Training method and system of cascade speech dialogue system based on capability guidance

Abstract

This invention provides a training method and system for a capability-guided cascaded speech dialogue system. The method includes: determining the domain-specific capabilities of the recognition and synthesis modules within the cascaded speech dialogue system; guiding routing rules between the recognition and synthesis modules based on the dominant capabilities in the capability priors; constructing or generating a teacher-student model's understanding-advantage guidance route based on the routing rules; generating training samples under different routes; using the training samples to perform supervised iterative training of the student model via online policy distillation, maintaining the optimizer and learning rate states and entering the next iteration of training to obtain the trained recognition and synthesis modules. This invention selects teacher directions based on domain capability differences, reducing low-quality pseudo-supervision; through cycle-by-cycle reconstruction and filtering, subsequent training data is updated with model checkpoints; and through policy distillation constraints, the risk of the updated module deviating from its original capabilities is reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent voice, and more particularly to a training method, system, electronic device, and storage medium for a capability-guided cascaded voice dialogue system. Background Technology

[0002] Cascaded voice dialogue systems typically break down voice interaction into three functional modules: a perception and understanding module maps acoustic observations to content representations; a decision-making module parses the content and determines the response; and a response generation module converts the response into an acoustic signal output. Simply put, these three functional modules are usually implemented using ASR (Automatic Speech Recognition), a language model, and TTS (Text-to-Speech). In practical implementation, methods include machine speech chains and their subsequent joint automatic speech recognition and speech synthesis training, semi-supervised training using unpaired speech and text, online synthesis and filtering toolchains such as SpeeChain, and modular schemes that cascade automatic speech recognition, language models, and speech synthesis to form a voice dialogue system.

[0003] These technologies utilize a near-inverse speech-text interface between automatic speech recognition and speech synthesis: for data containing only speech, the recognition module generates pseudo-text and trains the synthesis module; for data containing only text, the synthesis module generates speech and trains the recognition module, thus performing teacher-student training. Some methods generate pseudo-data all at once before training, or alternately update the two modules in a preset order; modular speech dialogue systems are mainly responsible for connecting the modules during inference, and do not dynamically arrange the training direction based on the differences in module capabilities.

[0004] In the process of realizing this invention, the inventors discovered at least the following problems in the related technology: 1. Teacher-student directions and update order are usually fixed in advance. When the strength of the two modules changes under different language, domain or acoustic conditions, the weak module may pass the error to the other module. 2. The pseudo-labels or synthesized speech constructed at one time remain unchanged in subsequent training and cannot reflect the model's ability update. 3. The lack of a unified cross-module intelligibility assessment, a screening mechanism for different routes and constraints on the original ability preservation during the update process can easily introduce low-quality supervision or lead to ability shift.

[0005] In summary, the aforementioned shortcomings are mainly caused by fixed bidirectional or alternating training strategies, static pseudo-data construction methods, and screening mechanisms that rely solely on the confidence of a single module. Existing solutions do not first compare the recognition error of the recognition module against real speech with the intelligibility error of the speech generated by the synthesis module on probes in the same domain, nor do they select "stronger modules as teachers and update their corresponding weaker modules" accordingly. Furthermore, the teacher checkpoints and screening results are not refreshed with the training cycle and typically lack two-sided distillation constraints relative to the original model. Summary of the Invention

[0006] In order to at least address the problems of error propagation during model training, the inability of one-time data to reflect model capability updates, and the easy introduction of low-quality supervision leading to capability shift in existing technologies.

[0007] In a first aspect, embodiments of the present invention provide a training method for a capability-guided cascaded voice dialogue system, comprising: Determine the prior domain division capabilities of the recognition module and the synthesis module within the cascaded voice dialogue system, and guide the routing rules between the recognition module and the synthesis module based on the dominant capabilities in the prior capabilities; Based on the routing rules, construct or generate the understanding advantage guidance route of the teacher-student model, and generate training samples under different routes; The student model is trained using the training samples under supervised iterative training with online policy distillation. After each training parameter update, a checkpoint for the next round of training is published, and the optimizer and learning rate are maintained while entering the next iteration of training to obtain the trained recognition module and synthesis module.

[0008] Secondly, embodiments of the present invention provide a training agent system for a capability-guided cascaded voice dialogue system, comprising: The advantage prior guidance module is used to determine the domain division capability prior of the recognition module and the synthesis module in the cascaded voice dialogue system, and guide the routing rules between the recognition module and the synthesis module based on the advantage capability in the capability prior. The iterative online optimization training module is used to construct or generate the understanding advantage guidance route of the teacher-student model based on the routing rules, generate training samples under different routes, and use the training samples to perform supervised iterative training of the student model with online policy distillation. After each training parameter update, the checkpoint for the next round of training is published, and the optimizer and learning rate are kept in state and the next iteration of training is entered to obtain the trained recognition module and synthesis module.

[0009] Thirdly, an electronic device is provided, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of a training method for a capability-guided cascaded voice dialogue system according to any embodiment of the present invention.

[0010] Fourthly, embodiments of the present invention provide a storage medium storing a computer program thereon, characterized in that, when the program is executed by a processor, it implements the steps of the training method for a capability-guided cascaded voice dialogue system according to any embodiment of the present invention.

[0011] Fifthly, embodiments of the present invention provide a computer program product, including a computer program / instructions, characterized in that, when the computer program / instructions are executed by a processor, they implement the steps of the training method for a capability-guided cascaded voice dialogue system according to any embodiment of the present invention.

[0012] The beneficial effects of this invention are as follows: This method selects more reliable teacher directions based on differences in domain capabilities, reducing low-quality pseudo-supervision; through cycle-by-cycle reconstruction and screening, subsequent training data is updated with model checkpoints; and the risk of the update module deviating from its original capabilities is reduced through distillation constraints based on the same strategy. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart of a training method for a capability-guided cascaded voice dialogue system provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of capability advantage assessment and scheduling routing for a training method of a capability-guided cascaded voice dialogue system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the understanding constraint supervision mode of a training method for a capability-guided cascaded voice dialogue system provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the generation supervision mode of a training method for a capability-guided cascaded voice dialogue system provided in an embodiment of the present invention; Figure 5This is a schematic diagram of the iterative online optimization of a training method for a capability-guided cascaded voice dialogue system provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the overall process of a training method for a capability-guided cascaded voice dialogue system provided in an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating the data resources and TTS prompt material sources for a training method of a capability-guided cascaded voice dialogue system provided in an embodiment of the present invention. Figure 8 This is a schematic diagram of the intelligibility-based advantage prior of a training method for a capability-guided cascaded voice dialogue system provided in an embodiment of the present invention. Figure 9 This is a schematic diagram illustrating the understanding quality of a training method for a capability-guided cascaded voice dialogue system according to an embodiment of the present invention. Figure 10 This is a schematic diagram illustrating the generation quality of a training method for a capability-guided cascaded voice dialogue system according to an embodiment of the present invention. Figure 11 This is a schematic diagram of the structure of a training agent system for a capability-guided cascaded voice dialogue system according to an embodiment of the present invention; Figure 12 This is a schematic diagram of an embodiment of an electronic device for training a capability-guided cascaded voice dialogue system, provided by an embodiment of the present invention. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] like Figure 1 The diagram shows a flowchart of a training method for a capability-guided cascaded voice dialogue system according to an embodiment of the present invention, including the following steps: S11: Determine the prior domain division capabilities of the recognition module and the synthesis module within the cascaded voice dialogue system, and guide the routing rules between the recognition module and the synthesis module based on the dominant capabilities in the prior capabilities. S12: Based on the routing rules, construct or generate the understanding advantage guidance route of the teacher-student model, and generate training samples under different routes; S13: Use the training samples to perform supervised iterative training of the student model with online policy distillation. After each training parameter update, publish the checkpoint for the next round of training, maintain the optimizer and learning rate state, and enter the next loop of iterative training to obtain the trained recognition module and synthesis module.

[0017] This method addresses the common practice in cascaded speech dialogue to improve performance. Existing techniques typically involve expanding training data, increasing model parameter size, or employing teacher-student training, pseudo-labeling, and alternating update methods (as mentioned above, involving the head and tail modules). However, these methods treat "which side to train" and "when to refresh supervision" as pre-defined engineering parameters. Simply put, cascaded speech dialogue includes: an understanding module C, a decision-making module, and a response generation module G. Without altering the core decision-making strategy, the peripheral understanding and generation modules are independently evaluated and optimized, focusing on iterative performance improvements for these two peripheral modules. These update strategies are pre-fixed, indiscriminately treating weaker modules as teachers, passing their errors to the corresponding student modules, and generating synthetic supervision samples all at once. This fixes the teacher model and selected samples; however, the modules themselves are constantly changing during optimization, making further training with this data ineffective.

[0018] This method abandons the ideas of existing technologies and rethinks the technical solution. First, it compares two heterogeneous modules using the same intelligibility index. Then, it dynamically determines the teacher-student direction according to the domain. Finally, it combines online supervision reconstruction, route-specific screening, checkpoint exchange, and original capability preservation into a closed loop to optimize cascaded voice dialogue.

[0019] Simply put, this method consists of two stages: Phase 1 involves acquiring advantageous priors: Based on reserved domain exploration samples, a unified understandability index is used to evaluate the two modules. The obtained domain-specific priors, combined with available one-sided data, determine the modules to be updated.

[0020] Phase 2 is iterative online optimization: The current teacher model constructs supervision signals, candidate samples are selected by path-specific rules, and the selected student modules are updated based on the selected data.

[0021] Thus, when the understanding module C has a performance advantage, samples containing only speech are generated into pseudo-labels by the understanding module C and used to update the generation module G; when the generation module G has a performance advantage, samples containing only content are rendered into speech by the generation module G, and after filtering, used to update the understanding module C.

[0022] For step S11, the cascaded voice dialogue system includes: an initial perception and understanding recognition module. The initial response generation synthesis module A decision-making module is located between the two interfaces. This module maintains its position throughout the dialogue strategy. The recognition module maps acoustic information y (e.g., waveform or feature sequence) to text content x; the synthesis module renders the text content x and acoustic context p into a speech acoustic waveform or acoustic feature sequence. The candidate domain set is denoted as... M represents the number of domains. This method can be used to determine the prior domain partitioning capabilities of the recognition and synthesis modules within a cascaded speech dialogue system when two interfaces possess modal duality and can be compared using a unified task-level intelligibility probe.

[0023] As one implementation method, determining the prior domain division capabilities of the recognition module and the synthesis module within the cascaded voice dialogue system, and determining the routing rules between the recognition module and the synthesis module based on the prior capabilities, includes: The recognition module's comprehension ability is evaluated using a speech one-sided data pool, and the synthesis module's generation ability is evaluated using a text one-sided data pool. The comprehension ability and the generation ability include a uniform corpus-level character error rate. Based on the understanding ability and the generation ability, the advantageous capabilities of the cascaded voice dialogue system are determined, and the routing rules for training between the recognition module and the synthesis module are determined through the advantageous capabilities.

[0024] In this embodiment, the method performs calculations based on a single-sided dataset pool: in, This is a single-sided data pool containing only speech data. A one-sided data pool of text containing only content and with a generation context.

[0025] In Phase 1, the prior capabilities of the recognition and synthesis modules are calculated by domain. Specifically, this method uses a uniform corpus-level character error rate as the capability verification content.

[0026] This method uses ASR to assess the capabilities of the comprehension module C and TTS to assess the capabilities of the generation module G. For the speech domain D, let the reserved probe set be: in: For reference text, This is the actual waveform. The input waveform for TTS. To detect the number of speech sentences in a set of words. Prompt waveform. As acoustic context: depending on the corpus, it can be either a segment of the same sentence extracted by VAD (Voice Activity Detection) or a reference sample pre-selected from the speaker pool.

[0027] For the reference text set With the model output hypothesis set The corpus-level character error rate (corpus CER) is defined as: in, , , These are the character-level replacement, deletion, and insertion error counts after text normalization specific to the corpus; This is the normalized length of the reference text.

[0028] Understanding the module-side error definition: in, To obtain from waveform The predicted transcribed text, This represents the intelligibility error on dataset D. To measure the intelligibility of the generated speech, this method first performs speech synthesis, then transcription: in, For synthesized speech, It is a frozen ASR evaluator for domain D. Output the transcribed text to the evaluator. To reduce the intelligibility error of the generated side synthesis module.

[0029] The aforementioned errors—the intelligibility error on the understanding module side and the intelligibility error on the generation-side synthesis module—are treated as capability priors and processed using the same detection method and normalization. Selection and Understanding Module They are independent of each other.

[0030] As one implementation method, the routing rules include: If the recognition module has an advantage in understanding, then the synthesis module is updated and trained. If the synthesis module has a superior generation capability, then the recognition module undergoes update training.

[0031] In this embodiment, the routing component will and Record the changes in the dominance table and apply the following deterministic rules: in, This includes areas where there is an advantage in understanding, therefore updates are performed on the synthesis module G used for generation; and This includes domains where the generation side has an advantage, therefore the understanding and recognition module C is updated. If the performance of both is equal, neither of the two unidirectional routes is activated. Figure 2 The instantiation implementation of this advantage-first acquisition mechanism in speech tasks is summarized, which realizes the routing rules between the recognition module and the synthesis module based on the advantage capabilities.

[0032] In simple terms, steps S12 and S13 include constructing supervisory signals, filtering path-specific samples, updating selected modules, and generating corresponding checkpoints. When a module is understood to have performance advantages, it is enabled. Update G; when the generated module has a performance advantage, G will... Rendered as speech, the filtered synthesized speech content sample pairs are used to update C. The process state includes capability priors, activation paths, current module checkpoints, optimizer and scheduler states, and stopping step budget. When the budget threshold is reached, the final optimized model pair is output. .

[0033] Specifically, for step S12, this method uses the construction of a teacher-student model to generate training samples for different routes.

[0034] As one implementation method, it includes: using the understanding advantage of the recognition module as the routing rule, constructing a teacher-student model understanding advantage guidance route when performing update training on the synthesis module, using the recognition module as the teacher model, using the synthesis module as the student model, and using the teacher model to generate online pseudo-text of real speech; A first judgment is made based on the confidence level of the online pseudo-text; The online pseudo-text is input into the student model to generate diagnostic speech, and then the teacher model uses loop tracing of the diagnostic speech to determine the loop error. A second judgment is made based on the loop error. When both the first and second judgments are true, the online pseudo-text and the real speech are used as training samples.

[0035] In this embodiment, if the rule is based on the understanding advantage of the recognition module, when updating the synthesis module, a guidance path for the understanding advantage of the teacher-student model is constructed, and an understanding-supervised mode is implemented. In this domain, the route sets the recognition module C as an ASR (Automatic Speech Recognition) teacher model and the generation module G as a TTS (Text-to-Speech) student model. For real waveforms... The parameters are initial module Generate cached online pseudo-text ,in, Represents the number of valid non-special text tokens. Figure 3 To summarize the routing process: diagnostic synthesis is used to verify the consistency of pseudo-text, while SFT (Supervised Fine-Tuning) uses the original waveform as the acoustic target.

[0036] The posterior probability of token n is: And the confidence level of the cache sequence is: in, Cache token The posterior probability, The token confidence threshold. The proportion of high-confidence tokens. For indicator functions. When If the sample passes the validation, the first judgment is made based on the confidence level of the online pseudo-text. The threshold for accepting sequences.

[0037] For sample lines that have been allowed through the cache confidence rules, the established understanding teacher model generates online pseudo-text: in, This is the established ASR teacher model; after the teacher model is updated, This may change. The currently generated student model then outputs a diagnostic waveform, which is then passed to the teacher model for further interpretation and recognition. in, For the current TTS student model, It is obtained from the VAD or the speaker's prompt waveform. To diagnose speech, This is a cyclic transcription of text.

[0038] Furthermore, the loop closure error was determined by transcribing the text in a loop: in, This represents the character-level Levinstein edit distance between normalized strings a and b. At that time, the sample row passes the second judgment check for loop closure error. This is the closure error threshold.

[0039] The generated training sample row will only be accepted if both of these conditions are met: Among them, online pseudo-text and real speech are used as training samples to diagnose waveforms. Used only for sample screening. The qualified sample set obtained after screening. It is then fed into the generation module optimizer.

[0040] Supervised training for understanding the advantage-guided path step S13 utilizes raw speech tokens. The generation module learns based on real waveforms: in, The loss is adjusted by supervision on the generation side.

[0041] As another implementation method, it includes: the routing rule is the generation capability advantage of the synthesis module; when performing update training on the recognition module, a generation advantage guidance route of the teacher-student model is constructed, the synthesis module is used as the teacher model, the recognition module is used as the student model, the teacher model is used to generate synthesized speech of the source text, and the first judgment of duration gating is performed. The synthesized speech is input into the student model to generate synthesized text, and a second judgment on the content error is made based on the source text and the synthesized text. A third judgment on the readability of the synthesized speech is made; When the first judgment, the second judgment, and the third judgment are all true, the source text and the synthesized speech are used as training samples.

[0042] In this embodiment, if the generative advantage of the synthesis module is used as a rule to update the training of the recognition module, a generative advantage guidance path is constructed for the teacher-student model to perform generative supervision mode. For In a specific domain, this link uses module G as the TTS teacher model and module C as the ASR student model. Before online synthesis, each source text... With prompt waveform Apply duration gating. Figure 4 This paper summarizes how the link converts text-only sample lines into speech and filters synthesized speech-text paired data for training in the understanding task. Definitions: In the formula, This represents the normalized number of characters in the source text. This is the initial diagnostic synthesized waveform; Output duration, in seconds; The number of seconds occupied by a single character.

[0043] Hint conditions It is the acoustic context corresponding to the j-th sample line, obtained from VAD or taken from the speaker pool. When At that time, this sample passes the first judgment check of the duration gating. This is the time-proportion threshold.

[0044] For each sample line that passes the duration rule validation, the selected teacher-generator model generates speech in real time, and then the student model for the current comprehension task decodes the speech: in, For the selected TTS teacher model, The generated synthetic waveform, For the current ASR student model, The recognition hypothesis results output by the student model. The content error between the two is: same, To normalize the character-level Levenstein edit distance between strings a and b, when... At that time, the sample row passed the second judgment check for content error. This is the error threshold for the student model.

[0045] This method uses the published ReadScore discrimination rule, and its output is denoted as: in, The selected autoregressive generative model is used as a scorer for the synthesized waveform. Target text and identification hypothesis The calculated no-reference difference value. When The sample line is verified through a third-party readability check. Set the ReadScore reception threshold.

[0046] The synthesized waveform will only be accepted if all three conditions are met.

[0047] For the supervised training of step S13 of generating the advantage-guided route, the source text and synthesized speech are used as training samples. The dataset obtained after screening is defined with the following training objective for the understanding task: in, To understand the supervised fine-tuning loss on the side.

[0048] The above steps illustrate the training for understanding and generating dominant guidance paths. However, this method does not train only once, nor is it a mindless repetitive iteration. Instead, after each parameter update following selection, the optimizer and learning rate states are continuously retained to determine the latest relevant model checkpoints and whether the stopping condition is met. The construction of supervision signals and the selection of model-dependent samples both utilize the newly updated teacher model; therefore, the supervision distribution generated in subsequent iterations will differ from that in earlier iterations. Figure 5 The complete loop process is shown: the upper path updates the recognition module C; the reverse lower path uses the updated recognition module C to update the generation module G. There are no cross-exchange arrows in the diagram.

[0049] As one implementation method, the supervised iterative training of the student model using the training samples for online policy distillation includes: Determine the supervised loss of the student model. If the supervised loss does not meet the expectations, proceed to the next iteration of training. If the supervised loss meets the expectations, obtain the trained recognition module and synthesis module.

[0050] In this embodiment, the above describes the training iteration steps, primarily focused on understanding or generating advantageous guided routes. The following explanation focuses on iterative online optimization. After the routing rules are completed, the second stage executes an online loop: determining the roles of the teacher and student models, constructing supervision signals using corresponding modules, employing route-specific selection predicates, and updating the selected student model. The following provides instantiation definitions for the above operations in autoregressive speech recognition (ASR) and text-to-speech (TTS) scenarios. Let... Let y be the text token sequence, y be the waveform, and p be the prompt waveform. This refers to the discrete speech token sequence output by the speech segmenter q. Their probabilistic decomposition form is: in, , These are the trainable parameter subsets for the understanding module and the generation module, respectively. , This is the autoregressive prefix for both.

[0051] When any autoregressive module is selected to perform an update, this method introduces OPD (On-Policy Distillation, or distillation based on the current model-generated trajectory) as a retention constraint. Let... This represents the modules that need to be updated. For conditional inputs from its reserved domain, and The output sequence trajectory obtained from sampling the current module. For each prefix segment generated by the student model. The parameters are The method freezes the probability distribution of the original module and the current module, respectively, for the output of the next token. The objective of this method is to minimize the following: in, The length of the sampling trajectory; Represents the parameters of the original teacher model; The KL divergence transmitted from teacher to student; The primary supervised loss for the selected module; For deviation constraint weights. When At that time, the trajectory is a content token sequence generated based on acoustic observation conditions; when At this point, the trajectory is a sequence of speech tokens generated based on the text content and corresponding acoustic context conditions. In the experiments of this paper, both modules use this loss form. After the update is completed, the module re-enters the online loop process; when the preset stopping iteration budget is reached, the final optimized model pair of the recognition module and the synthesis module after training is output.

[0052] Finally, the training method of the capability-guided cascaded voice dialogue system proposed in this application is summarized in a comprehensive manner, such as... Figure 6As shown, the process involves the following steps: Step 1: Prepare reserved pairing probes for each domain, as well as one-sided data pools for speech and text; Step 2: Evaluate recognition and synthesis intelligibility using a uniform character error rate caliber and determine the dominant module; Step 3: Select the "recognition teacher to synthesis student" route based on the understanding advantage guidance route or generate the "synthesis teacher to recognition student" route based on the generated advantage guidance route; Step 4: Generate pseudo-text or synthesized speech online using the current teacher and perform route-specific filtering; Step 5: Update the student with the filtered data and add the same strategy distillation to maintain constraints; Step 6: Publish new checkpoints, maintain the optimizer and learning rate states, and enter the next loop, outputting the optimized module pair after reaching the step budget. Specifically, the understanding advantage guidance route retains the real waveform as the training target for the synthesis module, only refreshing the pseudo-text and sample selection; the generated advantage guidance route refreshes the complete synthesized waveform and uses the source text as the recognition supervision label. Therefore, the supervision used in subsequent loops can be updated as the module capabilities change.

[0053] As can be seen from this implementation method, this method selects more reliable teacher directions based on differences in domain capabilities, reducing low-quality pseudo-supervision; through cycle-by-cycle reconstruction and screening, subsequent training data is updated with model checkpoints; and through distillation constraints based on the same strategy, the risk of the update module deviating from the original capability is reduced.

[0054] Experiments are conducted to illustrate this method. Regarding the model and the unimodal dataset, the initial modules are Whisperlarge v3 and the weights of the official CosyVoice2 model, respectively. Figure 7 The diagram summarizes all datasets used for training, capability detection, and final evaluation. The training set uses only text data from RAMC, and speech data from CV-JA and JVS. For the Japanese Common Voice dataset: training uses the CV-JA validation set, removing all speech segments from the officially defined development and test sets; the test set uses CV-JA test GE5, where GE5 represents speech samples longer than 5 seconds. In the JVS dataset, jvs001-jvs100 are speaker numbers; jvs001-jvs090 are used as the training set; the subset jvs091-jvs100, with no overlap between speakers, is reserved as the first-stage detection set. The first-stage detection of RAMC uses conversation segments from the official development set, which have no overlap with the official test conversation samples used in the final evaluation.

[0055] Each iteration starts with 512 original samples. When updating the generation module G, the token confidence threshold is... Sequence confidence threshold Round trip verification threshold When updating the understanding module C, the duration threshold is... s / character (0.8 seconds per character), student model error threshold The internal difference parameter of ReadScore is set to 0.1, and the external receive threshold is... Whisper is trained using BF16 precision, an 8-way FSDP distributed strategy, and an effective batch size of 64. The peak learning rate is set to... The preheating steps are 20; after preheating, a multiplicative decay coefficient of 0.995 is used, and the total number of optimizer iterations is 200. Supervised fine-tuning of CosyVoice2 only updates the autoregressive large language model part, while the streaming module and vocoder remain frozen. This large language model uses FP16 accuracy, 8-way DeepSpeed ​​Stage 2, and a maximum dynamic batch processing limit of 480 frames; the peak learning rate is... The warm-up steps were 60, the decay coefficient was 0.998, and the total number of optimizer steps was 600. In the online iterative experiments, the optimizer and learning rate scheduler states were continuously retained across loops without re-executing the warm-up. For each module, both the online iterative experimental group and the corresponding single static benchmark experimental group started from the same official initial weights and used the exact same optimization step budget and learning rate scheduling strategy. The static benchmark experimental group only performed a global shuffle on the pre-built Chinese-Japanese bilingual supervised dataset once; it did not regenerate or select supervised samples in multiple loops.

[0056] This method employs bidirectional OPD constraints. When updating the understanding module C, the frozen official Whisper is used as the teacher model; each loop uses 64 Japanese audio samples and 4 insertion events, with bias constraint weights. A greedy online strategy is used to generate a maximum of 256 new text tokens. When updating the generation module G, the frozen official CosyVoice2 is used as the teacher model; each round uses 64 Mandarin texts and 8 insertion events, with bias constraint weights. .

[0057] Regarding the dual model exchange mechanism: the first loop of both modules directly uses the official original model. Subsequently, in the initial iteration steps... , At this point, branch G uses an iteration step of Whisper weights; the C branch uses an iteration step of CosyVoice2 weights.

[0058] The final evaluation metrics for TTS include: character error rate (CER) calculated by Qwen3 ASR 1.7B, speaker cosine similarity, and UTMOS; the first-stage routing decision only uses the intelligibility character error rate (CER).

[0059] like Figure 8 The diagram shows a set of first-stage detection results for each activation mode. On the Japanese JVS detection set, Whisper large v3 yields results on real speech waveforms. =5.94%; Generate side evaluator Using Qwen3 ASR1.7B, the speech synthesized by CosyVoice2 was obtained. =12.09%. At this point, the understanding module has a performance advantage, and the routing rule selection update generation module is used.

[0060] On the official RAMC development probe set with completely non-overlapping samples, the generation side achieves a lower CER, and the routing rule selection update understanding module is updated.

[0061] The online iterative process terminated after Whisper completed 200 optimization steps and 37 data loops; and after CosyVoice2 completed 600 optimization steps and 14 data loops. For example... Figure 9 The performance of the understanding module on the final test set is presented. Compared with the original weights in the official dataset, the optimized Whisper reduced the character error rate (CER) from 15.23% to 15.10% on the CV-JA dataset and from 14.18% to 10.77% on the RAMC dataset. While maintaining the same number of optimization steps and learning rate scheduling strategy, the online optimization scheme also achieved improvements compared to a single static baseline construction: CV-JA (18.75% → 15.10%) and RAMC (11.63% → 10.77%). This comparison is at the system level: teacher model updates, online sample selection, and retention of constraint data all affect the supervision signal. This result demonstrates the benefits of the entire online closed-loop scheme, rather than just the effect of simple iterative iteration. Figure 9 , Figure 10 The marker “(OPD)” indicates that only the OPD constraint branch uses the reserved constraint data source.

[0062] Figure 10 The performance of the generative modules on the final test set is presented. Compared with the official original weights, after online iterative optimization, the CER of CosyVoice2 for CV-JA decreased from 14.85% to 11.18%, and the CER of RAMC decreased from 4.78% to 4.04%. However, compared with the static baseline with matching steps, the performance of iterative online optimization on CV-JA is close but slightly worse (11.18% vs. 11.03%), while the difference is more significant on RAMC (4.04% vs. 1.93%). Speaker cosine similarity and UTMOS metrics are basically the same as the static baseline, indicating that the performance gap on RAMC is mainly reflected in the intelligibility CER metric, rather than a serious degradation of the generative model as a whole. Therefore, the comparison results between the online solution and the static baseline are strongly correlated with the specific modules.

[0063] In summary, this method proposes a two-stage capability-guided iterative online optimization process for the peripheral understanding and generation modules of a cascaded spoken dialogue system. First, capability priors are obtained through domain-specific intelligibility detection, and then one of two options is chosen: understanding-supervised optimization based on pure speech data or generation-supervised optimization based on pure text content data. In the generation-supervised branch, the generation module converts text content into speech, and selects qualified synthesized speech-text pairing data for training the understanding module. During multiple rounds of iterative iteration, the two branches construct and filter supervision signals online, exchange iterative model weights, and apply bidirectional OPD constraints. On both test sets, Whisper large v3 and CosyVoice2 achieved a reduction in character error rate (CER) compared to the official initial weights. Under the condition of perfectly aligned optimization step budget and learning rate scheduling, the iterative online version of Whisper outperforms the single static baseline on both test sets. The iterative online version of CosyVoice2 performs close to but slightly worse than the static baseline on the Japanese task, and weaker than the static baseline on the Mandarin task.

[0064] like Figure 11 The diagram shows a structural schematic of a training agent system for a capability-guided cascaded voice dialogue system according to an embodiment of the present invention. This system can execute the training method for the capability-guided cascaded voice dialogue system described in any of the above embodiments and is configured in a terminal.

[0065] This embodiment provides a training agent system 10 for a capability-guided cascaded voice dialogue system, which includes: an advantage prior guidance module 11 and an iterative online optimization training module 2.

[0066] The advantage prior guidance module 11 is used to determine the domain division capability prior of the recognition module and the synthesis module in the cascaded speech dialogue system, and guides the routing rules between the recognition module and the synthesis module based on the advantage capability in the capability prior; the iterative online optimization training module 12 is used to construct the understanding advantage guidance route of the teacher-student model or generate the advantage guidance route based on the routing rules, generate training samples under different routes, and use the training samples to perform supervised iterative training of the student model with online policy distillation. After each training parameter update, the checkpoint of the next round of training is released, and the optimizer and learning rate are kept in state and the next loop of iterative training is entered to obtain the trained recognition module and synthesis module.

[0067] This invention also provides a non-volatile computer storage medium storing computer-executable instructions that can execute the training method of the capability-guided cascaded voice dialogue system in any of the above method embodiments. In one embodiment, the non-volatile computer storage medium of the present invention stores computer-executable instructions, which are configured as follows: Determine the prior domain division capabilities of the recognition module and the synthesis module within the cascaded voice dialogue system, and guide the routing rules between the recognition module and the synthesis module based on the dominant capabilities in the prior capabilities; Based on the routing rules, construct or generate the understanding advantage guidance route of the teacher-student model, and generate training samples under different routes; The student model is trained using the training samples under supervised iterative training with online policy distillation. After each training parameter update, a checkpoint for the next round of training is published, and the optimizer and learning rate are maintained while entering the next iteration of training to obtain the trained recognition module and synthesis module.

[0068] As a non-volatile computer-readable storage medium, it can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the methods in the embodiments of the present invention. One or more program instructions are stored in the non-volatile computer-readable storage medium, and when executed by a processor, they perform the training method of the capability-guided cascaded voice dialogue system in any of the above method embodiments.

[0069] Figure 12 This is a schematic diagram of the hardware structure of an electronic device for a training method of a capability-guided cascaded voice dialogue system provided in another embodiment of this application, as shown below. Figure 12 As shown, the device includes: One or more processors 1210 and memory 1220, Figure 12 Taking a processor 1210 as an example, the device for training a capability-guided cascaded voice dialogue system may further include an input device 1230 and an output device 1240.

[0070] The processor 1210, memory 1220, input device 1230, and output device 1240 can be connected via a bus or other means. Figure 12 Taking the example of a connection between China and Israel via a bus.

[0071] The memory 1220, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the training method of the capability-guided cascaded voice dialogue system in the embodiments of this application. The processor 1210 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions, and modules stored in the memory 1220, thereby implementing the capability-guided cascaded voice dialogue system training method described in the above embodiments.

[0072] The memory 1220 may include a program storage area and a data storage area, wherein the program storage area may store the operating system and applications required for at least one function; the data storage area may store data, etc. Furthermore, the memory 1220 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the memory 1220 may optionally include memory remotely located relative to the processor 1210, and these remote memories may be connected to the mobile device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0073] Input device 1230 can receive input numerical or character information. Output device 1240 may include display devices such as a display screen.

[0074] The one or more modules are stored in the memory 1220, and when executed by the one or more processors 1210, they execute the training method of the capability-guided cascaded voice dialogue system in any of the above method embodiments.

[0075] The above-described product can perform the methods provided in the embodiments of this application, and has the corresponding functional modules and beneficial effects for performing the methods. Technical details not described in detail in this embodiment can be found in the methods provided in the embodiments of this application.

[0076] Non-volatile computer-readable storage media may include a stored program area and a stored data area, wherein the stored program area may store an operating system and an application program required for at least one function; the stored data area may store data created based on the use of the device, etc. Furthermore, the non-volatile computer-readable storage medium may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, the non-volatile computer-readable storage medium may optionally include memory remotely located relative to the processor, and these remote memories may be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0077] This invention also provides an electronic device comprising: at least one processor and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the steps of the training method of the capability-guided cascaded voice dialogue system according to any embodiment of this invention.

[0078] The electronic devices described in this application exist in various forms, including but not limited to: (1) Mobile communication devices: These devices are characterized by their mobile communication capabilities and primarily aim to provide voice and data communication. These terminals include smartphones, multimedia phones, feature phones, and low-end phones.

[0079] (2) Ultra-mobile personal computer devices: These devices fall under the category of personal computers, possessing computing and processing capabilities, and generally also have mobile internet access features. These terminals include PDAs, MIDs, and UMPCs, such as tablet computers.

[0080] (3) Portable entertainment devices: These devices can display and play multimedia content. This category includes audio and video players, handheld game consoles, e-book readers, as well as smart toys and portable car navigation devices.

[0081] (4) Other electronic devices with data processing functions.

[0082] In this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, without necessarily requiring or implying any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising" or "including" include not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0083] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0084] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A training method for a capability-guided cascading voice dialogue system, comprising: Determine the prior domain division capabilities of the recognition module and the synthesis module within the cascaded voice dialogue system, and guide the routing rules between the recognition module and the synthesis module based on the dominant capabilities in the prior capabilities; Based on the routing rules, construct or generate the understanding advantage guidance route of the teacher-student model, and generate training samples under different routes; The student model is trained using the training samples under supervised iterative training with online policy distillation. After each training parameter update, a checkpoint for the next round of training is published, and the optimizer and learning rate are maintained while entering the next iteration of training to obtain the trained recognition module and synthesis module.

2. The method according to claim 1, wherein, The determination of the domain division capability prior of the recognition module and the synthesis module within the cascaded voice dialogue system, and the determination of the routing rules between the recognition module and the synthesis module based on the capability prior, include: The recognition module's comprehension ability is evaluated using a speech one-sided data pool, and the synthesis module's generation ability is evaluated using a text one-sided data pool. The comprehension ability and the generation ability include a uniform corpus-level character error rate. Based on the understanding ability and the generation ability, the advantageous capabilities of the cascaded voice dialogue system are determined, and the routing rules for training between the recognition module and the synthesis module are determined through the advantageous capabilities.

3. The method according to claim 2, wherein, The routing rules include: If the recognition module has an advantage in understanding, then the synthesis module is updated and trained. If the synthesis module has a superior generation capability, then the recognition module undergoes update training.

4. The method according to claim 3, wherein, The step of constructing or generating advantageous guidance routes for the teacher-student model based on the routing rules, and generating training samples under different routes, includes: The routing rule leverages the understanding capabilities of the recognition module. When updating and training the synthesis module, a route is constructed based on the understanding advantages of the teacher-student model. The recognition module is used as the teacher model, and the synthesis module is used as the student model. The teacher model is then used to generate online pseudo-text of real speech. A first judgment is made based on the confidence level of the online pseudo-text; The online pseudo-text is input into the student model to generate diagnostic speech, and then the teacher model uses loop tracing of the diagnostic speech to determine the loop error. A second judgment is made based on the loop error. When both the first and second judgments are true, the online pseudo-text and the real speech are used as training samples.

5. The method according to claim 3, wherein, The step of constructing or generating advantageous guidance routes for the teacher-student model based on the routing rules, and generating training samples under different routes, further includes: The routing rule leverages the generation capabilities of the synthesis module. When updating and training the recognition module, a generation advantage guidance route is constructed based on the teacher-student model. The synthesis module is used as the teacher model, and the recognition module is used as the student model. The teacher model is used to generate synthesized speech from the source text, and the first judgment of duration gating is performed. The synthesized speech is input into the student model to generate synthesized text, and a second judgment on the content error is made based on the source text and the synthesized text. A third judgment on the readability of the synthesized speech is made; When the first judgment, the second judgment, and the third judgment are all true, the source text and the synthesized speech are used as training samples.

6. The method according to claim 1, wherein, The supervised iterative training of the student model using the training samples for online policy distillation includes: Determine the supervised loss of the student model. If the supervised loss does not meet the expectations, proceed to the next iteration of training. If the supervised loss meets the expectations, obtain the trained recognition module and synthesis module.

7. A training system for a capability-guided cascading voice dialogue system, comprising: The advantage prior guidance module is used to determine the domain division capability prior of the recognition module and the synthesis module in the cascaded voice dialogue system, and guide the routing rules between the recognition module and the synthesis module based on the advantage capability in the capability prior. The iterative online optimization training module is used to construct or generate the understanding advantage guidance route of the teacher-student model based on the routing rules, generate training samples under different routes, and use the training samples to perform supervised iterative training of the student model with online policy distillation. After each training parameter update, the checkpoint for the next round of training is published, and the optimizer and learning rate are kept in state and the next iteration of training is entered to obtain the trained recognition module and synthesis module.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method described in any one of claims 1-6.

9. An electronic device comprising: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the steps of the method according to any one of claims 1-6.

10. A storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1-6.

Images