A streaming voiceprint recognition method, electronic device and storage medium

Abstract

The application provides a streaming voiceprint recognition method, which establishes a continuous accumulation state including a full-window voiceprint vector accumulation and a full-window count for a current speech, acquires a to-be-processed speech segment ending at a current processing time from an audio stream and extracts a current voiceprint vector according to a preset processing period, and divides the current voiceprint vector into a full-window voiceprint vector or an under-length window voiceprint vector according to an actual audio length; when the under-length window and the full-window count are zero, temporarily matches the current voiceprint vector without updating the accumulation state, when the full window, adds the current voiceprint vector to the accumulation and generates an accumulated recognition vector, when the full-window count is greater than zero, matches the accumulated recognition vector with a registered voiceprint template, and performs a multi-threshold decision combining a similarity, an interval value and a confidence, and outputs a current processing period recognition result containing a round number, a user identifier, a confidence and a state.

Description

Technical Field

[0001] This application relates to the fields of speech signal processing and biometric recognition technology, and in particular to a streaming voiceprint recognition method, electronic device and storage medium for embodied intelligent devices such as robots, smart speakers and vehicle terminals. Background Technology

[0002] In human-computer voice interaction, embodied intelligent devices typically need to identify the current speaker's identity in real time in order to perform personalized responses, access control, recall of conversation memories, or confirmation of sensitive actions. To obtain real-time output, existing solutions often divide the microphone audio into fixed-length scrolling windows, extract the speaker embedding vector independently for each window, and compare the similarity of this embedding vector with the registered voiceprint template one by one.

[0003] In the single-window determination method described above, phoneme content within a short window, environmental noise, far-field reverberation, speaker posture, or volume changes can all cause similarity fluctuations, leading to frequent jumps in recognition results within the same utterance. Especially when the window is not yet filled or the effective speech duration is insufficient, the statistical pooling and cepstral mean normalization processes in the embedding model may be based on insufficient frame estimation, thus causing systematic bias in the short-window embedding.

[0004] If we simply apply equal-weighted averaging to the embedding vectors of each window, or assign non-zero weights to short windows based on their length, the bias vectors from the early stages of short sentences will be included in the subsequent cumulative results, affecting the stability of long utterances. If we repeatedly extract embedding vectors from all audio from the start of the utterance in each processing cycle, although we can utilize the entire information, the computational load will increase with the duration of the utterance, which is not conducive to deployment on low-computing-power edge devices.

[0005] Furthermore, existing solutions often rely on a single similarity threshold to determine identity. When the candidate identity corresponding to the highest similarity score is close to the second candidate identity, the similarity threshold alone is insufficient to distinguish between high-scoring correct matches and high-scoring incorrect matches. Simultaneously, the recognition results typically lack confidence information that downstream agents can consume based on risk levels. Therefore, it is necessary to provide a voiceprint recognition scheme suitable for streaming processing, edge deployment, and capable of suppressing false recognition. Summary of the Invention

[0006] The technical problem to be solved by this application is to improve the stability and accuracy of speaker recognition results within the same discourse under the constraints of streaming, far-field and low computing power, reduce the contamination of accumulated evidence by short windows, and provide downstream services with recognition results with state cause and confidence.

[0007] In a first aspect, this application provides a streaming voiceprint recognition method, comprising: A continuous accumulation state is established for the current discourse, which includes at least the accumulated amount of the full-window voiceprint vector and the full-window count; According to the preset processing cycle, the speech segment to be processed with the current processing time as the end time is obtained from the audio stream corresponding to the current speech, and the current voiceprint vector is extracted. Based on whether the actual audio duration in the speech segment to be processed reaches the preset window duration, the current voiceprint vector is determined to be a full-window voiceprint vector or a short-window voiceprint vector. When the current voiceprint vector is a short-window voiceprint vector and the full-window count is zero, the speaker identity matching is performed using the current voiceprint vector as a temporary identification vector, and the continuous accumulation state is not updated based on the current voiceprint vector. When the current voiceprint vector is a full-window voiceprint vector, add the current voiceprint vector to the cumulative amount, update the full-window count, and generate a cumulative recognition vector based on the updated cumulative amount and full-window count; When the full window count is greater than zero, the speaker recognition result is generated by matching the cumulative recognition vector with the registered voiceprint template. Output the speaker recognition result for the current processing cycle.

[0008] In a second aspect, this application provides an electronic device, including a processor, a memory, and an audio access unit, wherein the memory stores a computer program, the audio access unit is used to receive an audio stream corresponding to the current utterance, and the computer program, when executed by the processor, implements the streaming voiceprint recognition method according to any one of claims 1 to 14.

[0009] Thirdly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the streaming voiceprint recognition method according to any one of claims 1 to 14.

[0010] The technical effects are as follows: by adding only the full-window voiceprint vector to the cumulative amount during the discourse lifecycle, the update weight of the short window is zero, which avoids the pollution of the cumulative state by the bias vector in the early stage of short sentences or when there is insufficient effective speech; by using the current voiceprint vector for temporary matching when there is no full-window evidence, the real-time output capability in short discourse scenarios can be maintained; by using the cumulative recognition vector for matching after full-window evidence exists, long discourses can obtain more stable identity determination over time.

[0011] Furthermore, through orthogonal constraints of similarity threshold, interval threshold, and confidence threshold, it is possible to determine whether candidate identities are similar, whether candidate identities are exclusive, and whether current evidence is sufficient. Through four-factor confidence output, downstream businesses can consume identification results according to action risk levels. Through discourse start signals, silence fallback, and round number management, the identification results can be kept consistent with the upper-level dialogue rounds and the ambiguity of attribution when clearing is reduced. Attached Figure Description

[0012] Figure 1 A schematic diagram of the streaming voiceprint recognition processing architecture provided in this application; Figure 2 A schematic diagram of the single-beat processing flow of the streaming voiceprint recognition method provided in this application; Figure 3 A schematic diagram illustrating the accumulation of centroids across a full window and the fallback effect of a short window provided for this application; Figure 4 A schematic diagram of the multi-threshold decision and confidence output logic provided for this application; Figure 5 A timing diagram of the round control provided in this application. Detailed Implementation

[0013] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described content is merely an exemplary illustration for ease of understanding and is not intended to limit the scope of protection of this application. Based on the content described in this application, other implementations obtained by those skilled in the art without inventive effort should all fall within the scope of protection of this application.

[0014] It should be noted that, where there is no conflict, the various technical features described in this application can be combined with each other. For ease of description, the accompanying drawings only show the parts of the structure or process related to this application. The division of steps, process sequence, size scale, and connection relationship in the drawings are all schematic and should not be construed as limiting this application.

[0015] It should also be noted that the terms "first," "second," etc., used herein are used only to distinguish the same or similar objects and do not indicate a specific order, primary or secondary relationship, or quantity limitation. The terms "comprising," "including," or any other variations thereof, as used herein, are intended to cover non-exclusive inclusion, such that a method, apparatus, or medium that includes a list of elements may include not only the elements already listed but also other elements not expressly listed.

[0016] The flowcharts shown in the accompanying drawings are for illustrative purposes only and are not required to include all the steps shown, nor are they required to be performed in the strict order shown. In some embodiments, some steps may be split, combined, executed in parallel, or selectively executed based on business triggering conditions.

[0017] In this application, a speech can represent a single voice interaction unit defined by the business side, such as a sentence, user input within a single round, or an audio segment between two speech initiation signals. The audio stream corresponding to the current speech can come from a robot microphone array, smart speaker, vehicle terminal, conference terminal, or other devices with audio acquisition capabilities.

[0018] In this application, the preset processing cycle can be triggered once every 500 milliseconds, or it can be set to other cycles according to the device's computing power, end-to-end latency requirements, or business scenarios. The preset window duration can be 2.5 seconds, or it can be 2 seconds, 3 seconds, or other durations suitable for stable extraction of the embedded model.

[0019] In this application, the voiceprint vector can be a fixed-length vector output by a pre-trained speaker embedding model. For example, after L2 normalization, this voiceprint vector can be matched with a registered voiceprint template using cosine similarity. This application does not limit the network structure, number of parameters, training data source, or output dimension of the speaker embedding model.

[0020] In this application, the actual audio duration can be the effective speech duration retained after silence gating in the speech segment to be processed. Optionally, silence gating can be implemented based on an energy threshold, a neural network speech activity detection model, or a combination of both.

[0021] In this application, a full-window voiceprint vector refers to the voiceprint vector corresponding to a speech segment whose actual audio duration reaches the preset window duration; a short-window voiceprint vector refers to the voiceprint vector corresponding to a speech segment whose actual audio duration does not reach the preset window duration. Short-window voiceprint vectors can be used for temporary matching, but are not used to update the cumulative amount of full-window voiceprint vectors.

[0022] See Figure 1The streaming voiceprint recognition process can be executed in stages, including audio access, voiceprint vector extraction, streaming evidence accumulation, multi-threshold decision, confidence assessment, round control, and result output. Audio access is used to receive audio streams and speech start signals sent by the service side; voiceprint vector extraction is used to extract the current voiceprint vector according to a preset processing cycle; streaming evidence accumulation is used to maintain the cumulative amount and full-window count in the speech pool; multi-threshold decision is used to output the identity or status reason based on similarity, interval value, and confidence; confidence assessment is used to generate confidence levels that can be consumed in a tiered manner; and round control is used to reset the continuous accumulation state and update the round number when speech start or silence fallback is triggered.

[0023] For example, the output can be generated at a fixed pace, consisting of a round number, user identifier, confidence level, and status. When no trusted identity exists, the user identifier can use a preset sentinel value, such as 0; the status can indicate reasons such as insufficient evidence, insufficient interval, no match, multiple overlapping users, or silent reset. Thus, downstream agents can obtain an availability description of the identity information while receiving it.

[0024] See Figure 2 This application provides a streaming voiceprint recognition method, including: A continuous accumulation state is established for the current discourse, which includes at least the accumulated amount of the full-window voiceprint vector and the full-window count; According to the preset processing cycle, the speech segment to be processed with the current processing time as the end time is obtained from the audio stream corresponding to the current speech, and the current voiceprint vector is extracted. Based on whether the actual audio duration in the speech segment to be processed reaches the preset window duration, the current voiceprint vector is determined to be a full-window voiceprint vector or a short-window voiceprint vector. When the current voiceprint vector is a short-window voiceprint vector and the full-window count is zero, the speaker identity matching is performed using the current voiceprint vector as a temporary identification vector, and the continuous accumulation state is not updated based on the current voiceprint vector. When the current voiceprint vector is a full-window voiceprint vector, add the current voiceprint vector to the cumulative amount, update the full-window count, and generate a cumulative recognition vector based on the updated cumulative amount and full-window count; When the full window count is greater than zero, the speaker recognition result is generated by matching the cumulative recognition vector with the registered voiceprint template. Output the speaker recognition result for the current processing cycle.

[0025] See Figure 2 Within a processing cycle, the streaming voiceprint recognition method may include the following steps.

[0026] Step S101: Establish a continuous accumulation state for the current discourse; At the start of the current utterance, the system establishes or resets the continuous accumulation state. The continuous accumulation state includes at least the accumulated amount A of the full-window voiceprint vector and the full-window count N. The accumulated amount A can be initialized to a zero vector, and the full-window count N can be initialized to zero. Optionally, the continuous accumulation state may also include the round number, effective speech duration, the most recent accumulated recognition vector, the previous output state, the silence duration, and the parameter configuration version.

[0027] Step S102: Obtain the speech segment to be processed and extract the current voiceprint vector; The system retrieves a speech segment to be processed from the audio stream corresponding to the current utterance, with the current processing time as the end time, according to a preset processing cycle. This segment can be a scrolling window speech segment with a length not exceeding the preset window duration. Subsequently, the system inputs the speech segment to be processed into the speaker embedding extraction model to obtain the current voiceprint vector e_i. For example, e_i is a unit vector for calculating cosine similarity with the registered voiceprint template.

[0028] Step S103: Determine whether the actual audio duration has reached the preset window duration. The system performs silence gating on the speech segments to be processed, determining the actual audio duration of the valid speech. If the actual audio duration reaches the preset window duration, the current voiceprint vector is determined to be a full-window voiceprint vector; if the actual audio duration does not reach the preset window duration, the current voiceprint vector is determined to be an under-window voiceprint vector. By setting the judgment criterion to the actual audio duration, low-quality voiceprint vectors caused by excessively high silence ratios within the window can be prevented from accumulating.

[0029] Step S104: Use the current voiceprint vector as a temporary recognition vector and do not update the continuously accumulated state. When the current voiceprint vector is a short-window voiceprint vector and the full-window count N is zero, it indicates that the current utterance has not yet formed a full-window evidence that can be accumulated. At this time, the system can use the current voiceprint vector as a temporary identification vector for speaker identity matching to meet the immediate output requirements of short utterances or early utterances. At the same time, the system does not add this short-window voiceprint vector to the cumulative amount A, nor does it update the full-window count N, so that the update weight of the short-window for the continuous accumulation state is zero.

[0030] Step S105: Add the cumulative amount and generate the cumulative recognition vector. When the current voiceprint vector is a full-window voiceprint vector, the system adds the current voiceprint vector to the cumulative amount A and increments the full-window count N by one. Subsequently, the system generates an average vector c = A / N based on the updated cumulative amount A and the full-window count N, and normalizes the average vector c to obtain the cumulative recognition vector. =c / ‖c‖. In some embodiments, if N is greater than zero and the current voiceprint vector is a short-window voiceprint vector, the system does not update the cumulative amount A and the full-window count N, and continues to use the generated cumulative recognition vector for matching.

[0031] Step S106: Matching based on accumulated recognition vectors When the full-window count N is greater than zero, the system matches the accumulated recognition vector with the registered voiceprint templates to generate speaker recognition results. Specifically, the system can calculate the cosine similarity between the accumulated recognition vector and each registered voiceprint template to obtain the first similarity s1, the first candidate identity u1, and the second similarity s2, and determine the margin value margin=s1-s2. In the short sentence stage where the full-window count N is zero, the system can use the current voiceprint vector as a temporary recognition vector to perform the same similarity ranking.

[0032] Step S107: Output the speaker recognition result for the current processing cycle: The system outputs the speaker recognition result for the current processing cycle. This result may include the round number, user identifier, confidence level, and status. When the judgment condition is met, the user identifier is the first candidate identity u1; when the judgment condition is not met, the user identifier is the preset sentinel value, and the status is used to explain why a trusted identity was not output.

[0033] See Figure 3 The streaming evidence accumulation in this application can be understood as a dual-path processing of a short-window temporary identification path and a full-window cumulative identification path. For the i-th scrolling window after the start of the speech, if its voiceprint vector is e_i and the actual audio duration corresponding to the scrolling window reaches the preset window duration, then e_i is included in the full-window set F; if the actual audio duration does not reach the preset window duration, then it is not included in F.

[0034] In some embodiments, the streaming evidence accumulation processing maintains the cumulative amount A = Σe_i of the full-window voiceprint vector and the full-window count N = |F|. When N is greater than zero, the system calculates c = A / N and normalizes c to obtain the accumulated identification vector. Since the short window contributes nothing to A and N, the short window voiceprint vector will not pollute the subsequent accumulation results by equal weight averaging or weighting by window length.

[0035] For example, in the short sentence limit, if F is empty, the system uses the latest voiceprint vector as a temporary recognition vector, which can maintain timeliness similar to single-window recognition of the whole sentence. In the long sentence scenario, as the number of full windows increases, the cumulative recognition vector is determined by multiple high-quality full-window voiceprint vectors, which can reduce random fluctuations in single windows. The marginal computation of each processing cycle mainly consists of one vector addition, count update, and normalization, and the computational cost is basically independent of the total duration of the speech.

[0036] Optionally, in deployments with sufficient computing power, the system can also maintain an audio buffer that has been silence-gated since the start of the speech, and re-extract the entire voiceprint vector from this audio buffer in each processing cycle. This recalculated audio buffer can share the same external output interface as the full-window centroid accumulation file, and its activation can be determined by device computing power, latency constraints, or session configuration.

[0037] See Figure 4 After obtaining the matching similarity ranking, the system can record the first similarity as s1, the first candidate identity as u1, and the second similarity as s2, and define the interval value as margin = s1 - s2. The system outputs the first candidate identity u1 only when s1 reaches the similarity threshold T_sim, margin reaches the interval threshold T_margin, and confidence conf reaches the confidence threshold T_conf.

[0038] The similarity threshold is used to determine whether the recognition vector is sufficiently similar to a registered voiceprint template; the interval threshold is used to determine whether the first candidate identity is exclusive to the second candidate identity, thus forming a false recognition suppression defense; and the confidence threshold is used to filter low-evidence judgments in the early stages of accumulation or when there is insufficient evidence. The three types of thresholds can be configured as process-level default parameters or can be overridden by session-level parameters to adapt to different operating points of high security or high recall.

[0039] In some embodiments, when the registered voiceprint template library contains only one registered voiceprint template, the system can convert the interval value into the difference between s1 and T_sim to maintain exclusiveness judgment in a single registered user scenario. Optionally, when the similarity of at least two candidate identities reaches the overlap detection threshold, the system can output the status as multiple people overlapping, or refuse to output a specific user identifier.

[0040] In some embodiments, the confidence score `conf` can be determined based on the duration saturation factor `f_dur`, the similarity factor `f_sim`, the interval factor `f_margin`, and the centroid modulus factor `f_norm`. For example, `conf = f_dur·f_sim·f_margin·f_norm`. `f_dur` can be equal to `min(D / T_sat, 1)`, where `D` is the cumulative effective speech duration obtained by counting the newly added effective speech duration after silence gating in the current utterance, and `T_sat` is the duration saturation value. By using the newly added effective speech duration instead of simply adding the durations of each scroll window, the artificial increase in duration caused by overlapping scroll windows can be avoided.

[0041] Optionally, f_sim can be a monotonically increasing function of s1-T_sim, f_margin can be a monotonically increasing function of margin, and f_norm can be the magnitude of the average vector c before normalization. The centroid magnitude factor is used to reflect the consistency of the direction of each full-window speaker vector. When multiple full-window speaker vectors come from the same speaker and the environment is stable, the magnitude of the average vector approaches 1; when other speakers are mixed in, multiple speakers overlap, or there is strong noise, the magnitude of the average vector decreases. Since c has already been obtained in the process of accumulating recognition vector generation, the calculation of the centroid magnitude factor does not require additional model inference.

[0042] In some embodiments, confidence level and identity determination can be decoupled. That is, even if the system outputs a preset sentinel value without outputting a specific identity, it can still output the confidence level and the reason for the status to downstream services. Downstream services can consume the identification results in a tiered manner based on the confidence level. For example, a higher confidence level allows high-risk personalized actions, a medium confidence level only allows low-risk actions, and a lower confidence level ignores the identity information.

[0043] See Figure 5 The continuous accumulation state can have a lifecycle based on the utterance. When a new utterance is detected, the service side can send a utterance start signal through the control channel corresponding to the audio stream. In response to this utterance start signal, the system clears the accumulated amount A, sets the full window count N to zero, and increments or updates the round number.

[0044] In some embodiments, to improve fault tolerance, the system can also set a fallback mechanism for silence. When the duration of no valid voice messages exceeds the fallback silence duration, the system automatically clears the continuously accumulated state. The fallback silence duration can be configured according to the scenario, for example, set to 6 seconds, or it can be overridden by session-level parameters.

[0045] Optionally, the round number can be output along with the recognition result of each processing cycle, enabling downstream agents to align the recognition result with their own dialogue rounds and avoid ambiguity at the moment the accumulated state is cleared. The business side can define its own utterance start signal to represent each sentence, each round, or other granularities. The recognition system of this application maintains a predictable state lifecycle based on this signal and the silence fallback rule.

[0046] In some embodiments, the system can also perform intra-speaker change detection. For example, the system can trigger intra-speaker change detection based on the decrease in the centroid magnitude of the cumulative identification vector relative to historical values; it can also trigger intra-speaker change detection based on the inconsistency between the immediate candidate identity corresponding to the temporary identification vector and the cumulative candidate identity corresponding to the cumulative identification vector over multiple consecutive processing cycles. When intra-speaker change detection is triggered, the system clears the cumulative amount A, sets the full-window count N to zero, and updates the round number.

[0047] In some embodiments, silence gating can be implemented using an energy threshold, a neural network speech activity detection model, or a combination of both. Optionally, the recognition path can use a first silence threshold, and the registration path can use a second silence threshold, with the first and second silence thresholds being different. For example, the far-field recognition path can use a lower threshold to preserve weak speech, while the registration quality control path can use a higher threshold to improve template quality.

[0048] In some embodiments, registered voiceprint templates can be stored in isolation according to tenant, user ID, and model version. During matching, the system calculates cosine similarity within the set of registered voiceprint templates corresponding to the current tenant and the current model version. The storage layer can store only user IDs and voiceprint templates, and the mapping between user IDs and real identities is maintained by the business side.

[0049] Optionally, the embedding extraction model can be replaced with any deep neural network model capable of outputting fixed-length speaker vectors. Different models can have different numbers of parameters, vector dimensions, or training data sources. The cumulative amount can also be applied using an exponentially decaying weighted form, provided business requirements are met, so that earlier full-window evidence decays over time.

[0050] In some embodiments, this application also provides an electronic device, including a processor, a memory, and an audio access unit. The memory stores a computer program, and the audio access unit is used to receive an audio stream corresponding to the current utterance. When the computer program is executed by the processor, it implements the streaming voiceprint recognition method according to any one of claims 1 to 14. The electronic device may be a robot, a smart speaker, a vehicle terminal, a conference terminal, an edge computing device, or a server.

[0051] In some embodiments, this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the streaming voiceprint recognition method according to any one of claims 1 to 14. The computer-readable storage medium may include a magnetic disk, optical disk, read-only memory, random access memory, flash memory, or other non-transitory storage medium capable of storing a computer program.

[0052] The storage medium can be the robot's internal storage unit as described in the previous embodiments, such as the robot's hard drive or memory. Alternatively, the storage medium can be the robot's external storage device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card.

[0053] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0054] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A streaming voiceprint recognition method, characterized in that, include: A continuous accumulation state is established for the current discourse, which includes at least the accumulated amount of the full-window voiceprint vector and the full-window count; According to the preset processing cycle, the speech segment to be processed with the current processing time as the end time is obtained from the audio stream corresponding to the current speech, and the current voiceprint vector is extracted. Based on whether the actual audio duration in the speech segment to be processed reaches the preset window duration, the current voiceprint vector is determined to be a full-window voiceprint vector or a short-window voiceprint vector. When the current voiceprint vector is a short-window voiceprint vector and the full-window count is zero, the speaker identity matching is performed using the current voiceprint vector as a temporary identification vector, and the continuous accumulation state is not updated based on the current voiceprint vector. When the current voiceprint vector is a full-window voiceprint vector, add the current voiceprint vector to the cumulative amount, update the full-window count, and generate a cumulative recognition vector based on the updated cumulative amount and full-window count; When the full window count is greater than zero, the speaker recognition result is generated by matching the cumulative recognition vector with the registered voiceprint template. Output the speaker recognition result for the current processing cycle.

2. The streaming voiceprint recognition method according to claim 1, characterized in that, The speech segment to be processed is a scrolling window speech segment with the current processing time as the end time and a length not exceeding the preset window duration. The current voiceprint vector is a fixed-length vector extracted from the speech segment to be processed by the speaker embedding extraction model and then normalized.

3. The streaming voiceprint recognition method according to claim 1, characterized in that, The actual audio duration is the effective audio duration obtained after gating the speech segment to be processed with mute; When the effective speech duration does not reach the preset window duration, the current voiceprint vector is determined as a short-window voiceprint vector, and the update weight of the short-window voiceprint vector relative to the cumulative amount is set to zero.

4. The streaming voiceprint recognition method according to claim 1, characterized in that, The cumulative recognition vector is generated based on the updated cumulative amount and full-window count, including: The average vector of the full-window voiceprint vector is generated based on the cumulative amount and the full-window count, and the average vector is normalized to obtain the cumulative recognition vector.

5. The streaming voiceprint recognition method according to claim 1, characterized in that, When the current voiceprint vector is a short-window voiceprint vector and the full-window count is greater than zero, the continuous accumulation state is not updated based on the current voiceprint vector, and the generated accumulated recognition vector is matched with the registered voiceprint template.

6. The streaming voiceprint recognition method according to claim 1, characterized in that, Matching based on the recognition vector and the registered voiceprint template includes: Calculate the cosine similarity between the recognition vector and each registered voiceprint template, determine the first similarity, the first candidate identity, and the second similarity according to the cosine similarity, and determine the interval value based on the difference between the first similarity and the second similarity.

7. The streaming voiceprint recognition method according to claim 6, characterized in that, The generated speaker recognition results include: When the first similarity reaches the similarity threshold, the interval value reaches the interval threshold, and the confidence level reaches the confidence threshold, the first candidate identity is output. When any one of the first similarity, the interval value, and the confidence level fails to reach the corresponding threshold, a preset user identifier indicating no trustworthy identity and the corresponding reason for the status are output.

8. The streaming voiceprint recognition method according to claim 7, characterized in that, When the number of registered voiceprint templates is one, the interval value is determined by the difference between the first similarity and the similarity threshold; The state reason includes multiple people overlapping when the similarity of at least two candidate identities reaches the overlap detection threshold.

9. The streaming voiceprint recognition method according to claim 7, characterized in that, The confidence level is determined based on the duration saturation factor, similarity factor, interval factor, and centroid modulus factor. The duration saturation factor is determined based on the effective speech duration obtained by accumulating the newly added effective speech duration in the current speech after the silence gating statistics, so as to avoid the duration being artificially increased due to overlapping scroll windows; The similarity factor is a monotonically increasing function of the difference between the first similarity and the similarity threshold, the interval factor is a monotonically increasing function of the interval value, and the centroid modulus factor is determined based on the modulus of the average vector of the full-window voiceprint vector before normalization.

10. The streaming voiceprint recognition method according to any one of claims 1 to 9, characterized in that, The continuously accumulating state has a lifecycle based on the discourse; In response to the speech start signal sent by the service side, the accumulated amount is cleared, the full window count is set to zero, and the round number is updated; When the duration of no valid voice messages exceeds the minimum silence buffer time, the continuously accumulated state is automatically cleared.

11. The streaming voiceprint recognition method according to claim 10, characterized in that, The output of the speaker recognition result for the current processing cycle includes the output consisting of round number, user identifier, confidence score, and status; At least one of the similarity threshold, interval threshold, confidence threshold, preset window duration, preset processing cycle, and silent fallback duration supports either a process-level default configuration or a session-level overriding configuration.

12. The streaming voiceprint recognition method according to claim 10, characterized in that, Based on the decrease in the centroid magnitude of the cumulative identification vector relative to historical values, or based on the inconsistency between the immediate candidate identity corresponding to the temporary identification vector and the cumulative candidate identity corresponding to the cumulative identification vector in multiple consecutive processing cycles, the speaker change detection in the discourse is triggered. When the speaker change detection within the utterance is triggered, the cumulative amount is cleared, the full window count is set to zero, and the round number is updated.

13. The streaming voiceprint recognition method according to any one of claims 1 to 11, characterized in that, The mute gating is implemented using an energy threshold, a neural network speech activity detection model, or a combination of both. Furthermore, the identification path uses a first silence threshold, while the registration path uses a second silence threshold, and the first silence threshold is different from the second silence threshold.

14. The streaming voiceprint recognition method according to any one of claims 6 to 11, characterized in that, The registered voiceprint templates are stored in isolation according to tenant, user identifier, and model version; During matching, the cosine similarity is calculated within the set of registered voiceprint templates corresponding to the current tenant and the current model version.

15. An electronic device, characterized in that, The device includes a processor, a memory, and an audio access unit. The memory stores a computer program, and the audio access unit is used to receive an audio stream corresponding to the current utterance. When the computer program is executed by the processor, it implements the streaming voiceprint recognition method according to any one of claims 1 to 14.

16. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the streaming voiceprint recognition method according to any one of claims 1 to 14.

Images