record security verification fails, output a record exception and block the downgrade verification path; if the record security verification passes but the record compatibility verification fails, output a record incompatibility or insufficient evidence abstention status; if both verifications pass, proceed to the main verification.

[0012] In the main verification, the segment-level comparable acoustic content fingerprint is recalculated for each current audio segment to generate segment-level content difference features and cross-segment sequence integrity features. Based on the segment-level content difference features, the cross-segment sequence integrity features, and the recorded verification status, a comprehensive verification is performed by an anomaly decision-maker to output a structured verification result.

[0013] If no candidate external content record is obtained and there is no state where a candidate external content record has been obtained but the record security verification has failed, a downgrade permission determination is made based on the verification policy commitment information; if the determination meets the downgrade conditions, a self-contained downgrade verification is performed, and a verification result marked as downgraded is output. The verification result is not considered as a state of consistency between the source and content of the external content record; if the determination does not meet the downgrade conditions, a state of insufficient evidence and abstention or a state of downgrade not allowed is output.

[0014] As a preferred embodiment of the present invention, determining the target synchronization offset includes: selecting multiple probe windows from the current audio segment, generating multiple candidate synchronization offsets based on the probe windows; traversing each candidate synchronization offset to perform in-band audio watermark detection, and calculating the synchronization score corresponding to each candidate synchronization offset based on the watermark detection confidence, payload bit error rate estimation, error correction decoding status, payload fragment consistency, authentication verification consistency, and candidate offset stability; selecting a candidate synchronization offset that meets both the synchronization score threshold and the candidate offset stability condition as the target synchronization offset; and outputting an insufficient evidence abstention status when no candidate synchronization offset meets the conditions.

[0015] The process of recovering the source token includes: extracting candidate payload fragments under the target synchronization offset, determining whether the number of valid fragments meets the minimum fragment number requirement; if the minimum fragment number requirement is not met, rejecting the source token and outputting a token recovery failure status or an insufficient evidence abandonment status; if the minimum fragment number requirement is met, reassembling the candidate payload based on the fragment sequence number, and sequentially performing error correction decoding, integrity verification, and authentication verification on the candidate payload; if all three verifications pass, outputting the source token and the source token recovery confidence level; if any one verification fails, rejecting the source token and outputting a token recovery failure status or an insufficient evidence abandonment status.

[0016] The verification policy commitment information in the source token is a normalized encoded digest of the verification policy object.

[0017] As a preferred embodiment of the present invention, the candidate external content record includes a record identifier, a token binding field, fragment index information, fragment-level comparable acoustic content fingerprint or comparable fingerprint digest, adjacent fragment acoustic transfer digest, sequence commitment, feature version, quantization rule, verification strategy object, verification strategy commitment digest, discrimination parameter version, evidence threshold field, status priority table, downgrade permission field, record version, validity period field, revocation field, and signature or message authentication code; the signature or message authentication code covers the other fields in the candidate external content record.

[0018] As a preferred embodiment of the present invention, the record compatibility verification includes the processability verification of the record version, feature version, quantization rule and discrimination parameter version; if the record version, feature version, quantization rule or discrimination parameter version is not supported by the verification end, the record incompatibility status or the evidence insufficient abandonment status is output, and the source and content consistency status is not output.

[0019] The record security verification includes signature or message authentication code verification, token binding relationship verification, verification policy commitment relationship verification, validity period verification, and revocation status verification; wherein:

[0020] The verification strategy commitment relationship verification includes: reading the verification strategy object from the candidate external content record, encoding the verification strategy object according to a preset normalized encoding method, and recalculating the verification strategy commitment digest based on the encoding result; comparing the recalculated verification strategy commitment digest with the verification strategy commitment information in the source token and the verification strategy commitment digest in the candidate external content record that is covered by the signature or message authentication code; if the three are inconsistent, the candidate external content record is determined to have failed record security verification; if the comparison cannot be completed because the normalized encoding method is not supported or the normalized encoding result cannot be uniquely determined, a record incompatibility status or insufficient evidence abstention status is output.

[0021] If the signature or message authentication code verification, the token binding relationship verification, the verification policy commitment relationship verification, the validity period verification, and the revocation status verification all pass, the candidate external content record is determined to have passed the record security verification; if any one of the verifications fails, the candidate external content record is determined to have failed the record security verification, an abnormal record status is output, and the execution of a self-contained degradation verification path that can output a status of consistency between the source and content or a degradation pass status is prohibited.

[0022] As a preferred embodiment of the present invention, the signature or message authentication code verification includes: verifying the signature or message authentication code in the candidate external content record; if the signature or message authentication code verification fails, the candidate external content record is determined to be a record security verification failure.

[0023] The token binding relationship verification includes: comparing the token binding information in the source token with the token binding field in the candidate external content record; if the two do not satisfy the predetermined binding relationship, the candidate external content record is determined to be a record security verification failure.

[0024] The validity period verification includes: determining whether the candidate external content record has expired based on the validity period field in the candidate external content record; if the candidate external content record has expired, the candidate external content record is determined to be a record security verification failure.

[0025] The revocation status verification includes: determining whether the candidate external content record has an abnormal revocation status based on the revocation field in the candidate external content record; if an abnormal revocation status is determined, the candidate external content record is determined to have failed the record security verification.

[0026] As a preferred embodiment of the present invention, the generated fragment-level content difference features include:

[0027] Based solely on the feature version covered by the signature or message authentication code in the candidate external content record and the quantization rule, recalculate the segment-level comparable acoustic content fingerprint or comparable fingerprint digest for each current audio segment;

[0028] The recalculated fragment-level comparable acoustic content fingerprint or comparable fingerprint summary is compared with the fragment-level comparable acoustic content fingerprint or comparable fingerprint summary in the candidate external content record to generate fragment-level content difference features.

[0029] As a preferred embodiment of the present invention, the generation of cross-segment sequence integrity features includes:

[0030] For each current audio segment, the recalculated segment-level comparable acoustic content fingerprint or comparable fingerprint summary is compared with the segment-level comparable acoustic content fingerprints or comparable fingerprint summaries of multiple candidate record segments in the candidate external content records, and the corresponding fingerprint distance is calculated; candidate record segments whose fingerprint distance meets the matching threshold are determined as candidate matching segments for the corresponding current audio segment; if no candidate record segment meets the matching threshold, the corresponding current audio segment is marked as an unmatchable segment;

[0031] Based on the candidate matching segments, corresponding fingerprint distances, and unmatched markers for each current audio segment, a matching index sequence is determined through dynamic programming with index monotonicity constraints. The path cost of the dynamic programming includes at least fingerprint distance cost, matching index jump cost, repeated matching index cost, index reversal cost, and unmatched cost, and is corrected based on the comparison results between the acoustic transfer summaries between adjacent current audio segments and the acoustic transfer summaries between adjacent segments of candidate record segments. The weights and matching thresholds used in the dynamic programming are determined by the verification strategy object or the discriminant parameter version.

[0032] The matching index sequence is arranged according to the time order of the current audio segment, and the continuous matching, skipping, repetition, reverse order and unmatched cases of candidate record segments are identified according to the matching index sequence; for adjacent matched current audio segments, their acoustic transfer summaries are compared with the acoustic transfer summaries of adjacent segments of the corresponding candidate record segments to obtain the acoustic transfer summaries comparison results of adjacent segments;

[0033] The candidate record segment sequence determined according to the matching index sequence is compared with the sequence commitment covered by the signature or message authentication code in the candidate external content record to obtain the sequence commitment verification result; the sequence commitment is generated jointly based on the ordered segment index information, the segment-level comparable acoustic content fingerprint or comparable fingerprint digest, and the acoustic transfer digest of adjacent segments.

[0034] The cross-segment sequence integrity feature is generated based on the matching index sequence, the acoustic transfer summary alignment results of adjacent segments, and the sequence commitment verification results.

[0035] When the matching index sequence indicates that the candidate record segment has skipped, repeated, or reversed sequences that exceed the allowed range of the verification strategy object, the acoustic transfer summaries of adjacent segments are inconsistent, or the candidate record segment sequence determined according to the matching index sequence cannot meet the sequence commitment, a sequence integrity feature is generated to characterize at least one sequence abnormality among segment deletion, insertion, repetition, rearrangement, or splicing; when the number of current audio segments that can participate in the matching is insufficient to complete the sequence integrity judgment, a sequence integrity feature is generated to characterize insufficient sequence verification evidence.

[0036] As a preferred embodiment of the present invention, the threshold or model parameters of the anomaly decision device correspond to the version of the discrimination parameters, and all thresholds or model parameters are determined before the start of the verification phase, and no adjustment is made for a single stream of speech to be verified during the verification process;

[0037] The anomaly decision-maker takes the fragment-level content difference features, the cross-fragment sequence integrity features, the source token recovery confidence, the synchronization score or synchronization confidence, the number of valid fragments, and the record verification status as input, and combines the evidence threshold field and the status priority table to output the structured verification results.

[0038] As a preferred embodiment of the present invention, the degradation permission determination includes three degradation conditions: 1) no candidate external content record available for primary verification has been obtained; 2) no candidate external content record has been obtained but the record security verification has failed; and 3) corresponding degradation permission information is obtained from the local trusted policy set according to the verification policy commitment information in the source token, and the degradation permission information indicates that degradation verification is allowed. If all three degradation conditions are met simultaneously, the degradation condition is deemed to be met.

[0039] The self-contained degradation verification includes: recovering the microfingerprint tag of the corresponding audio segment from the subsequent current audio segment of the streaming speech to be verified, recalculating the current microfingerprint of the corresponding current audio segment, and comparing the current microfingerprint with the recovered microfingerprint tag; if the recovered evidence length is insufficient, the tag recovery confidence is lower than a preset threshold, the microfingerprint error rate is higher than a preset threshold, or the number of delayed segments is insufficient, then an insufficient evidence abstention status is output; if the comparison is inconsistent and the evidence is sufficient, then a local anomaly prompt with a degradation status mark is output; if the comparison is consistent and the evidence is sufficient, then a degradation pass result with a degradation status mark is output.

[0040] The verification results marked as degraded must not overwrite the abnormal status of the record corresponding to the security verification failure.

[0041] This invention also provides a streaming voice content binding anti-downgrade verification system, comprising:

[0042] The segmentation module is used to acquire the streaming speech to be verified and divide it into multiple current audio segments;

[0043] The token recovery module is used to perform in-band audio watermark detection on each current audio segment, determine the target synchronization offset, and recover the source token. The source token includes a record identifier or equivalent index, token binding information, and verification strategy commitment information.

[0044] The recording verification module is used to attempt to obtain candidate external content records corresponding to the multiple current audio segments based on the recording identifier or equivalent index; if obtained, the candidate external content records are subjected to recording security verification and recording compatibility verification; if the recording security verification fails, a recording exception is output and the downgrade verification path is blocked; if the recording security verification passes but the recording compatibility verification fails, a recording incompatibility or insufficient evidence abstention status is output; if both verifications pass, the main verification module is triggered.

[0045] The main verification module is used to recalculate the segment-level comparable acoustic content fingerprint for each current audio segment according to the verification parameters protected by the signature in the candidate external content record, and generate segment-level content difference features and cross-segment sequence integrity features; based on the segment-level content difference features, the cross-segment sequence integrity features and the recorded verification status, it performs comprehensive verification through an anomaly decision-maker and outputs a structured verification result.

[0046] The downgrade verification module is used to determine downgrade permissions based on the verification policy commitment information when no candidate external content record is obtained or when there is no candidate external content record that has been obtained but the record security verification has failed. If the downgrade conditions are met, a self-contained downgrade verification is performed, and a verification result marked as downgraded is output. The verification result is not considered as a state of consistency between the source and content of the external content record. If the downgrade conditions are not met, a state of insufficient evidence and abstention or downgrade not allowed is output.

[0047] The beneficial effects of this invention are:

[0048] This invention binds a source token, an external content record, a fragment-level comparable acoustic content fingerprint or comparable fingerprint digest, an adjacent fragment acoustic transfer digest, a matching index sequence, a sequence commitment, and a verification strategy object. This enables the verification end to identify content anomalies such as local replacement, splicing, fragment rearrangement, fragment duplication, or fragment missing based on fragment-level content difference features and sequence integrity features, even when the source token can be correctly recovered. This reduces the risk of outputting a source and content consistency status simply because the source token is correct.

[0049] This invention stores the verification policy object or verification policy commitment digest in the external content record and enables the source token to carry the verification policy commitment information. The verification end recalculates the verification policy commitment digest through normalized encoding and performs a consistency comparison with the verification policy commitment information in the source token and the verification policy commitment digest in the external content record that is covered by the signature or message authentication code. This enables the detection of anomalies such as the replacement of the verification policy object, inconsistent policy commitment relationships, or uncertain policy encoding, reducing the risk of verification policy replacement attacks.

[0050] This invention ensures that the signature or message authentication code covers at least the fragment-level comparable acoustic content fingerprint or comparable fingerprint digest, adjacent fragment acoustic transfer digest, sequence commitment, verification strategy object or verification strategy commitment digest, feature version, quantization rule, discriminant parameter version, evidence threshold field, state priority table, and downgrade permission field. This ensures that the verification end can only use the protected interpretation rules and discriminant parameters for verification, preventing attackers from changing the verification results by replacing quantization rules, distance thresholds, dynamic programming weights, model parameters, or state priorities. This reduces the risk of misjudgment caused by interpretation rule obfuscation, threshold replacement, and model version replacement.

[0051] This invention outputs an abnormal recording status when security verification fails and prohibits entry into self-contained degradation verification paths that could output a source-content consistency status or a degradation pass status. When there is record incompatibility, insufficient evidence to recover the source token, insufficient synchronization scoring, or insufficient number of valid fragments, it outputs a record incompatibility status or an insufficient evidence abstention status, and does not output a source-content consistency status. Therefore, it avoids recording security verification failures being overwritten by low-intensity degradation verification results, reducing the risk of erroneous deterministic output under low-evidence conditions.

[0052] When no candidate external content record is available and there is no record of security verification failure, this invention does not directly downgrade to pass. Instead, it obtains the corresponding downgrade permission information from the local trusted policy set based on the verification policy commitment information in the source token, and determines whether to allow entry into the self-contained downgrade verification path based on the downgrade permission information. The self-contained downgrade verification output carries a downgrade status flag and is not considered as a source-content consistency status based on the external content record. Thus, it can provide controlled and limited verification capabilities when external content records are temporarily unavailable, while avoiding unconditional downgrade to pass.

[0053] This invention, during the registration phase, embeds the micro-fingerprint tag of the i-th source audio segment into the (i+1)th to (i+k)th subsequent source audio segments after fragmentation. When the verification end does not obtain external content records, it can recover the micro-fingerprint tag from the subsequent current audio segments and compare it with the recalculated current micro-fingerprint. Therefore, under the condition of low-capacity in-band watermarking where a single audio segment can only carry a short payload, limited content consistency verification can be achieved without relying on external content records. Furthermore, it avoids erroneous passes under low-evidence conditions by judging evidence length, tag recovery confidence, micro-fingerprint error rate, and the number of delayed segments. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the overall system architecture of the present invention.

[0055] Figure 2 This is a schematic diagram of the registration and watermark embedding process.

[0056] Figure 3 This is a schematic diagram of the main verification process based on external content records.

[0057] Figure 4 This is a schematic diagram of the protected verification strategy object and the external content record structure.

[0058] Figure 5 This is a schematic diagram of a state machine designed to prevent degradation.

[0059] Figure 6 This is a schematic diagram of the probe-based synchronization process.

[0060] Figure 7 This is a schematic diagram of fragment-level fingerprint and sequence integrity feature generation.

[0061] Figure 8 This is a schematic diagram of an anomaly decision device with fixed discrimination parameters before verification.

[0062] Figure 9 This is a schematic diagram of a delayed microfingerprint tag with self-contained downgrade verification.

[0063] Figure 10 This is a schematic diagram of the protected verification strategy and the anti-degradation verification closed loop. Detailed Implementation

[0064] In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0065] Combination Figures 1-10 The present invention provides the following embodiments:

[0066] Example 1:

[0067] This invention provides a method for preventing downgrades in streaming voice content binding, comprising the following steps.

[0068] Step 1: Segmentation of the streaming speech segment to be verified

[0069] In this embodiment, the streaming audio to be verified is acquired, and the streaming audio to be verified is divided into multiple current audio segments according to the preset segment duration and preset overlap ratio consistent with the registration stage.

[0070] Step 2: In-band audio watermark detection to determine target synchronization offset

[0071] Each current audio segment undergoes in-band audio watermark detection, and a probe-based synchronization scheme is used to determine the target synchronization offset, specifically including:

[0072] (1) Select from multiple current audio segments A probe window is provided, and multiple candidate synchronization offsets are generated based on the probe window;

[0073] Optionally, the number of probe windows The number of probe windows ranges from 4 to 64, with 8 to 16 being preferred. The probe windows can be evenly distributed across the streaming speech to be verified, or they can be adaptively selected based on the speech activity detection results, segment energy, watermark detection confidence, or the temporal coverage position of the probe windows in the streaming speech to be verified.

[0074] The step size for coarse-grained candidate offsets is, for example, 50 milliseconds, and the step size for subsequent fine-grained searches on candidate offsets is, for example, 5 milliseconds.

[0075] (2) For each candidate synchronization offset Calculate the corresponding synchronization score using the following formula:

[0076]

[0077] in, The confidence level for watermark detection. For payload bit error rate estimation, In error correction decoding state, To ensure load fragmentation consistency, To ensure consistency in authentication verification, The instability of candidate offsets corresponding to multiple probe windows. These are the weighting coefficients for the corresponding items. The value range is 0.10 to 0.25. The value range is 0.10 to 0.25. The value range is 0.15 to 0.30. The value range is 0.10 to 0.25. The value range is 0.20 to 0.40. The value range is 0.05 to 0.20.

[0078] Candidate offset instability and related stability indices are calculated as follows:

[0079] ① Solving for the local optimal offset: Let the period of the audio segment be... For candidate synchronization offset ,exist Each probe window contains... Perform a local fine-grained search around the center to obtain the first... Local optimal offset of each probe window Due to the synchronization offset For the period, offset 0 and close Since the offsets are adjacent in physical location, the cyclic distance is used to calculate the offset deviation. The cyclic distance between any two offsets a and b is defined as: .

[0080] ② Determining the offset: Based on the local optimal offset of all probe windows Calculate the cyclic midpoint offset Alternatively, the offset that minimizes the sum of the loop distances of all windows can be used as the representative offset. The deviation of the p-th probe window is: .

[0081] ③Support window statistics: Set candidate offset tolerance , will satisfy The probe window is denoted as the support window, and the number of support windows is counted. With support rate : ;in This is an indicator function; it takes the value 1 when the condition is true and 0 when the condition is false.

[0082] ④ Instability quantification: Calculate the instability of candidate offsets This characterizes the degree of dispersion of the local optimal offset of each probe window around the representative offset. The optional normalization calculation method is: ;in The scale is used to normalize the dispersion. The closer to 0, the more concentrated the offset is under multi-window conditions, and the stronger the stability of the candidate synchronization offset; the closer to 1, the more dispersed the offset is, and the candidate offset is likely caused by noise, local pseudo-peaks, or audio splicing boundaries.

[0083] (3) Target synchronization offset selection

[0084] Get synchronized rating Not lower than the synchronous scoring threshold Candidate synchronization offsets that meet the candidate offset stability conditions are selected as target synchronization offsets. The synchronization score and stability must be met simultaneously, and the synchronization score should not be the sole criterion for selection.

[0085] Explanation of Synchronous Scoring Threshold: Synchronous Scoring Threshold This is a discriminant parameter used to distinguish between sufficient and insufficient synchronous evidence. Its value corresponds to the weight configuration, scoring item normalization method, and watermark detection model version. It is registered as a fixed parameter in the verification strategy object and belongs to the fixed discriminant parameter corresponding to the evidence threshold field or the discriminant parameter version. In an implementation where both the scoring item and weight are normalized to the [0,1] interval and the weight sum is 1, The value can be between 0.60 and 0.85, for example, 0.70; the specific value can be determined based on the receiver operating characteristic curve on the independently calibrated sample set, and the false acceptance rate should be controlled to not exceed the preset upper limit.

[0086] Stability determination criteria include:

[0087] Supported window number meets ,in This is the preset minimum number of supported windows;

[0088] Support rate meets ,in This is the preset minimum support rate;

[0089] Candidate offset instability satisfies ,in This is the preset threshold for candidate offset instability.

[0090] Optionally, an additional peak interval discrimination condition can be added: requiring the difference between the highest and second-highest synchronization scores among all candidate offsets to satisfy... This is to eliminate multi-peak ambiguous scenarios where multiple candidate offset scores are close and cannot be reliably distinguished. This represents the minimum interval between the highest and second-highest scores, ranging from 0.02 to 0.10.

[0091] When there are multiple candidate synchronization offsets that meet all verification conditions, the final target synchronization offset is selected according to priority rules: the offset with the higher synchronization score is selected first; if the synchronization scores are the same, the offset with the lower instability of the candidate offset is selected first; if they are still the same, the offset with the higher support rate is selected first; or predetermined rules such as smaller offset values ​​can be used to achieve deterministic disambiguation.

[0092] (4) Abnormal handling If there is no candidate synchronization offset that meets all the conditions, it is determined that the synchronization evidence is insufficient, the insufficient evidence abstention status is output, and the source token recovery process is terminated to avoid outputting incorrect verification results when the synchronization is unreliable.

[0093] Step 3: Restore the source token and restore the source token confidence level

[0094] Extract candidate payload fragments under target synchronization offset and determine whether the number of valid fragments meets the minimum fragment number requirement. If the minimum fragment number requirement is not met, reject the source token and output a token recovery failure status or an insufficient evidence abstention status. If the minimum fragment number requirement is met, reassemble the candidate payload based on the fragment sequence number and perform error correction decoding, integrity verification, and authentication verification on the reassembled candidate payload in sequence. If the error correction decoding, integrity verification, and authentication verification all pass, output the source token and the source token recovery confidence. If any one of the verifications fails, reject the source token and output a token recovery failure status or an insufficient evidence abstention status.

[0095] In this embodiment, the source token includes at least a record identifier for obtaining external content records or an equivalent index that can index external content records, token binding information, and verification policy commitment information; wherein, the verification policy commitment information is a normalized encoded digest or a digest truncation value of the verification policy object.

[0096] Optionally, each audio segment carrying the watermark payload carries a watermark payload of 4 bits to 64 bits, such as 16 bits.

[0097] Optionally, error correction decoding can employ at least one of Reed-Solomon codes, BCH codes, LDPC codes, convolutional codes, and parity check codes; integrity verification and authentication verification can employ at least one of cyclic redundancy check, message authentication code truncation value, digital signature truncation value, or their verification digest; the in-band audio watermark channel can be implemented by at least one of spread spectrum watermarking, quantization index modulation watermarking, phase modulation watermarking, and neural audio watermarking. This embodiment does not limit the underlying watermarking method.

[0098] Step 4: Attempt to retrieve candidate external content records

[0099] Based on the record identifier or equivalent index in the source token, attempt to retrieve candidate external content records corresponding to multiple current audio segments.

[0100] In this embodiment, the candidate external content record includes at least a record identifier, a token binding field, fragment index information, a fragment-level comparable acoustic content fingerprint or comparable fingerprint digest, an adjacent fragment acoustic transfer digest, a sequence commitment, a feature version, a quantization rule, a verification strategy object, a verification strategy commitment digest, a discrimination parameter version, an evidence threshold field, a status priority table, a downgrade permission field, a record version, a validity period field, a revocation field, and a signature or message authentication code; wherein, the signature or message authentication code at least covers the other fields in the candidate external content record.

[0101] Table 1 shows the fields of an external content record in one embodiment. This table is for illustrative purposes only and does not limit the field order or encoding method.

[0102]

[0103] Optionally, external content records may not be in the form of a single file, but may be in the form of database rows, key-value records, signature objects, transparent log entries, or distributed hash table entries; as long as the record is bound to the source token and the signature or message authentication code covers the fields related to the verification strategy, it belongs to the candidate external content record referred to in this embodiment.

[0104] Step 5: Record security verification and record compatibility verification

[0105] Upon obtaining candidate external content records, perform record security verification and record compatibility verification on the candidate external content records to determine the record verification status.

[0106] (1) Record security verification

[0107] Record security verification includes signature or message authentication code verification, token binding relationship verification, verification policy commitment relationship verification, validity period verification, and revocation status verification; if any of the above verifications fails, the record security verification is deemed to have failed.

[0108] Specifically, the verification of the verification strategy commitment relationship includes: reading the verification strategy object from the candidate external content record, encoding the verification strategy object according to the preset normalized encoding method, and recalculating the verification strategy commitment digest based on the encoding result; comparing the recalculated verification strategy commitment digest with the verification strategy commitment information in the source token and the verification strategy commitment digest covered by the signature or message authentication code in the candidate external content record; if the three are inconsistent, the candidate external content record is determined to have failed record security verification; if the comparison cannot be completed because the normalized encoding method is not supported or the normalized encoding result cannot be uniquely determined, the record incompatibility status or insufficient evidence abstention status is output.

[0109] The verification strategy commitment summary is calculated using the following formula:

[0110]

[0111] in, For hash functions, For deterministic normalized coding, To verify the policy object; normalized encoding is used to specify the field order, data type, numerical precision, string encoding, and default value handling method to ensure that the same verification policy object receives a consistent verification policy commitment summary on both the registration and verification ends.

[0112] (2) Record compatibility verification

[0113] Record compatibility verification includes the processing compatibility verification of record version, feature version, quantization rule and discrimination parameter version; if the record version, feature version, quantization rule or discrimination parameter version is not supported by the verification end, the record incompatibility status or insufficient evidence abandonment status will be output, and the source and content consistency status will not be output.

[0114] Step 6: Degradation Prevention Branch Processing

[0115] Branch control is executed according to state priority, with the highest priority being the recording of security verification failure states, specifically including:

[0116] (1) If the security verification fails, output the record exception and block the downgrade verification path, prohibiting the execution of the self-contained downgrade verification path that can output the source and content consistent status or the downgrade passed status.

[0117] (2) If the security verification of the record passes but the record compatibility verification fails, output the record incompatibility or insufficient evidence abstention status;

[0118] (3) If both the security verification and the compatibility verification of the record are passed, then proceed to the main verification, that is, execute steps 7 to 9 below;

[0119] (4) If no candidate external content record is obtained, and there is no state where a candidate external content record has been obtained but the record security verification has failed, then the downgrade permission determination shall be made according to the verification policy commitment information in the source token.

[0120] (5) If the downgrade authority determines that the downgrade conditions are not met, output a status of insufficient evidence and abstention or a status of downgrade not allowed.

[0121] Step 7: Recalculate fingerprints and generate fragment-level content difference features.

[0122] In the main verification, the segment-level comparable acoustic content fingerprint or comparable fingerprint summary is recalculated for each current audio segment only according to the verification strategy corresponding to the feature version, quantization rule and discrimination parameter version covered by the signature or message authentication code in the candidate external content record. The recalculated fingerprint is then compared with the segment-level comparable acoustic content fingerprint or comparable fingerprint summary in the candidate external content record to generate segment-level content difference features.

[0123] It should be noted that feature parameters or quantization rules that are not covered by the signature or message authentication code must not be used during the verification process.

[0124] Optionally, the preprocessing before recalculating the fingerprint includes at least one of sampling rate unification, channel merging, amplitude normalization, loudness normalization, speech activity detection, silence processing, and time-frequency transformation; the segment-level comparable acoustic content fingerprint can be generated by one or more of log-Mel spectrum features, log-Mel spectrum features compressed by discrete cosine transform, spectral dynamic features, overlapping window features, energy transfer features, and neuroacoustic embedding; the fingerprint distance type can be Euclidean distance, cosine distance, Hamming distance, correlation distance, or dynamic time-warped distance, with the specific type specified by the verification strategy object.

[0125] Step 8: Generate sequence integrity features across segments

[0126] Based on the matching index sequence, the acoustic transfer summary comparison results of adjacent segments, and the sequence commitment verification results, cross-segment sequence integrity features are generated, as specifically implemented below.

[0127] (1) Generation of matching index sequences

[0128] For each current audio segment, the recalculated segment-level comparable acoustic content fingerprint or comparable fingerprint summary is compared with the segment-level comparable acoustic content fingerprints or comparable fingerprint summaries of multiple candidate record segments in the candidate external content records, and the corresponding fingerprint distance is calculated. Candidate record segments whose fingerprint distance meets the matching threshold are determined as candidate matching segments of the current audio segment. If there are no candidate record segments that meet the matching threshold, the current audio segment is marked as an unmatchable segment.

[0129] Furthermore, based on the candidate matching segments, corresponding fingerprint distances, and unmatched markers for each current audio segment, a matching index sequence is determined through dynamic programming with index monotonicity constraints. The path cost of the dynamic programming includes at least fingerprint distance cost, matching index jump cost, repeated matching index cost, index reversal cost, and unmatched cost. The path cost is adjusted based on the comparison between the acoustic transfer summaries of adjacent current audio segments and the acoustic transfer summaries of adjacent segments between candidate record segments. The weights and matching thresholds used in the dynamic programming are determined by the verification strategy object or the discriminant parameter version.

[0130] The matching index sequence is arranged according to the time order of the current audio segment, and the continuous matching, skipping, repetition, reversal and unmatchable cases of candidate record segments are identified according to the matching index sequence; for adjacent matched current audio segments, their acoustic transfer summaries are compared with the acoustic transfer summaries of adjacent segments of the corresponding candidate record segments to obtain the acoustic transfer summaries comparison results of adjacent segments.

[0131] (2) Serial commitment verification

[0132] Verify whether the candidate record fragment sequence determined by the matching index sequence satisfies the sequence commitment covered by the signature or message authentication code in the candidate external content record; if not, generate a sequence integrity feature that represents sequence tampering.

[0133] In this embodiment, the sequence commitment is generated jointly by ordered fragment index information, fragment-level comparable acoustic content fingerprints or comparable fingerprint summaries, and acoustic transfer summaries of adjacent fragments. For example, a leaf node hash can be calculated for each fragment, and then a Merkle root can be generated from all leaf node hashes as the sequence commitment.

[0134] The leaf node hash of a single fragment is calculated using the following formula:

[0135]

[0136] in, For hash functions, This is a byte concatenation operation. For segment number, For the first Fingerprint summary of each fragment, For the first A transfer summary of a fragment.

[0137] Sequence commitments are calculated using the following formula:

[0138]

[0139] in, Calculate the function for the root of the Merkle tree. For the 1st to the 1st The hash value of the leaf node corresponding to each segment.

[0140] Alternatively, sequence commitments can also use hash chains, sequenced message authentication code sequences, signature digest sequences, or combinations thereof to replace Merkle tree roots; digital signatures or certificate chains can be used for public verification scenarios, while message authentication codes can be used for internal verification scenarios.

[0141] (3) Components of sequence integrity

[0142] Sequence integrity features may include at least one of the following: monotonicity violation, index jump, duplicate index, missing index, number or proportion of unmatched segments, inconsistent transfer summaries, segment rearrangement score, segment duplication score, splicing score, and segment replacement score.

[0143] (4) Sequence anomaly determination

[0144] When the matching index sequence indicates that the candidate record segment has skipped, repeated, or reversed sequences that exceed the allowed range of the verification strategy object, the acoustic transfer summary comparison of adjacent segments is inconsistent, or the candidate record segment sequence determined by the matching index sequence cannot meet the sequence commitment, a sequence integrity feature is generated to represent at least one sequence abnormality among segment deletion, insertion, repetition, rearrangement, or splicing; when the number of current audio segments that can participate in the matching is insufficient to complete the sequence integrity judgment, a sequence integrity feature is generated to represent insufficient evidence for sequence verification.

[0145] Step 9: The anomaly detector outputs the structured verification results.

[0146] An anomaly discriminator, which corresponds to the discriminant parameter version and has fixed thresholds or model parameters before the start of the verification phase, is used to output structured verification results based on fragment-level content difference features, cross-fragment sequence integrity features, source token recovery confidence, synchronization score or synchronization confidence, number of valid fragments, and recorded verification status, combined with evidence threshold fields and status priority tables.

[0147] It should be noted that all thresholds or model parameters of the anomaly discriminator are determined before the start of the verification phase. During the verification process, the discriminant parameters are not adjusted for individual streams of speech to be verified. The model weights, normalization parameters, output thresholds, and state mappings are all determined by the discriminant parameter version.

[0148] In one embodiment of a rule-based threshold model, the anomaly decision-maker judges the following in descending order of priority: record anomaly, record not obtained / downgraded permission determination, record incompatibility, insufficient evidence and abstention, partial tampering, sequence tampering, and consistency between source and content; wherein, if the downgraded permission determination does not meet the downgrade conditions, the output is an insufficient evidence and abstention status or a downgrade not allowed status.

[0149] The structured verification result may include at least one of the following: status code, confidence level, evidence length, source token recovery confidence level, record verification status, target synchronization offset, feature version, quantization rule, discriminant parameter version, index of affected audio segments, affected time interval, anomaly type, and degraded status flag.

[0150] Optionally, the anomaly discriminator can be deployed as a rule-based model or as a machine learning model, such as a logistic regression model, a quantile calibration model, an ordinal-preserving regression calibration model, a lightweight neural network model, or a combination thereof; the training or calibration process of the model is completed offline, and the validation phase only reads the fixed parameters corresponding to the discriminant parameter version.

[0151] Step 10: Self-contained downgrade verification path

[0152] (1) Triggering conditions

[0153] Self-contained degradation verification requires the simultaneous fulfillment of the following three degradation conditions: no candidate external content record available for primary verification has been obtained; there is no record in a state where candidate external content records have been obtained but security verification has failed; and the corresponding degradation permission information is obtained from the local trusted policy set based on the verification policy commitment information in the source token, and this degradation permission information indicates that degradation verification is permitted. If all three degradation conditions are met simultaneously, the degradation condition is deemed to be met.

[0154] (2) Verification process

[0155] When the downgrade conditions are met, the following processing is performed: recover the microfingerprint label of the corresponding audio segment from the subsequent current audio segment of the streaming speech to be verified; recalculate the current microfingerprint of the corresponding current audio segment and compare the current microfingerprint with the recovered microfingerprint label; if the recovered evidence length is insufficient, the label recovery confidence is lower than the preset threshold, the microfingerprint error rate is higher than the preset threshold, or the number of delayed segments is insufficient, output an insufficient evidence abandonment status; if the comparison is inconsistent and the evidence is sufficient, output a local anomaly prompt with a downgrade status mark; if the comparison is consistent and the evidence is sufficient, output a downgrade pass result with a downgrade status mark.

[0156] (3) Output constraints

[0157] The downgrade verification result carries a downgrade status marker and is not considered as a consistency status between the source and content of the external content record. It must not overwrite the record exception status corresponding to the record security verification failure.

[0158] In this embodiment, the micro-fingerprint tag is used during the registration phase for the first... Each source audio segment is pre-generated, and after being segmented, it is delayed before being embedded into the first source audio segment. The first to the second Tags for subsequent source audio segments, It is a positive integer, for example, 1 to 8, preferably 2 to 4. The microfingerprint tag may include at least one of the following: fragment index digest, local acoustic fingerprint digest, feature version, parity bit, error correction check bit, and authentication digest.

[0159] Optionally, microfingerprint tags can be generated only for certain key segments (such as segments with strong voice activity, stable energy, or sensitive content); the size can also be adjusted according to the in-band watermark capacity and verification latency requirements. The value of .

[0160] The core technical mechanism of this embodiment includes the following:

[0161] (1) Verification strategy full signature binding mechanism. The signature or message authentication code of the external content record not only covers the fragment-level content verification field, but also the feature version, quantification rule, verification strategy object or verification strategy commitment summary, discrimination parameter version, evidence threshold field, status priority table and downgrade permission field; the verification end ensures that the verification strategy has not been tampered with through the three-party commitment consistency check, and can only use the protected interpretation rules and discrimination parameters for verification, thereby reducing the risk of strategy replacement and rule obfuscation attacks.

[0162] (2) Degradation Prevention Status Priority Mechanism. A strict status priority rule is established: Record Anomaly > Record Not Acquired / Degradation Permission Determination > Record Incompatibility > Insufficient Evidence and Abandonment > Partial Tampering > Sequence Tampering > Source and Content Consistency. Record security verification failure has the highest priority, directly blocking the degradation verification path and prohibiting the use of weak verification results to cover the security verification failure status; if the degradation permission determination does not meet the degradation conditions, an insufficient evidence and abandonment status or a degradation not allowed status is output; in low-evidence scenarios, an insufficient evidence and abandonment status is output to avoid erroneous deterministic output.

[0163] (3) Three-level content integrity verification mechanism. A three-level verification system is constructed: fragment level, adjacency level, and sequence level. Fragment level is based on acoustic content fingerprint to detect local replacement tampering; adjacency level is based on acoustic transfer summary to detect splicing breakpoints; sequence level is based on matching index sequence and sequence commitment to detect global tampering such as rearrangement, deletion, and duplication. The matching algorithm adopts dynamic programming with index monotonicity constraints to reduce the impact of single fragment mismatch on global decision.

[0164] (4) Fixed parameter anomaly judgment mechanism. The threshold, model weights, and normalization parameters of the anomaly judge are all determined by the version of the discrimination parameters and are fixed before the start of the verification phase. During the verification process, the parameters are not temporarily adjusted for a single sample to be verified to prevent the relaxation of verification requirements by adjusting the discrimination criteria.

[0165] (5) Restricted self-contained degradation verification mechanism. Self-contained degradation verification must simultaneously meet three triggering conditions: "no candidate external content record has been obtained, there is no state where the candidate external content record has been obtained but the record security verification has failed, and the degradation permission information indicates that degradation verification is allowed"; the degradation verification result is forced to carry a degradation status mark, and is not regarded as a state of consistency between the source and content of the external content record, is not confused with the main verification conclusion, and does not cover the record abnormal state corresponding to the record security verification failure.

[0166] Example 2:

[0167] To more clearly illustrate this invention, the specific implementation process of the streaming voice content binding and registration method is described below. It should be noted that the segmentation rules in the registration phase are consistent with those in the verification phase.

[0168] Step 1: Segmenting the source speech into segments

[0169] Obtain the source audio to be registered, unify it to a preset sampling rate, and classify it according to a preset segment duration. and overlap ratio The source audio to be registered is divided into multiple source audio segments.

[0170] Optionally, the preset sampling rate can be 16kHz, 22.05kHz, 24kHz, or 48kHz; segment duration The overlap can be from 0.25 seconds to 8 seconds, preferably from 1 second to 2 seconds; It can be 0% to 75%, preferably 25% to 50%.

[0171] Step 2: Generate the verification policy object and the verification policy commitment summary

[0172] First, a structured verification strategy object is generated. In this embodiment, the verification strategy object includes at least the feature version, quantization rule, fingerprint distance type, matching index generation rule, dynamic programming weight, local anomaly threshold, sequence anomaly threshold, evidence length threshold, source token confidence threshold, synchronization scoring threshold, state priority table, downgrade permission field, model identifier, and model parameter summary.

[0173] Then, normalized coding is performed on the verification policy object, and the verification policy commitment summary is calculated using the following formula:

[0174]

[0175] in, For hash functions, For deterministic normalized coding, To verify the policy object; normalized encoding is used to specify the field order, data type, numerical precision, string encoding, and default value handling method to ensure that the same verification policy object receives a consistent verification policy commitment summary on both the registration and verification ends.

[0176] Next, the verification policy object or its trusted copy is registered to the local trusted policy set accessible to the verification end. It should be noted that the external content record can store the complete verification policy object, or it can only store the verification policy commitment summary and obtain the complete verification policy object through the local trusted policy set.

[0177] It should also be noted that the local trusted policy set stores the verification policy object or its trusted copy; the verification end obtains the corresponding verification policy object based on the verification policy commitment information in the source token, and determines the downgrade permission information based on the downgrade permission field therein; the downgrade permission information is consistent with the content indicated by the downgrade permission field in the candidate external content record.

[0178] Step 3: Generate source token

[0179] Generate a source token, which includes at least a record identifier or equivalent index for obtaining external content records, token binding information, and verification policy commitment information; wherein, the verification policy commitment information is a normalized encoded digest or digest truncation value of the verification policy object.

[0180] Optionally, the source token may also include at least one of the following: issuer identifier, voice object identifier, version number, random number, validity period, key identifier, and authentication information.

[0181] Step 4: Source Token Sharding and Watermark Embedding

[0182] The source token is fragmented, and error correction coding and authentication processing are performed on the fragmented payload to obtain an in-band audio watermark payload segment; then the payload segment is embedded into at least a part of the source audio segment to generate watermarked source speech.

[0183] In this embodiment, the payload frame structure includes at least a fragment sequence number field and a payload value field, as well as at least one of an integrity verification field or an authentication field; when carrying multiple types of fragments simultaneously, the payload frame also includes a frame type field.

[0184] Optionally, the error correction coding may employ at least one of Reed-Solomon codes, BCH codes, LDPC codes, convolutional codes, and parity check codes; the underlying watermark may employ at least one of spread spectrum watermarks, quantization index modulation watermarks, phase modulation watermarks, and neural audio watermarks.

[0185] Step 5: Calculate fragment fingerprints and adjacent acoustic transfer summaries

[0186] After preprocessing each source audio segment, a segment-level comparable acoustic content fingerprint or comparable fingerprint summary is calculated, and an acoustic transfer summary between adjacent source audio segments is calculated.

[0187] Optionally, the preprocessing includes at least one of sampling rate unification, loudness normalization, and speech activity detection; the fingerprint feature sources include one or more of log-Mel spectrum features, log-Mel spectrum features compressed by discrete cosine transform, spectral dynamic features, overlapping window features, energy transfer features, and neuroacoustic embeddings.

[0188] Step 6: Generate sequence commitments

[0189] Sequence commitments are generated by combining ordered fragment index information, fragment-level comparable acoustic content fingerprints or comparable fingerprint summaries, and acoustic transfer summaries of adjacent fragments.

[0190] For example, the leaf node hash can be calculated segment by segment, and then a Merkle root can be generated as a sequence commitment. For the specific calculation method, please refer to the description in step 8 of the verification method, which will not be repeated here.

[0191] Optionally, the sequence commitment can be a hash chain, a Merkle root, a sequenced message authentication code sequence, a signature digest sequence, or a combination thereof; digital signatures are used for public verification scenarios, and message authentication codes are used for internal verification scenarios.

[0192] Step 7: Generate external content records

[0193] Generate an external content record, which includes at least the record identifier, token binding field, fragment index information, fragment-level comparable acoustic content fingerprint or comparable fingerprint digest, adjacent fragment acoustic transfer digest, sequence commitment, feature version, quantization rule, verification strategy object, verification strategy commitment digest, discrimination parameter version, evidence threshold field, status priority table, downgrade permission field, record version, validity period field, and revocation field.

[0194] For ease of explanation, the feature version, quantification rule, discrimination parameter version, evidence threshold field, status priority table, and downgrade permission field mentioned above can be collectively referred to as the verification rule and permission configuration field; the record version, validity period field, and revocation field can be collectively referred to as the lifecycle management field.

[0195] Optionally, external content records can take the form of database rows, key-value records, signature objects, transparent log entries, or distributed hash table entries, and are not limited to a single file.

[0196] Step 8: Generate a signature or message authentication code

[0197] Generate signatures or message authentication codes for external content records, so that the content verification fields, verification rules and permission configuration fields, and lifecycle management fields in the external content records are all covered by the signatures or message authentication codes.

[0198] Optionally, a digital signature can be generated using the issuer's private key, or a message authentication code can be generated using a shared key; the signature or message authentication code must at least cover all business fields in the candidate external content record except for itself.

[0199] Step 9: Establishing the relationship between record storage and retrieval

[0200] The signed external content records are stored in at least one of the following: external content record repository, local trusted storage, distributed storage, or transparent log, and a searchable mapping relationship is established between the record identifier or equivalent index in the source token and the external content records.

[0201] Step 10: Self-contained downgrade verification and matching embedding

[0202] For the first Each source audio segment generates a micro-fingerprint tag. The micro-fingerprint tag is then fragmented and embedded with a delay according to a preset multiplexing strategy. The first to the second A subsequent source audio segment, It is a positive integer, for example, 1 to 8, preferably 2 to 4.

[0203] The preset reuse strategy is used to allocate frame type, embedding slot, redundancy count, or priority between source token fragments and microfingerprint tag fragments. Optionally, microfingerprint tags can be generated only for some key segments, and adjustments can be made based on watermark capacity and latency requirements. The value of .

[0204] Example 3:

[0205] This embodiment provides a streaming voice content binding anti-downgrade verification system, including:

[0206] The segmentation module is used to acquire the streaming speech to be verified and divide it into multiple current audio segments;

[0207] The token recovery module is used to perform in-band audio watermark detection on each current audio segment, determine the target synchronization offset, and recover the source token. The source token includes a record identifier or equivalent index, token binding information, and verification strategy commitment information.

[0208] The recording verification module is used to attempt to obtain candidate external content records corresponding to the multiple current audio segments based on the recording identifier or equivalent index; if obtained, the candidate external content records are subjected to recording security verification and recording compatibility verification; if the recording security verification fails, a recording exception is output and the downgrade verification path is blocked; if the recording security verification passes but the recording compatibility verification fails, a recording incompatibility or insufficient evidence abstention status is output; if both verifications pass, the main verification module is triggered.

[0209] The main verification module is used to recalculate the segment-level comparable acoustic content fingerprint for each current audio segment according to the verification parameters protected by the signature in the candidate external content record, and generate segment-level content difference features and cross-segment sequence integrity features; based on the segment-level content difference features, the cross-segment sequence integrity features and the recorded verification status, it performs comprehensive verification through an anomaly decision-maker and outputs a structured verification result.

[0210] The downgrade verification module is used to determine downgrade permissions based on the verification policy commitment information when no candidate external content record is obtained or when there is no candidate external content record that has been obtained but the record security verification has failed. If the downgrade conditions are met, a self-contained downgrade verification is performed, and a verification result marked as downgraded is output. The verification result is not considered as a state of consistency between the source and content of the external content record. If the downgrade conditions are not met, a state of insufficient evidence and abstention or downgrade not allowed is output.

[0211] The product carriers and deployment forms corresponding to the above methods and systems are as follows:

[0212] In one embodiment, a streaming voice processing device is provided, including one or more processors and a memory connected to the one or more processors, the memory storing a computer program that, when executed by the one or more processors, implements the above-described verification method or registration method.

[0213] In one embodiment, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the above-described verification method or registration method.

[0214] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the above-described verification method or registration method.

[0215] In terms of engineering deployment, hardware devices can be deployed as servers, clients, streaming media platform nodes, mobile terminals, evidence collection workstations, edge computing devices, or embedded devices; service forms can take the form of software development kits, cloud service interfaces, or platform verification services.

[0216] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for preventing downgrades in streaming voice content binding, characterized in that, include: Obtain the streaming audio to be verified and divide it into multiple current audio segments; In-band audio watermark detection is performed on each current audio segment to determine the target synchronization offset and recover the source token. The source token includes a record identifier or equivalent index, token binding information, and verification policy commitment information. The verification policy commitment information is a normalized encoded digest or digest truncation value of the verification policy object. Based on the record identifier or equivalent index, attempt to obtain candidate external content records corresponding to the multiple current audio segments; if obtained, perform record security verification and record compatibility verification on the candidate external content records; if the record security verification fails, output a record exception and block the downgrade verification path; if the record security verification passes but the record compatibility verification fails, output a record incompatibility or insufficient evidence abstention status; if both verifications pass, proceed to the main verification. In the main verification, the segment-level comparable acoustic content fingerprint is recalculated for each current audio segment to generate segment-level content difference features and cross-segment sequence integrity features; Based on the fragment-level content difference characteristics, the cross-fragment sequence integrity characteristics, and the recorded verification status, a comprehensive verification is performed by an anomaly decision-maker, and a structured verification result is output. If no candidate external content record is obtained and there is no candidate external content record that has been obtained but the security verification failed, the downgrade permission is determined according to the verification policy commitment information; if the downgrade condition is met, self-contained downgrade verification is performed, and the verification result marked as downgraded is output. The verification result is not considered as the consistency between the source and content of the external content record. If the conditions for downgrading are not met, the output will be either "insufficient evidence, abstention" or "downgrading not allowed". The self-contained degradation verification includes: recovering the microfingerprint tag of the corresponding audio segment from the subsequent current audio segment of the streaming speech to be verified, recalculating the current microfingerprint of the corresponding current audio segment, and comparing the current microfingerprint with the recovered microfingerprint tag; if the recovered evidence length is insufficient, the tag recovery confidence is lower than a preset threshold, the microfingerprint error rate is higher than a preset threshold, or the number of delayed segments is insufficient, then an insufficient evidence abstention status is output; if the comparison is inconsistent and the evidence is sufficient, then a local anomaly prompt with a degradation status mark is output; if the comparison is consistent and the evidence is sufficient, then a degradation pass result with a degradation status mark is output.

2. The method for preventing degradation of streaming voice content binding according to claim 1, characterized in that, Determining the target synchronization offset includes: selecting multiple probe windows from the current audio segment, generating multiple candidate synchronization offsets based on the probe windows; performing in-band audio watermark detection on each candidate synchronization offset, and calculating the synchronization score corresponding to each candidate synchronization offset based on the watermark detection confidence, payload bit error rate estimation, error correction decoding status, payload fragment consistency, authentication verification consistency, and candidate offset stability; selecting the candidate synchronization offset that meets both the synchronization score threshold and the candidate offset stability condition as the target synchronization offset; and outputting an insufficient evidence abstention status when no candidate synchronization offset meets the conditions. The process of recovering the source token includes: extracting candidate payload fragments under the target synchronization offset, determining whether the number of valid fragments meets the minimum fragment number requirement; if the minimum fragment number requirement is not met, rejecting the source token and outputting a token recovery failure status or an insufficient evidence abandonment status; if the minimum fragment number requirement is met, reassembling the candidate payload based on the fragment sequence number, and sequentially performing error correction decoding, integrity verification, and authentication verification on the candidate payload; if all three verifications pass, outputting the source token and the source token recovery confidence level; if any one verification fails, rejecting the source token and outputting a token recovery failure status or an insufficient evidence abandonment status. The verification policy commitment information in the source token is a normalized encoded digest of the verification policy object.

3. The method for preventing degradation of streaming voice content binding according to claim 1, characterized in that, The candidate external content record includes a record identifier, a token binding field, fragment index information, fragment-level comparable acoustic content fingerprint or comparable fingerprint digest, adjacent fragment acoustic transfer digest, sequence commitment, feature version, quantization rule, verification strategy object, verification strategy commitment digest, discrimination parameter version, evidence threshold field, status priority table, downgrade permission field, record version, validity period field, revocation field, and signature or message authentication code; the signature or message authentication code covers the other fields in the candidate external content record.

4. The method for preventing degradation of streaming voice content binding according to claim 3, characterized in that, Record compatibility verification includes the processing compatibility verification of record version, feature version, quantization rule and discrimination parameter version; if the record version, feature version, quantization rule or discrimination parameter version is not supported by the verification end, the record incompatibility status or insufficient evidence abstention status will be output, and the source and content consistency status will not be output. The record security verification includes signature or message authentication code verification, token binding relationship verification, verification policy commitment relationship verification, validity period verification, and revocation status verification; wherein: The verification strategy commitment relationship verification includes: reading the verification strategy object from the candidate external content record, encoding the verification strategy object according to a preset normalized encoding method, and recalculating the verification strategy commitment digest based on the encoding result; comparing the recalculated verification strategy commitment digest with the verification strategy commitment information in the source token and the verification strategy commitment digest in the candidate external content record that is covered by the signature or message authentication code; if the three are inconsistent, the candidate external content record is determined to have failed record security verification; if the comparison cannot be completed because the normalized encoding method is not supported or the normalized encoding result cannot be uniquely determined, a record incompatibility status or insufficient evidence abstention status is output. If the signature or message authentication code verification, the token binding relationship verification, the verification policy commitment relationship verification, the validity period verification, and the revocation status verification all pass, the candidate external content record is determined to have passed the record security verification; if any one of the verifications fails, the candidate external content record is determined to have failed the record security verification, an abnormal record status is output, and the execution of a self-contained degradation verification path that can output a status of consistency between the source and content or a degradation pass status is prohibited.

5. The method for preventing degradation of streaming voice content binding according to claim 4, characterized in that, The signature or message authentication code verification includes: verifying the signature or message authentication code in the candidate external content record; if the signature or message authentication code verification fails, the candidate external content record is determined to be a record security verification failure. The token binding relationship verification includes: comparing the token binding information in the source token with the token binding field in the candidate external content record; if the two do not satisfy the predetermined binding relationship, the candidate external content record is determined to be a record security verification failure. The validity period verification includes: determining whether the candidate external content record has expired based on the validity period field in the candidate external content record; if the candidate external content record has expired, the candidate external content record is determined to be a record security verification failure. The revocation status verification includes: determining whether the candidate external content record has an abnormal revocation status based on the revocation field in the candidate external content record; if an abnormal revocation status is determined, the candidate external content record is determined to have failed the record security verification.

6. The method for preventing degradation of streaming voice content binding according to claim 3, characterized in that, The generated fragment-level content difference features include: Based solely on the feature version covered by the signature or message authentication code in the candidate external content record and the quantization rule, recalculate the segment-level comparable acoustic content fingerprint or comparable fingerprint digest for each current audio segment; The recalculated fragment-level comparable acoustic content fingerprint or comparable fingerprint summary is compared with the fragment-level comparable acoustic content fingerprint or comparable fingerprint summary in the candidate external content record to generate fragment-level content difference features.

7. The method for preventing degradation of streaming voice content binding according to claim 6, characterized in that, The methods for generating cross-segment sequence integrity features include: For each current audio segment, the recalculated segment-level comparable acoustic content fingerprint or comparable fingerprint summary is compared with the segment-level comparable acoustic content fingerprints or comparable fingerprint summaries of multiple candidate record segments in the candidate external content records, and the corresponding fingerprint distance is calculated; candidate record segments whose fingerprint distance meets the matching threshold are determined as candidate matching segments for the corresponding current audio segment; if no candidate record segment meets the matching threshold, the corresponding current audio segment is marked as an unmatchable segment; Based on the candidate matching segments, corresponding fingerprint distances, and unmatched markers for each current audio segment, a matching index sequence is determined through dynamic programming with index monotonicity constraints. The path cost of the dynamic programming includes at least fingerprint distance cost, matching index jump cost, repeated matching index cost, index reversal cost, and unmatched cost, and is corrected based on the comparison results between the acoustic transfer summaries between adjacent current audio segments and the acoustic transfer summaries between adjacent segments of candidate record segments. The weights and matching thresholds used in the dynamic programming are determined by the verification strategy object or the discriminant parameter version. The matching index sequence is arranged according to the time order of the current audio segment, and the continuous matching, skipping, repetition, reverse order and unmatched cases of candidate record segments are identified according to the matching index sequence; for adjacent matched current audio segments, their acoustic transfer summaries are compared with the acoustic transfer summaries of adjacent segments of the corresponding candidate record segments to obtain the acoustic transfer summaries comparison results of adjacent segments; The candidate record segment sequence determined according to the matching index sequence is compared with the sequence commitment covered by the signature or message authentication code in the candidate external content record to obtain the sequence commitment verification result; the sequence commitment is generated jointly based on the ordered segment index information, the segment-level comparable acoustic content fingerprint or comparable fingerprint digest, and the acoustic transfer digest of adjacent segments. The cross-segment sequence integrity feature is generated based on the matching index sequence, the acoustic transfer summary alignment results of adjacent segments, and the sequence commitment verification results. When the matching index sequence indicates that the candidate record segment has skipped, repeated, or reversed sequences that exceed the allowed range of the verification strategy object, the acoustic transfer summaries of adjacent segments are inconsistent, or the candidate record segment sequence determined according to the matching index sequence cannot meet the sequence commitment, a sequence integrity feature is generated to characterize at least one sequence abnormality among segment deletion, insertion, repetition, rearrangement, or splicing; when the number of current audio segments that can participate in the matching is insufficient to complete the sequence integrity judgment, a sequence integrity feature is generated to characterize insufficient sequence verification evidence.

8. The method for preventing degradation of streaming voice content binding according to claim 3, characterized in that, The threshold or model parameters of the anomaly decision device correspond to the version of the discrimination parameters, and all thresholds or model parameters are determined before the start of the verification phase. No adjustments are made to a single stream of speech to be verified during the verification process. The anomaly decision-maker takes the fragment-level content difference features, the cross-fragment sequence integrity features, the source token recovery confidence, the synchronization score or synchronization confidence, the number of valid fragments, and the record verification status as input, and combines the evidence threshold field and the status priority table to output the structured verification results.

9. The method for preventing degradation of streaming voice content binding according to claim 1, characterized in that, The degradation permission determination includes three degradation conditions: 1) no candidate external content record available for primary verification has been obtained; 2) no candidate external content record has been obtained but security verification has failed; and 3) corresponding degradation permission information is obtained from the local trusted policy set according to the verification policy commitment information in the source token, and the degradation permission information indicates that degradation verification is allowed. If all three degradation conditions are met simultaneously, the degradation condition is deemed to be met. The self-contained degradation verification includes: recovering the microfingerprint tag of the corresponding audio segment from the subsequent current audio segment of the streaming speech to be verified, recalculating the current microfingerprint of the corresponding current audio segment, and comparing the current microfingerprint with the recovered microfingerprint tag; if the recovered evidence length is insufficient, the tag recovery confidence is lower than a preset threshold, the microfingerprint error rate is higher than a preset threshold, or the number of delayed segments is insufficient, then an insufficient evidence abstention status is output; if the comparison is inconsistent and the evidence is sufficient, then a local anomaly prompt with a degradation status mark is output; if the comparison is consistent and the evidence is sufficient, then a degradation pass result with a degradation status mark is output. The verification results marked as degraded must not overwrite the abnormal status of the record corresponding to the security verification failure.

10. A streaming voice content binding anti-downgrade verification system, characterized in that, include: The segmentation module is used to acquire the streaming speech to be verified and divide it into multiple current audio segments; The token recovery module is used to perform in-band audio watermark detection on each current audio segment, determine the target synchronization offset, and recover the source token. The source token includes a record identifier or equivalent index, token binding information, and verification strategy commitment information. The recording verification module is used to attempt to obtain candidate external content records corresponding to the multiple current audio segments based on the recording identifier or equivalent index; if obtained, the candidate external content records are subjected to recording security verification and recording compatibility verification; if the recording security verification fails, a recording exception is output and the downgrade verification path is blocked; if the recording security verification passes but the recording compatibility verification fails, a recording incompatibility or insufficient evidence abstention status is output; if both verifications pass, the main verification module is triggered. The main verification module is used to recalculate the segment-level comparable acoustic content fingerprint for each current audio segment according to the verification parameters protected by the signature in the candidate external content record, and generate segment-level content difference features and cross-segment sequence integrity features. Based on the fragment-level content difference characteristics, the cross-fragment sequence integrity characteristics, and the recorded verification status, a comprehensive verification is performed by an anomaly decision-maker, and a structured verification result is output. The downgrade verification module is used to determine the downgrade permission based on the verification policy commitment information when no candidate external content record is obtained or when there is no state where a candidate external content record has been obtained but the record security verification has failed. If the downgrade condition is met, self-contained downgrade verification is performed, and the verification result marked as downgraded is output. The verification result is not used as a status of consistency between the source and content of the external content record. If the conditions for downgrading are not met, the status will be either "insufficient evidence, abstention" or "downgrading not allowed".

Images