Intelligent evaluation method for speech rehabilitation training effect of children
By presenting training and generalization probe terms in a mixed manner, and combining forced alignment and output scoring, the problem of inaccurate training effect evaluation in existing technologies is solved, and reliable evaluation and effective scheduling of training effect are achieved.
Patent Information
- Application Number
- CN202611586567.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-14
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies cannot effectively distinguish the effects of training on children's speech rehabilitation from the influence of other factors, resulting in unreliable training schedules and an inability to accurately assess the actual benefits of training.
An intelligent evaluation method is adopted, which presents training items and generalization probes in a mixed manner, applies prompts step by step, and combines forced alignment and output scoring to separate the influence of training effect on other factors and generate task configuration parameters for the next training cycle.
It achieves the separation of training effect and practice effect, the correction of natural development and daily state, ensures that training scheduling is based on reliable gain estimation, and improves the accuracy and reliability of training effect evaluation.
Smart Images

Figure CN123058675A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of speech rehabilitation, and in particular to an intelligent assessment method for the effectiveness of speech rehabilitation training in children. Background Technology
[0002] Rehabilitation training for children with speech disorders is conducted in courses. At the end of each training cycle, the therapist needs to determine whether the current target phoneme has been mastered, whether progress can be made to the next phoneme, and whether the training vocabulary needs to be changed. Training resources and children's attention are both limited, and the number of tasks that can be presented in a single session is not large. Every move forward or pause in the scheduling decision determines the training benefits for several subsequent cycles.
[0003] Currently, the assessment of training effectiveness mainly relies on three methods. First, therapists administer standardized scales periodically and manually score the tests; the scores are used to complete training records and adjust treatment plans. Second, automatic speech processing techniques, such as forced alignment, are used to score the degree of match between the child's spoken language and the target text, thus replacing some of the manual listening and discrimination work. Third, performance is age-normalized based on developmental norms for different age groups to determine the child's current level relative to their peers. These methods are feasible for obtaining scores from a single test.
[0004] The effects of training cannot be directly observed; what can be observed is the change in performance over time. This change is the result of multiple factors, and current methods cannot separate the portion truly attributable to training. Because the vocabulary used for assessment and the vocabulary used for training share the same origin, an increase in assessment scores may come from improved phonemic ability or simply from repeated practice and memorization of that set of vocabulary. Because the acoustic model upon which automatic scoring relies is continuously updated to adapt to children's pronunciation, the scoring scale itself shifts with the training process; the more familiar the model is with the child, the higher the score for the same output. Because developmental norms are taken from external population samples, the natural developmental amount subtracted by age interpolation does not match the child's own developmental rate; the training gains of slow-developing individuals are underestimated, while those of fast-developing individuals are overestimated. Because the performance in a single conversation is also affected by the level of cooperation on that day and physiological conditions such as nasal congestion and lack of sleep, a decline in performance may indicate a decline in ability or simply a fluctuation in the child's condition on that day. These sources are in the same direction and shape as the training effect, making the gain estimates used to advance the scheduling unreliable. Measurement errors are treated as progress or regression, and training stops at the wrong phonemes or moves forward prematurely. Summary of the Invention
[0005] In order to separate the training effect from the practice effect, the shift in the rating scale, the natural development and the changes in the state of the day, and to make the advancement criteria of the training schedule based on reliable gain estimates, this application provides an intelligent assessment method for the effect of speech rehabilitation training in children.
[0006] This application provides an intelligent assessment method for the effectiveness of speech rehabilitation training in children, which adopts the following technical solution: An intelligent assessment method for the effectiveness of speech rehabilitation training in children includes: S1. Determine a set of task items including training items and generalization probes, wherein the generalization probes are words containing target phonemes that have not been used for training, and each of the generalization probes is presented at most once; S2. The training items and the generalization probe items are presented in the same presentation format. The speech signals produced by the trainee for each task item in the task item set are collected. The training items that are not produced correctly are prompted level by level along the prompting level sequence. The prompting level sequence includes the lowest prompting level that represents no prompting. The prompting level at which the training item is first produced correctly is recorded. S3. Force alignment is performed on the speech signal based on the target text of each task item to obtain the output score; S4. Determine the validity weight of each output score, and determine the change of the output score statistics after weighting the training item and the generalization probe item relative to the previous training cycle as the first gain and the second gain, respectively. S5. Determine the generalization parameter based on the first gain and the second gain, and generate training task configuration parameters including the target phoneme, training vocabulary, and cue start level for the next training cycle based at least on the generalization parameter and the recorded cue level, wherein the words in the training vocabulary are used as training items in the next training cycle.
[0007] Optionally, in S4, a natural developmental gain is subtracted when determining the first gain and the second gain. The natural developmental gain is determined based on the trainee's natural developmental baseline and the time interval between two adjacent training cycles. The natural developmental baseline characterizes the natural rate of change of the trainee's output score with the developmental process.
[0008] Optionally, the task item set further includes prospective detection items, which are words selected from the words corresponding to phonemes downstream of the target phoneme and not yet included in the training on the phoneme learning order constraint graph. The phoneme learning order constraint graph uses phonemes as nodes and directed edges based on the learning order dependencies between phonemes. In S2, the prospective detection items are also presented periodically at the lowest prompt level without providing correctness feedback. In each training session, the number of presentations of the prospective detection items is lower than the number of presentations of the generalization detection items. The number of times the same prospective probe is presented is at most a preset number of times before the corresponding phoneme enters training; in S4, the natural developmental baseline is determined based on the longitudinal change of the output score of each prospective probe; in S5, in response to the increase in the output score of any prospective probe exceeding the expected range determined based on the natural developmental baseline, the corresponding phoneme is marked as spontaneously learned on the phoneme learning sequence constraint graph, and the phonemes marked as spontaneously learned are skipped when generating the training task configuration parameters.
[0009] Optionally, S3 includes sub-steps S31-S33: S31. The output score of each task item is determined by an adaptive acoustic model, wherein the adaptive acoustic model performs speaker adaptive update based on the speech signal of the training item, and the speech signal of the generalized probe item is used only to determine the output score. S32. Store the collected speech signals, and periodically re-evaluate the stored historical speech signals while keeping the model version and feature extraction configuration of the adapted acoustic model unchanged. S33. The difference between two scores of the same historical speech signal is determined as the score drift. In step S4, the scoring drift is also deducted when determining the first gain and the second gain.
[0010] Optionally, the task set further includes a retention probe, which is a historical training item that has stopped training for a preset retention interval; in S2, the retention probe is presented at the lowest prompt level; in S4, the retention degree parameter of the phoneme corresponding to the retention probe is determined based on the output score of the retention probe and the output score when the retention probe stops training.
[0011] Optionally, step S5 includes sub-steps S51-S52: S51. Determine the comparison results of the generalization parameter and the preset generalization threshold, the comparison results of the retention parameter and the preset retention threshold, and the comparison results of the prompt level at which the training item is first correctly produced and the preset level threshold. S52. Based on the comparison results, determine the target phoneme for the next training cycle along the topological order of the phoneme learning order constraint graph, wherein the phoneme learning order constraint graph has phonemes as nodes and directed edges with the learning order dependency between phonemes.
[0012] Optionally, each training cycle includes at least one training session; the task item set also includes anchor items, which are words selected from historical training items that have stopped training and meet the anchoring conditions, the anchoring conditions including the retention parameter reaching a preset anchoring threshold and the syllable duration variation coefficient being lower than a preset variation threshold; in S2, the anchor items are presented before the training items and the generalization probe items in each training session; in S4, the median deviation of the output score of each anchor item from the anchoring baseline is determined as the session state offset, the anchoring baseline is determined based on the historical output scores of each anchor item, and the validity weight decreases as the session state offset increases.
[0013] Optionally, the intelligent evaluation method for the effect of speech rehabilitation training for children further includes: determining a minimum detectable change threshold based on the measurement standard error determined by weighting each output score according to the validity weight; in response to the absolute values of both the first gain and the second gain being lower than the minimum detectable change threshold, determining the target phoneme as having no significant change detected, postponing the evaluation of the target phoneme to the next training cycle, wherein in step S5, the training task configuration parameters for the next training cycle are generated to be the same as the current training task configuration parameters, and the presentation number of the generalized probe corresponding to the target phoneme in the next training cycle is reduced; in response to the absolute value of the first gain reaching the minimum detectable change threshold, restoring the presentation number of the generalized probe.
[0014] Optionally, in step S2, image signals of the trainee are also collected; in step S4, the posterior probability of the trainee being in each cooperation state in the current training session is determined based on the voice signal and the image signal, the cooperation state with the highest posterior probability is determined as the cooperation state of the trainee, and the effectiveness weight is determined based on the posterior probability of the determined cooperation state.
[0015] Optionally, S2 includes sub-steps S21-S23: S21. Each task item in the task item set is presented starting from the lowest prompt level. For the training item that is not correctly produced, prompts are applied level by level along the prompt level sequence starting from the prompt starting level. The prompt starting level is not higher than the prompt level at which the training item was first correctly produced in the previous training session. S22. Determine the independence index based on the cue level at which the training item is first correctly produced. The value of the cue level in the cue level sequence increases with the increase of cue intensity. The lower the cue level at which the item is first correctly produced, the higher the independence index. S23. In response to the independence index of any training item corresponding to the target phoneme reaching a preset independence threshold, increase the proportion of the generalization probe item corresponding to the target phoneme in the mixed presentation.
[0016] In summary, this application includes at least one of the following beneficial technical effects: 1. Training items and untrained one-time generalization probes are mixed in the same session and processed through the same scoring channel, making direct training results and untrained ability transfer a pair of comparable observations. The training effect is thus separated from the practice effect on specific terms. The criteria for scheduling advancement are all derived from the data generated by the trainees themselves during the training process, without relying on external control groups or manual evaluation.
[0017] 2. Natural development is calibrated by longitudinal observation of the untrained downstream phonemes of the trainee and deducted from the gain. The shift of the scoring scale is measured by the difference between two scores of the same historical speech and deducted. The ability maintenance is given by the unaided retest after training is stopped. This allows the observed performance changes to be restored layer by layer to the part caused by training, avoiding the underestimation of the benefits of slow-developing individuals and the counting of model familiarity as an improvement.
[0018] 3. The retired training items that have been thoroughly mastered are reused as a session state reference to correct the impact of daily fluctuations. Whether the gain exceeds the measurement noise is determined by the threshold determined by the data itself and is linked to the consumption and recovery of the detection items. The degree of cooperation participates in the weighting in the form of a posteriori probability. The detection terms are paid according to the marginal information content and merged in multiple device copies with only increasing sets, so that the one-time measurement resources are sustainable throughout the entire training cycle. Moreover, all the heavy mixing is corrected by the observations generated during the system's own operation, without the need to introduce any external data assets. Attached Figure Description
[0019] Figure 1 This is a schematic diagram of the overall process of an intelligent assessment method for the effectiveness of speech rehabilitation training in children, as described in one embodiment of this application.
[0020] Figure 2 This is a schematic diagram illustrating the migration of terms between various task roles in one embodiment of this application.
[0021] Figure 3 This is a schematic diagram illustrating how the output score is decomposed into a first gain and a second gain and then corrected progressively in one embodiment of this application. Detailed Implementation
[0022] The present application will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the application and are not intended to limit the scope of the application.
[0023] This application discloses an intelligent evaluation method for the effectiveness of speech rehabilitation training in children.
[0024] This method decomposes observed performance changes into a pair of comparable gains by presenting trained and untrained one-off terms in the same training session in a mixed manner, thus separating the training effect from the practice effect on specific terms. First, a set of task items containing training and generalization probes is determined. Then, these items are presented in a mixed manner with cues applied hierarchically along the cue sequence to collect speech. Next, forced alignment is performed based on the target text to obtain output scores. Then, validity weights are determined, and the first and second gains are calculated. Finally, training task configuration parameters for the next training cycle are generated based on the two gains. Therefore, without setting an external control group or relying on manual rating scales, the progression of training scheduling is based on reliable gain estimation.
[0025] To better understand the necessity of the above approach, let's first demonstrate the failure process of existing methods using a specific scenario. A five-year-old child is trained with alveolar fricatives as the target phoneme. The training vocabulary includes words such as "umbrella," "three," and "sour." Each training cycle consists of one week, with two training sessions per cycle. At the fourth week's review, the child's output score on these words improved from 62.0 to 68.0. If this change were used as the sole criterion, the assessment would conclude that the training was effective and could proceed to the next phoneme. However, the same set of words has been presented dozens of times over four weeks. The child may have only memorized the movement patterns of these words, while still failing to pronounce any words containing the target phoneme outside the training vocabulary. Promoting a new phoneme at this point would lead to repeated failures due to a lack of foundational knowledge; conversely, if the training is deemed ineffective and regressed, the real progress already made would be erased. Existing methods fail to distinguish between these two scenarios because they rely solely on trained terms as a source of observation, and performance on this source simultaneously includes two inseparable components: improved skills and familiarity with the terms.
[0026] Reference Figure 1 The following sections will provide a detailed explanation of the terminology, the determination of the task set, the mixed presentation and prompt hierarchy recording, the acquisition of output scores, the calculation of gain and successive correction, and the generation of training task configuration parameters.
[0027] The following section will first explain some of the technical terms involved in the embodiments of this application.
[0028] The target phoneme refers to the single phoneme that is the focus of training in the current training cycle. Intervention for articulation disorders in children usually does not correct all erroneous sounds at the same time, but rather proceeds one phoneme at a time in a certain order, and the phoneme at the current position of the progression is the target phoneme. In this embodiment, the target phoneme is the alveolar fricative, and the downstream phonemes related to it include the alveolar affricates and the retroflex series of phonemes.
[0029] A cue hierarchy sequence refers to a set of cues arranged from weakest to strongest external support, used to provide progressive assistance when a child cannot produce an answer independently. In this embodiment, the cue hierarchy sequence comprises four levels, with values increasing as the cue strength increases: Level 0 is the lowest cue level, representing no cueing applied, only the presentation of images or text awaiting the child's independent output; Level 1 provides semantic cue cues, offering semantic guidance related to the target word; Level 2 provides phonological cue cues, indicating the pronunciation position or key points of the target sound; and Level 3 provides a complete demonstration cue, where the trainer or system demonstrates the target word in its entirety for the child to imitate. Levels 0 to 3 form a monotonically increasing chain of support strength; the lower the level at which the child first correctly produces an answer, the less dependent they are on external support.
[0030] Output score is a quantitative score given to the accuracy of a child's single speech signal on the target sound. In this embodiment, the output score is normalized to a range of 0 to 100, where 100 points corresponds to the target phoneme being acoustically produced at a typical correct level for that age group, and 0 points corresponds to the target phoneme being missing or replaced with another phoneme. The output score is the only original observation in this method, and all subsequent gains, parameters, and decisions are derived from it.
[0031] Training cycles and training sessions are two nested time units in this method. A training session is a face-to-face or human-computer interaction training activity, representing the actual presentation of tasks and the actual collection of speech. A training cycle is the unit of evaluation and decision-making; each training cycle includes at least one training session, and gain is calculated and configuration is generated only at the end of a training cycle. In this embodiment, one week constitutes a training cycle, and each training cycle includes two training sessions, with each training session presenting approximately 30 trials. The separation of these two units ensures that occasional fluctuations in a single session do not directly trigger scheduling decisions.
[0032] The determination of the task item set is described in detail below. This method includes steps S1 to S5.
[0033] S1. Determine the set of task items, including training items and generalization probes. Generalization probes are words containing the target phoneme that were not used in training, and each generalization probe is presented at most once.
[0034] The training item refers to a lexical entry that is written into the training vocabulary, presented repeatedly in training sessions, and given prompts when a production error occurs, and is the direct object on which the intervention action acts. The generalization probe item refers to a lexical entry that also contains the target phoneme, but is never included in the training vocabulary, never receives prompts, and is presented at most once throughout the entire training process, and is a pure measurement object. The only difference between the two lies in whether they have been trained, so the difference between their performances exactly reflects the transfer amount of training on unpracticed materials, rather than the difference in presentation conditions or scoring conditions.
[0035] S1 includes sub-steps S11 to S13.
[0036] S11. Determine the target phoneme, and screen out all lexical entries containing the target phoneme from the lexicon to form a candidate lexical entry set. This sub-step establishes an index from phonemes to lexical entries, providing a common material selection range for training items and generalization probe items, so that the two types of task items come from a homogeneous lexical entry pool under the same phoneme constraint, thereby eliminating the systematic difficulty difference between the two types of lexical entries caused by different material selection ranges at the source. For example, when the alveolar fricative is taken as the target phoneme, the candidate lexical entry set includes about 120 lexical entries such as "umbrella", "three", "sour", "forest", "pine", "ear of grain", "broken" and the like.
[0037] S12. Select a part of lexical entries from the candidate lexical entry set as training items, and write them into the training vocabulary. The input of this sub-step is the candidate lexical entry set, and the output is a training item list. This list is not only the object that is presented repeatedly and given prompts step by step in subsequent sessions, but also the only range of lexical entries that children are allowed to repeatedly contact during the training period. The number of training items is determined according to the number of trials that a single session can carry, and 20 lexical entries are taken in this embodiment.
[0038] S13. Exclude the lexical entries that have been selected as training items from the candidate lexical entry set, select generalization probe items from the remaining lexical entries, and establish a single presentation mark for each generalization probe item. The single presentation mark is atomically set while the lexical entry is presented, and then the lexical entry permanently exits the to-be-presented queue. The output of this mark is queried by the presentation schedule of S2 to determine whether a certain lexical entry can still be presented, and is used by S4 to judge whether a certain observation can be included in the statistics of generalization probe items. In this embodiment, 6 generalization probe items are presented in a single session.
[0039] The constraint that any generalization probe item is presented at most once is an irreplaceable setting in this method. Once a lexical entry is presented, the child gets a practice opportunity on this lexical entry, and the performance obtained from any subsequent re-presentation contains both transfer and practice components, which is no different in nature from the training item. Therefore, single presentation is not a resource saving measure, but a necessary condition for the second gain to carry the pure transfer amount, and the one-time nature of the probe lexical entry thus becomes a scarce resource that must be specially managed by this method.
[0040] In some embodiments, the task set further includes lookahead probes. Lookahead probes are words selected from the words corresponding to phonemes downstream of the target phoneme and not yet in training, as shown in the phoneme learning order constraint graph. The phoneme learning order constraint graph uses phonemes as nodes and directed edges representing the order of phoneme acquisition.
[0041] Phoneme acquisition sequence constraint graphs depict the order in which phonemes are acquired within the same language. In the phonological development of Chinese children, plosives are generally acquired earlier than fricatives, fricatives are generally acquired earlier than affricates, and the front alveolar series is generally acquired earlier than the back alveolar series. Treating each phoneme as a node, the dependency relationship "A is usually acquired before B" is represented as a directed edge from A to B, resulting in a directed acyclic graph. Nodes downstream of a given node in the graph represent subsequent phonemes acquired after that node has been mastered. In this embodiment, when the front alveolar fricative is the target phoneme, its downstream nodes include front alveolar affricates, back alveolar fricatives, and back alveolar affricates. The graph can be constructed from publicly available phonological development norm literature or by inverting the order of phoneme acquisition based on historical case data accumulated by an institution. The graphs obtained from the two paths are usually consistent in topology, differing only in the presence or absence of individual edges.
[0042] The prospective detection items are sourced using a constraint graph: first, downstream nodes of the target phoneme are identified on the graph; then, phonemes already in training are excluded from the corresponding words of these nodes, and the remaining words are selected. Since these phonemes have never been intervened in during the current and previous training cycles, any changes in the child's performance on them can only stem from their own development; therefore, this set of observations constitutes a longitudinal baseline uncontaminated by training. The determination of whether a phoneme has not yet entered training is based on the historical records of the training vocabulary; any phoneme that has appeared in any historical training vocabulary is excluded. The constraint graph must exclude loops during construction; if statistical data generates two mutually referential edges, the earlier historical achievement point is taken as the upstream.
[0043] As an optional implementation, the task set also includes hold probes. Hold probes are historical training items that have stopped training for a preset hold interval.
[0044] The difference between retention probes and generalized probes is that the former were previously training items and have been practiced extensively, but training has been suspended for a period of time. Historical training items whose suspension period has reached a preset retention interval are retrieved and retested. Their performance reflects whether mastery is maintained after detachment from training support, rather than newly acquired abilities. In this embodiment, the preset retention interval is 4 weeks. This value ensures that the retest falls after a significant decline in the forgetting curve, sufficient to expose weak mastery, but not so long that the child has moved on to the next phoneme, thus missing the opportunity for intervention. The start point for the suspension period is the last time the term was presented as a training item. Achievement determination is performed during the task item arrangement phase of each session; achieving the goal means the term is included in the presentation queue for that session. After suspension, training items are transferred to the retention probe pool, constituting the term's first role transfer in this method.
[0045] In one alternative implementation, each training cycle includes at least one training session, and the task set further includes anchor terms. Anchor terms are selected from historical training items that have stopped training and meet anchoring criteria. Anchoring criteria include a retention parameter reaching a preset anchoring threshold and a syllable duration coefficient of variation below a preset variation threshold.
[0046] The purpose of anchoring is not to measure ability, but to measure the child's state on that day. When a child is drowsy, physically unwell, or emotionally resistant, even if their ability remains unchanged, their performance on all tasks will decline in the same direction; if this component is not measured separately, the low score for the day will be counted as invalid training. Retired training items that have been deeply mastered are suitable for this reference, because the child's ability on these items is stable, and the remaining fluctuations in their performance are almost entirely due to their state. The two anchoring conditions select such items from two perspectives: the retention parameter reaches a preset anchoring threshold (0.90 in this example) to ensure that the mastery of the item remains solid after training stops; the syllable duration variation coefficient is lower than a preset variation threshold (0.15 in this example) to ensure that the movement pattern of the item is automated and the duration of each output is highly consistent.
[0047] The coefficient of variation for syllable duration is determined by the following formula: in, This is the average duration of the target syllable in all previous outputs of this entry. The standard deviation is the corresponding value. For example, the target syllable durations of a certain historical training item in the most recent five outputs were 0.42 seconds, 0.45 seconds, 0.44 seconds, 0.43 seconds, and 0.46 seconds, respectively, with a mean of 0.44 seconds, a standard deviation of approximately 0.014 seconds, and a coefficient of variation of approximately 0.032, which is lower than 0.15. Combined with its retention parameter of 0.92, this term meets the anchoring criteria and is selected as an anchoring term. In this embodiment, six anchoring terms are maintained simultaneously, with two of them presented in each session. In addition to retired terms, the daily status can also be given by physiological signals such as heart rate variability or by the trainee's subjective rating. However, physiological signals require additional wearable devices and have poor compliance in young children, while subjective rating reintroduces the human factor that this method aims to eliminate. Therefore, using retired terms as a reference is more feasible in practice.
[0048] When the number of available generalization probes for any target phoneme is lower than a preset stock threshold, a controlled non-word generated according to the Chinese phoneme alignment rules can be optionally used as the generalization probe. The generalization degree parameters corresponding to the controlled non-words and the generalization degree parameters corresponding to the real words are listed and statistically analyzed separately.
[0049] Controlled non-words refer to syllable strings that are phonetically legal but not found in dictionaries, generated according to the rules of phonetic combination and tone system of Chinese. During generation, only combinations of initials and finals and tones actually existing in Chinese are allowed, while existing real words are excluded. Since non-words do not carry semantic meaning or have word frequency differences, children's performance on them is purely determined by their articulation ability, thus eliminating the interference of high-frequency words in real words being easier to pronounce correctly. In this embodiment, a preset stock threshold of 12 is set, meaning that when there are fewer than 12 unconsumed real word probes for a certain target phoneme, the system switches to non-words. After switching, the generalization parameters of the two types of probes are counted separately and not combined, because non-words lack semantic support, and their absolute scores are generally lower than those of real words. Mixing them would cause a step change in parameters at the switching point that is unrelated to ability.
[0050] The generalization detection item and the training item optionally satisfy a preset matching condition in at least one of the number of syllables, the categories of initials and finals, and the tone combination. The number of syllables refers to the number of syllables contained in a lexical item, the categories of initials and finals refer to the classification of the articulation place and articulation method of the remaining speech segments other than the target phoneme, and the tone combination refers to the arrangement of the tones of each syllable in a multi-syllable lexical item. For example, if the training item is the two-syllable word "three" with the target phoneme in the first syllable and the tone combination of level tone plus neutral tone, the matching generalization detection item can be the two-syllable word "loosened" which also has the target phoneme in the first syllable and the same tone combination. After the structure is controlled, the articulation loads of the two types of lexical items other than the target sound are roughly equivalent, so that the difference in children's performance on the two only comes from the single factor of "whether trained", without being mixed with the difference in the difficulty of the lexical items themselves. The matching dimensions can be selected according to the focus of the case: if the error of the target phoneme is mainly manifested as assimilation affected by adjacent sounds, the matching of initial and final categories is prioritized; if there is mutual interference between tones and articulation in children, the matching of tone combinations is prioritized.
[0051] With reference to Figure 2 , in some embodiments, visibility flags that migrate with the training role and the detection role are set for lexical items. The visibility flags of the generalization detection item, the retention detection item and the prospective detection item are set to be externally invisible, the visibility flag of the training item is set to be externally visible, and the visibility flag of a lexical item that is determined to have been externally contacted and converted into a training item is synchronously switched to be externally visible. The report for guardians is aggregated and output at the phoneme and skill levels, and candidate lexical items are only selected from lexical items outside the detection reserved set when generating the training vocabulary list.
[0052] The visibility flag is a binary attribute attached to each lexical item, which takes two values of "externally visible" and "externally invisible". Its value is not maintained manually, but automatically switches with the migration of the lexical item's role: it is set to invisible when the lexical item is selected as any type of detection item, and set to visible when selected as a training item or converted from a detection item to a training item. The detection reserved set refers to a set composed of all lexical items currently in an invisible state. The candidate pool of the training vocabulary list directly takes the complement of this set as the scope when generated, which fundamentally prevents the measurement words from being selected by the task generation module. For example, after the word "loosen" is selected as a generalization detection item, it is set to invisible, and no matter how the training vocabulary list is expanded later, it will not be included; if "loosen" is subsequently determined to have been externally contacted and converted into a training item, its flag is synchronously switched to visible, at which point it can normally appear in the training vocabulary list and home practice list. The report for guardians is aggregated and output at the phoneme and skill levels, and only presents aggregated conclusions such as "the independent production rate of the blade-alveolar fricative at the word-initial position", without listing specific lexical items, so as to prevent guardians from practicing the measurement words in advance at home according to the report.
[0053] In this method, terms do not belong to a fixed category of task items, but rather migrate between roles along several unidirectional paths. When candidate terms are initially assigned, they either become training terms or generalized probe terms. Training terms become maintenance probe terms after a training pause and maintenance interval, with the most firmly grasped and longest-lasting terms further becoming anchor terms. Generalized probe terms are consumed as soon as they are presented, and subsequently either become single-contact maintenance probe terms when the transition interval expires, or become training terms when determined to have been externally contacted. Each path is unidirectional; terms will not return to the consumed probe role. This irreversibility is the root cause of the need for specialized management of probe resources.
[0054] The following section provides a detailed explanation of the mixed layout of task items and the hierarchical recording of prompts.
[0055] S2. Present the training items and generalization probes in the same format, and collect the speech signals produced by the trainee for each task item. For training items that are not produced correctly, apply cues step by step along the cue hierarchy sequence. The cue hierarchy sequence includes the lowest cue level representing no cue applied. Record the cue level at which the training item is first correctly produced.
[0056] Mixed presentation using the same format means that two types of tasks use the same image style, the same instructions, and the same presentation interface, and appear randomly interspersed within the session. Neither the child nor the trainer can distinguish which type is currently being presented. If there are any discernible differences in the presentation format between the two types of tasks, the child will invest more effort in the type that is being presented, and the trainer will unconsciously give more waiting time for the training items. The difference in performance between the two types then incorporates differences in engagement and no longer purely reflects transfer. The randomization of the mixed presentation is completed all at the beginning of the session, and the presentation order within the same session is fixed to avoid the same terms being presented repeatedly due to re-arrangement during the session.
[0057] The initial correct output determination and output score share the same channel to ensure consistency between the determination criteria and subsequent statistical criteria. Each time a child produces an item, the system calculates the output score according to the S3 method. A correct output is determined when the score reaches a preset correct determination value; in this embodiment, this value is 60 points. The determination is completed immediately after each output, and the result drives the real-time decision of whether to apply the next level of prompt. Therefore, the scoring channel must provide results within hundreds of milliseconds. The recording granularity is at the term level, not the trial level: the same term may undergo multiple output attempts within a session; only the prompt level of the first correctly determined output is recorded, and subsequent repeated outputs do not update this record. If a training item is not correctly produced within a session up to the highest prompt level, it is recorded as unachieved and is weighted according to the level above the highest level in the independence index calculation.
[0058] In some embodiments, S2 includes sub-steps S21 to S23.
[0059] S21. Each task item is presented starting with the lowest cue level. For training items not produced correctly, cues are applied progressively up the cue level sequence, starting from the cue level, with the cue level not exceeding the cue level at which the training item was first correctly produced in the previous training session. Each presentation begins without cueing, allowing the child an opportunity for independent production before receiving any external support, avoiding the obscuring of their actual independent production ability by providing premature cues. The constraint that the cue level cannot exceed the first positive level of the previous session ensures that external support only decreases between sessions. Once a child has achieved correct production at a certain level, they will not be given stronger cues again, thus continuously and tentatively reducing their dependence on support. The input to this constraint is the first positive level record of the previous session, and the output is the cue level of the word in the current session. For example, if a training item was first correctly produced with a level 2 phoneme cue in the previous session, the cue for that word in the current session will start from level 2, instead of starting with a complete level 3 demonstration.
[0060] S22. Determine the independence index based on the cue level at which the training item is first correctly produced. The value of the cue level in the cue level sequence increases with the cue strength. The lower the cue level at which the item is first correctly produced, the higher the independence index. The independence index refines the binary result of "right or wrong" into an ordered quantity of "external support strength required to achieve correctness." This allows for the observation of tangible progress when a child cannot yet produce independently, but the required cue level has decreased from level three to level one, rather than being uniformly recorded as wrong under a binary caliber. In this embodiment, the independence index is the normalized complement of the mean of the first positive level in the two most recent training sessions relative to the highest level. That is, the closer the mean of the first positive level is to level zero, the closer the index is to 1. For example, if the first positive level of a training item in the two most recent sessions is level one and level zero, with a mean of 0.5, the independence index after normalization relative to the highest level of three is approximately 0.83. In addition to mapping by the first positive level, independence can also be measured by the number of prompts required to achieve the correctness in a single session. Both methods yield the same results when the prompt levels are equidistant, but the level mapping is more robust to cases where the intensity between levels is not equidistant.
[0061] S23. When the independence index of any training item corresponding to the target phoneme reaches a preset independence threshold, increase the proportion of the generalized probe item corresponding to the target phoneme in the mixed presentation. In this embodiment, the preset independence threshold is 0.80, corresponding to the child being able to produce the word with basically no prompt or only the weakest prompt. This sub-step takes the independence index as input and the ratio of the two types of task items in the next session as output: when the child cannot produce the target sound independently, ability transfer is objectively impossible. At this time, consuming one-time probe items will only result in a batch of inevitably low scores and permanent loss of these items; after the child can produce it independently, the proportion of probe items is increased, and the same batch of scarce items can fall at the point when they are truly informative. In this embodiment, after reaching the threshold, the number of generalized probe items in a single session is increased from 6 to 9, and the number of training items is reduced to 17 accordingly.
[0062] As an optional implementation, S2 also periodically presents prospective probes at the lowest prompt level without providing correctness feedback. In each training session, the number of prospective probes presented is lower than the number of generalized probes presented, and the same prospective probe is presented at most a preset number of times before its corresponding phoneme enters training. The presentation of prospective probes must simultaneously satisfy two conditions: no prompts and no feedback. Providing prompts would constitute interference with downstream phonemes, and providing correctness feedback would also promote learning; both would cause the observations to lose their "uninterrupted" nature. The presentation frequency is constrained by two factors: first, the natural development and changes of downstream phonemes are slow, and excessive sampling does not increase the amount of information; second, repeated attempts by children to pronounce sounds they cannot make will accumulate frustration. Therefore, in this embodiment, only 2 prospective probes are presented per session, lower than the 6 generalized probes presented in the same period, and the same prospective probe is presented at most 6 times before its corresponding phoneme enters training. After reaching this number, it is removed from the presentation queue to prevent it from gradually being practiced and losing its baseline properties through long-term repetition.
[0063] In S2, hold-up probes may optionally be presented at the lowest cue level. Hold-up probes are presented without any cueing and are interspersed with training and generalized probes in a mixed queue rather than clustered together, so as to prevent children from realizing that this batch of words is a retest and changing their answering strategy.
[0064] In S2, anchor terms can optionally be presented before training and generalization probe terms in each training session. Anchor terms are fixed at the beginning of the session because state references must be acquired before the child becomes fatigued or frustrated from continuous attempts in the session. If anchor terms are placed in the middle or at the end of the session, the deviation on them would include both the day's baseline state and the accumulated consumption of the session, thus no longer serving as a baseline.
[0065] In an optional implementation, image signals of the trainee are also acquired in S2. The image signals are acquired by a camera positioned directly in front of the child, sharing the same clock as the voice acquisition and being frame-aligned, so that facial and posture images at a given moment can be cross-indexed with voice segments at the same moment. In this embodiment, the image signal acquisition frame rate is 25 frames per second, and the resolution is sufficient to distinguish facial key points and gaze direction. After acquisition, only segments overlapping with each trial time window are retained, and the remaining segments are discarded when the session ends.
[0066] In some embodiments, a pool of generalized probes and a remaining stock are maintained. The number of generalized probes consumed in subsequent training sessions is determined based on the estimated uncertainty of the generalization parameter and the remaining stock, and generalized probes are consumed only when the reduction in estimated uncertainty by a new presentation reaches a preset incremental threshold.
[0067] The probe pool refers to the set of all generalized probes for a target phoneme that have not yet been presented; the remaining stock is the size of this set. Estimation uncertainty refers to the sampling fluctuation of the generalization parameter calculated from existing probe observations; the more observations, the smaller the fluctuation. Since probe terms are permanently consumed once presented, their expenditure must be based on marginal returns rather than a fixed rhythm: when existing observations are sufficient to define the generalization parameter within a sufficiently narrow range, presenting another term has almost no impact on the conclusion, and this presentation is a net loss. In this embodiment, the preset incremental threshold is 3%, meaning that each additional presentation must reduce the estimation uncertainty by more than 3% before consumption; simultaneously, the expenditure amount for each session is calculated based on the remaining stock and the expected number of sessions before the next advancement decision point, ensuring that the stock is evenly distributed throughout the training cycle and not exhausted early on. For example, if the remaining stock for a target phoneme is 40 and there are 8 expected sessions, the baseline expenditure amount for each session is 5, and the actual expenditure is then controlled by the incremental threshold. In comparison, the strategy of consuming at fixed intervals is simpler to implement, but it will consume a large amount of resources on observations that have lost their resolution when the gain is close to saturation.
[0068] In other embodiments, the generalized probes that have been presented are replaced by single-contact retention probes after a preset transition interval has elapsed. The retention parameter of the single-contact retention probes is determined based on the output score at the time of the first presentation. The probe pool is divided into a first sub-pool and a second sub-pool according to the syllable structure distance from the training items. The syllable structure distance is determined based on the number of syllables, the initial and final categories, and the tone combination. The first sub-pool is consumed before the second sub-pool, and the reserve amount of the second sub-pool is determined based on the expected number of remaining training sessions before the advancement decision point.
[0069] Once a generalized probe is presented, its value as a "zero-practice observation" is exhausted, but it does not lose all its informational value: after a period of time, when it is presented again, the ratio of its score to the score of the first presentation precisely reflects the retention of the material that was only exposed once in the child, and thus it can continue to serve as a special type of retention observation. In this embodiment, the preset transition interval is 8 weeks, which is longer than the 4 weeks for conventional retention probes, because the memory trace of a single exposure is inherently weaker, and if the interval is too short, the retest results will mainly reflect short-term memory rather than stable mastery.
[0070] Syllable structure distance refers to the degree of difference between the probe word and the training word in three dimensions: number of syllables, initial and final categories, and tone combinations. A word with identical syllable structure in all three dimensions has a zero distance; the more different the dimensions, the greater the distance. Words with close distances are structurally similar to the training words, making transfer most likely. Words with far distances require children to extend the target sound's movement pattern to different syllable segments and prosodic environments, representing a more rigorous transfer test. Therefore, the first sub-pool, containing close-distance words, is consumed first for daily tracking; the second sub-pool, containing far-distance words, is reserved based on the expected remaining number of sessions before the decision point, to be used when the decision is most critical. For example, if it is expected that there are 8 more sessions until the decision point, and 6 far-distance observations are needed for the decision, then at least 6 words in the second sub-pool are reserved and not used for daily expenditure.
[0071] As an optional implementation, the same training item is subjected to at most one cue trial within a single training session. The trial step size is determined from a preset step size range based on the cooperation status and the cumulative number of failures in the current training session. The independence index is determined as an interval estimate based on the cue level at which the trial fails and the cue level at which the first correct output is achieved.
[0072] Hint probing refers to attempting a lower hindrance level below the initial hindrance level determined in S21 to see if the child can complete the output with less support. A successful probing indicates that the child's independence is better than the previous session's record; a failed probing returns the child to the original initial level. Probing has costs for the child: each failure is a frustrating experience, and frequent failures can damage their willingness to participate. Therefore, in this embodiment, the same training item is probing at most once per session, and the probing step size is set to two levels. If the child cooperates well and has accumulated fewer than 3 failures in a day, they can attempt to cross two levels; otherwise, they can only cross one level. The results of the probing simultaneously provide a lower and upper bound on independence: the level at which the probing fails indicates that the child cannot yet complete the task independently at that level, and the level at which the child achieves the first correct output indicates that they can complete the task at that level. The interval between these two levels represents the true range of independence. For example, if a training item fails at level zero but is correctly produced for the first time at level one, its independence index falls within the interval between level one and level zero, reflecting the discriminative power of the observation more accurately than a single-point value.
[0073] In one alternative implementation, consumed records of generalized probes are maintained on multiple device replicas, with the consumed records being an increasing set. When replicas are merged, the union of the consumed records is taken. If the same generalized probe is detected to have been presented twice after merging, the observation presented the second time is reclassified as an observation of a single-contact maintained probe, and the generalization parameter within the affected statistical window is redefined based on the reclassified observation.
[0074] The presentation is physically irreversible: once a term has been presented to a child, no data rollback can undo the exposure. The corresponding data structure must possess the same monotonicity; therefore, consumed records are designed as an incremental set, where any copy can only add elements, not delete them. When multiple copies are merged, the union is taken, ensuring a unique merge result regardless of the merging order or offline duration. In offline scenarios, two devices may each present the same term; during merging, it is detected that the term was recorded as presented twice. While the second observation of this term loses its zero-practice attribute, it precisely meets the definition of "material that has only been exposed once and is retested after a period of time." Therefore, it is reclassified as a single-exposure probe observation and continues to be used, while simultaneously being removed from the previously included generalized probe statistics, and the generalization parameter is recalculated for the affected statistical window.
[0075] In some embodiments, the start and end times of unplanned interruptions are recorded. In the first training session after retraining, a set of retention probes covering the items during the interruption period is presented, and the trainee's individualized forgetting parameters are calibrated based on the measured retention degree parameters. The individualized forgetting parameters are used to determine the preset retention interval.
[0076] Training interruptions due to sick leave, holidays, or other reasons, while technically representing data gaps, can, from a measurement perspective, constitute a natural forgetting experiment with a known start and end time and no intervention during that period. In the first session after retraining, a set of retention probes covering the interruption period is presented. The resulting retention parameters, paired with the known interruption duration, form a set of data that can be fitted to determine the child's individual forgetting rate—the personalized forgetting parameter. This parameter then determines the preset retention interval: children who forget quickly have shorter retention intervals to allow problems to be detected before they worsen; children who forget slowly have longer intervals to allocate valuable session capacity for training. For example, a child who interrupted training for 3 weeks due to illness had a retention parameter of 0.78 measured in the first retraining session, significantly lower than the typical level of children who did not interrupt training during the same period. Based on this, the retention interval was shortened from 4 weeks to 2 weeks.
[0077] When an unplanned interruption exceeds a preset duration, causing multiple hold probes to expire simultaneously, the phonemes can be selectively sorted by their output score at the time of training stop and the number of downstream dependencies on the phoneme learning order constraint graph. Only the hold probes corresponding to the top-ranked phonemes are tested, and the test results propagate downstream along the phoneme learning order constraint graph as prior estimates. With longer interruptions, multiple batches of hold probes that should have expired dispersedly during the interruption period will expire simultaneously at the first retraining session, filling the entire session with tests and crowding out the training itself. The sorting key consists of two parts: a lower output score at the time of training stop indicates weaker initial mastery and a higher likelihood of current degradation, resulting in higher test value; a higher number of downstream dependencies on the constraint graph indicates that if the phoneme has degraded, it will block the advancement of more subsequent phonemes, also resulting in higher test value. Only the top-ranked phonemes are measured. The preservation status of the remaining phonemes is estimated by propagating the measured upstream results downstream along the constraint graph. That is, if the upstream phonemes are well preserved, the downstream phonemes that depend on them are also treated as well preserved, and are measured again when they expire or anomalies occur. The propagated estimates are marked as inferred values rather than measured values in the data and are not used in the judgments that require measurement support.
[0078] The following section provides a detailed explanation of how output scores are obtained.
[0079] S3. Force alignment of the speech signal is performed based on the target text of each task item to obtain the output score.
[0080] Forced alignment refers to the process of segmenting continuous speech signals along the time axis and mapping them to individual phonemes in the text, given the known text being read. Since the target word the child is to produce is determined at the time of presentation, and the text is known, there is no need for prior speech recognition and scoring. The speech can be segmented to the phoneme level directly, constrained by the target text. Then, the acoustic rendition of the target phoneme is measured within its corresponding time segment to determine its similarity to the typical form of that phoneme, and this is converted into a score. For example, when a child reads "three," the alignment result shows the target sound occupying a time segment from 0.12 seconds to 0.31 seconds. The acoustic score calculated and normalized for this segment is 68 points. This text-constrained approach ensures that even if the child pronounces the target sound as a substitute, the scoring channel will not misjudge it as reading a different word, but will accurately award a low score for that phoneme. In addition to acoustic likelihood conversion, the output score can also be given by phoneme-level posterior probability, that is, the posterior probability of being judged as the target phoneme rather than each competing phoneme in that time period is calculated and linearly mapped to the score interval. The ranking of the two methods is basically the same under the same scoring model.
[0081] In some embodiments, S3 includes sub-steps S31 to S33.
[0082] S31. Determine the output scores for each task item using an adapted acoustic model. The adapted acoustic model performs speaker adaptive updates based on the speech signals of the training items, while the speech signals of the generalization probe items are only used to determine the output scores. Children's speech differs significantly from adult speech in fundamental frequency, formants, and speech rate. The general acoustic model has a large bias in scoring children's speech, so the model must be adapted for this child. Speaker adaptation can be achieved by transforming the mean parameters of the model using maximum likelihood linear regression, or by adjusting the parameters of the general model based on the child's data using maximum a posteriori estimation. Both are common practices in the field. The key is that the data used for adaptation is strictly limited to the training items: if the speech of the generalization probe items is also involved in the update, the model will gradually become familiar with the acoustic forms of these terms, and the increase in its score will incorporate the component of model adaptation rather than the improvement of the child's ability. This isolation constraint keeps the model in an unadapted state for the generalization probe items, which structurally corresponds to the child's unpracticed state for these terms.
[0083] S32. Store the collected speech signals and, assuming the model version and feature extraction configuration of the adaptive acoustic model remain unchanged, periodically re-evaluate the stored historical speech signals. The input to this sub-step is existing recordings in the historical speech database, and the output is the new score of these recordings under the current model, which is then used in S33 to calculate the difference between the new and original scores. Maintaining the same model version and feature extraction configuration is a prerequisite for re-evaluation: if either of these changes, the difference between the two scores will incorporate the system shift brought about by the version upgrade, and will no longer simply reflect the drift accumulated by adaptive evaluation. In this embodiment, re-evaluation is performed once at the end of each training cycle, and the re-evaluation object is all historical speech since the most recent advancement decision.
[0084] S33. The difference between two scores for the same historical speech signal is defined as the score drift. Since the acoustic content of the same recording remains unchanged, any change in its score can only stem from a shift in the scoring scale itself. This difference explicitly measures the model familiarity component, which was previously hidden within the gain and indistinguishable from actual progress, making it a subtractable quantity. For example, if a recording from the first week was scored 64.0 points, and re-scored in the fourth week using the same model, it would be scored 64.6 points; the score drift for that period would be 0.6 points.
[0085] Reference Figure 3 The following section provides a detailed explanation of gain calculation and successive correction.
[0086] The output score is the only original observation in this method, from which all other parameters are derived. The observation scores are first weighted by validity weights and then aggregated by task category, decomposing into two parallel statistical branches: training items and generalization probe items. The first and second gains are obtained by subtracting each branch from the previous training cycle. Four calibration chains are injected into different locations: the conversation state offset and the posterior probability of cooperation are injected into the weighting stage, affecting each observation; the natural development gain and score drift are injected into the subtraction stage, affecting the periodic changes. The four chains are measured independently and are not dependent on each other, thus they can be superimposed one by one without coupling.
[0087] S4. Determine the validity weight of each output score, and determine the change of the output score statistics after weighting the training term and the generalization probe term relative to the previous training period as the first gain and the second gain, respectively.
[0088] The validity weight refers to the confidence coefficient assigned to each output score, used to suppress the impact of observations obtained under poor conditions or low cooperation during aggregation. The first gain is the weighted performance change on the training item, reflecting the direct effectiveness of training on the practiced material; the second gain is the weighted performance change on the generalization probe item, reflecting the transfer of ability to unpracticed material. Both are calculated using the same scoring channel, the same weighting caliber, and the same time window, thus forming a pair of comparable observations. Their comparison forms the basis for all subsequent judgments in this method.
[0089] S4 includes sub-steps S41 to S43.
[0090] S41. Determine the validity weights for each output score. The validity weights are synthesized from the base weights and two modulation factors: in, Based on the basic weights, preset according to task item categories. This is the session state offset for the current session. To set an upper limit on the offset that would cause the observation to fail. This represents the posterior probability of the cooperation state determined in the current conversation. The two modulation paths have different functions: the state offset is measured at the beginning of the conversation by the anchoring term, characterizing the physiological and emotional baseline of the day; the posterior probability of cooperation is jointly determined by speech and image, characterizing the child's actual engagement during the conversation. The former is determined at the beginning of the conversation and applies to the entire conversation, while the latter is updated as the conversation progresses and applies to the corresponding time period. Therefore, during synthesis, state reduction is applied first, followed by cooperation reduction. For example, if the base weight is 1.0, the daily state offset is 5.0 points, the upper limit of the offset is 25 points, the state reduction factor is 0.80, and the posterior probability of cooperation is 0.90, then the validity weight of this observation is 0.72. The weight value falls between 0 and 1; when it is zero, the observation is excluded from the statistics.
[0091] S42. Group by task category, and aggregate the output scores within this training cycle using validity weights to obtain weighted output score statistics for both training and generalization probe items. The input to this sub-step is the output score and corresponding weight for each trial, and the output is two scalar statistics at the category level, which are then used in S43 to differ from the statistics of the same category from the previous training cycle. Grouping is necessary: the absolute score levels of the two types of task items are inherently different, and merging and aggregating them would cause their changes to cancel each other out, resulting in a loss of comparable structure. In this embodiment, the statistics are taken as the weighted mean; alternatively, the weighted median can be used. The latter is more robust when extremely low scores occur in individual trials, but its discriminative power is slightly lower when the sample size is small.
[0092] S43. The changes of the two weighted statistics relative to the previous training period are respectively defined as the first gain and the second gain.
[0093] In some embodiments, in S4, a natural developmental gain is subtracted when determining the first and second gains. The natural developmental gain is determined based on the trainee's natural developmental baseline and the time interval between two adjacent training cycles. The natural developmental baseline characterizes the natural rate of change of the trainee's output score with developmental progress. Children's speech abilities naturally increase with age during training; this increase occurs even without intervention. If not subtracted, it would be included in the training gains, causing fast-developing children to be overestimated and slow-developing children to be underestimated. The natural developmental baseline is expressed as the rate of score change per unit time, multiplied by the actual time interval between two adjacent training cycles to obtain the natural developmental gain to be subtracted in that cycle. When the intervals between adjacent cycles are unequal, the actual interval is used for calculation instead of the number of cycles. For example, if two cycles are separated by 3 weeks due to a holiday, the natural developmental gain subtracted in that cycle is three times the normal value. Besides subtracting at a linear rate, in younger age groups where the developmental curve is significantly non-linear, a proportional calculation can also be used, i.e., replacing the difference with the ratio of the baseline-predicted expected score to the measured score. However, linear subtraction is sufficiently approximate within a short time window such as a training cycle.
[0094] In S4, the natural developmental baseline can optionally be determined based on the longitudinal changes in the output scores of each prospective probe. Since the phonemes corresponding to the prospective probes have never been intervened with, the changes in the child's performance on these phonemes can only come from natural development. These observations constitute a developmental curve derived from the child themselves. The longitudinal changes are regressed against time using the output scores of each prospective probe over several training cycles; the resulting slope is the natural developmental baseline. For example, if a child's prospective probe score increases from 21.0 to 25.0 in 10 weeks, the regression slope is approximately 0.4 points per week, meaning their natural developmental baseline is 0.4 points per week. Using an individual-derived baseline avoids mismatch between group norms and individuals; children with slow development will not be judged as having ineffective training simply because the group mean is applied. When observation points are sparse or unevenly distributed, ordinary least squares regression is prone to being dominated by individual outliers. In such cases, robust regression can be used to reduce the weight of outliers.
[0095] In S4, optionally, the rating shift can also be deducted when determining the first and second gains. At this point, both gains have had the natural development and rating scale shift components deducted, and their calculation formulas are as follows: in, For the gain of the corresponding category, and These are the weighted output scores for the current training cycle and the previous training cycle, respectively. This is the natural developmental baseline. The time interval between two adjacent training cycles. This represents the score drift during that period. For example, during a training cycle, the weighted statistic of the training item increases from 62.0 to 68.0, and the weighted statistic of the generalization probe item increases from 55.0 to 57.5. The natural developmental baseline is 0.4 points per week, the cycle interval is 1 week, and the score drift is 0.6 points. Therefore, the first gain is 6.0 - 0.4 - 0.6 = 5.0 points, and the second gain is 2.5 - 0.4 - 0.6 = 1.5 points. The drift is subtracted equally for both types of tasks because the shift in the scoring scale affects all scores under the same model, regardless of the term category. Regarding compensation methods, deducting from the gain afterward is equivalent to normalizing historical scores to the current scale during the scoring process. The former preserves the original scores for easier tracking and is more preferable in practice.
[0096] As an optional implementation, S4 also determines the retention parameter of the phoneme corresponding to the retained probe based on the output score of the retained probe and the output score when the retained probe stops training. The retention parameter is represented by the ratio of the retest score to the score at the time of cessation of training. A value close to 1 indicates that the mastery has been basically maintained after cessation of training, while a value significantly lower than 1 indicates that it has degenerated. For example, if the output score of a certain historical training item at the time of cessation of training is 76.0 points, and the retest score after 4 weeks is 70.0 points, the retention parameter is approximately 0.92. The score at the time of cessation of training is the score when the word was last presented as a training item. If the score itself is significantly affected by the state of that day, the weighted average of the last three scores is used to avoid distortion of the baseline itself. When the same word is retested multiple times, the parameter is updated according to the most recent retest, and historical values are retained to observe the degradation trend.
[0097] In an optional implementation, in S4, the median deviation of the output score of each anchor item relative to the anchor baseline is determined as the session state offset. The anchor baseline is determined based on the historical output scores of each anchor item, and the validity weight decreases as the session state offset increases. The anchor baseline is the average of the output scores of each anchor item after it meets the anchoring conditions, representing the stable performance of the term in a normal state. The session state offset is obtained by taking the median deviation of the measured scores of each anchor item relative to its respective baseline in a given session. Taking the median instead of the mean is to prevent the slow decay of individual anchor items from skewing the overall measurement, because decay only affects a few terms, while the state of the day will affect all terms in the same direction. The validity weight decreases as the offset increases. In this embodiment, it is reduced linearly, with no reduction when the offset is zero, and reduced to zero when it reaches 25 points. The reason for choosing retired terms that have been deeply mastered as a reference is that the capability components of such terms are saturated and have the smallest variance, and their remaining fluctuations are almost entirely attributed to the state. Compared to removing session data entirely based on state, weighted reduction retains some still valid information in low-state sessions, thus avoiding data waste.
[0098] In other embodiments, in step S4, the posterior probability of the trainee's cooperation status in the current training session is determined based on the speech and image signals. The cooperation status with the highest posterior probability is determined as the trainee's cooperation status, and the validity weight is determined based on the posterior probability of the determined cooperation status. In this embodiment, cooperation status is divided into three categories: high cooperation, medium cooperation, and low cooperation, corresponding to three typical behaviors: children's gaze task, attempting to produce according to instructions, children's intermittent distraction requiring repeated prompts, and children's refusal to answer, leaving their seats, or crying. The speech-side features used for determination include response delay, speech duration, and volume stability, while the image-side features include the duration ratio of gaze direction, facial orientation, and body posture. The features from both sides are aligned and fused on the same time window. For example, in a certain session, the speech-side response delay is normal, and the image-side gaze ratio is 0.85. After fusion, the posterior probability of high cooperation is 0.90, medium cooperation is 0.09, and low cooperation is 0.01. The cooperation status is determined to be high cooperation, and 0.90 is used as the cooperation modulation factor for this session. By using posterior probabilities instead of binary labels in the weighting, ambiguous periods between the two classes receive intermediate weights, rather than being entirely retained or discarded in an either-or manner. Fusion can employ Naive Bayes assuming conditional independence of features on both sides, or a discriminative model can be trained to directly provide the posterior; the latter performs better when the features on both sides are highly correlated.
[0099] In some embodiments, the intelligent assessment method for the effectiveness of speech rehabilitation training in children further includes: determining a minimum detectable change threshold based on the measurement standard error determined by weighting each output score according to validity weights. When the absolute values of both the first gain and the second gain are below the minimum detectable change threshold, the target phoneme is determined to be in a state of no significant change detected, and the assessment of the target phoneme is postponed to the next training cycle.
[0100] Measurement standard error refers to the dispersion of scores obtained from repeated measurements at the same true ability level, calculated by weighting the scores of each test within the current period according to validity weights. Minimum detectable change threshold refers to the magnitude of the difference between two measurements that is considered not to be caused by measurement noise, determined by the following formula: in, For weighted measurement standard error, This represents the normal quantile corresponding to the chosen confidence level. The variance from the difference of two independent measurements is twice the variance of a single measurement. (At a 95% confidence level) The threshold is 1.96. For example, when the weighted standard error of measurement is 1.5, the threshold is approximately 4.2. Continuing the previous example, if the first gain of 5.0 exceeds the threshold and the second gain of 1.5 is below it, neither is below the threshold simultaneously. Therefore, this training cycle does not enter the state of no significant change detected, and the evaluation proceeds as usual. If in a subsequent training cycle the first gain drops to 2.0 and the second gain drops to 0.8, both below the threshold, then the target phoneme is determined to be in the state of no significant change detected. This threshold is necessary because small gains may statistically be caused by measurement fluctuations. If this is directly used to advance or regress, the scheduling decision becomes a pursuit of noise. The threshold criterion is derived from the data of the current cycle itself, not from external empirical values, and therefore automatically adjusts with the data quality of the child in that period. Besides comparing by threshold, the confidence interval of the gain can also be directly calculated and observed to see if it contains zero. Both methods yield consistent conclusions at the same confidence level, making the threshold method simpler in implementation and interpretation.
[0101] As an optional implementation, the minimum detectable change threshold is determined jointly by the measurement standard error and the estimation standard error of the natural developmental baseline, and increases with the number of training cycles experienced during the delayed evaluation of the target phoneme. The natural developmental baseline itself is fitted from a finite number of prospective observations, and it inherently carries estimation errors. Subtracting the action introduces a new uncertainty into the gain; including this uncertainty in the threshold prevents the judgment from being too lenient due to ignoring the error of the action itself. Increasing the threshold with the number of delay cycles is to gradually tighten the conclusion as the same target phoneme is repeatedly delayed, avoiding the system from idling indefinitely in a state of long-term undetected activity. When the number of observation points for the prospective detection items used to determine the natural developmental baseline is lower than the preset lower limit, the natural developmental gain is determined based on population developmental norm data. In this embodiment, the preset lower limit is 8 observation points. If the number is lower than this, the estimation error of the regression slope is too large to be accepted. In this case, the population developmental norm of the same age group is temporarily used as a substitute, and the individual baseline is switched back after the individual observation points accumulate above the lower limit. This degradation path allows the method to continue operating even when there is insufficient longitudinal data in the early stages of training.
[0102] If a generalized probe is correctly produced at the lowest cue level upon first presentation, and both its reaction time and syllable duration coefficients of variation fall within the trainee's final performance range on items for which training has ceased, then the generalized probe may be optionally classified as having been externally touched. Only the output scores of generalized probes not classified as having been externally touched are used to determine the second gain; terms classified as having been externally touched are converted to training items.
[0103] The final performance range refers to the range of reaction time and syllable duration coefficient of variation for words that have been sufficiently trained to the point of cessation of training, representing the performance characteristics when the child has "mastered" the word. If a word that has never been trained is correctly produced without prompting upon first presentation, and the shortness of the reaction time and the stability of the duration are at the level of a mastered word, then the most reasonable explanation is that the word has been exposed to and practiced outside of training, such as a guardian leading practice at home in advance based on a report. Such words no longer possess the attribute of zero practice, and their performance would artificially inflate the transfer if included in the second gain calculation, so it must be removed from the second gain calculation; at the same time, since it has already been practiced, it is better to formally convert it into a training item to avoid wasting this exposure. For example, if a generalization probe is read correctly without prompting upon first presentation, with a reaction time of 0.6 seconds and a syllable duration coefficient of variation of 0.05, both of which fall within the performance range of the word that the child has mastered, it is determined that the child has been exposed to it externally. Using children's own performance characteristics as the criterion has the advantage of not relying on truthful feedback from families, as guardians often do not feel the need to report helping their children practice a few more words. In contrast, collecting practice records directly from families is simpler to implement, but its accuracy depends on the willingness of those who wish to report, which is difficult to guarantee in practice.
[0104] In other embodiments, when the deviation of any anchor term constitutes an outlier relative to the median deviation of all anchor terms, the anchor term is removed from the anchor term set and replaced by a replacement from historical training terms that meet the anchoring conditions. Outliers can be determined by the absolute median difference; that is, if the difference between the deviation of an anchor term and the median exceeds a certain multiple of the absolute median difference, it is considered an outlier. In this embodiment, a multiple of 3 is used. The replacement mechanism enables the term set serving as a state reference to have self-maintenance capabilities, so that the entire set will not drift when individual terms lose their reference eligibility due to forgetting.
[0105] In some embodiments, external cues are detected by analyzing the trainee's gaze direction and the speech of bystanders based on image signals. Only output scores for which no external cues were detected are used to determine the first and second gains. The independence index is determined only based on cues not marked as suspicious external cues. When observables disappear in the current training session, low-difficulty tasks that have stopped training are first presented to distinguish between insufficient ability and no response. Then, a self-test of the acquisition channel is performed through calibration signal re-sampling to distinguish between equipment failure and participation termination.
[0106] External cues refer to lip-reading demonstrations, whispered prompts, or nodding cues given by people other than the trainer during a child's response. These cues are not recorded in the cue hierarchy sequence, but they genuinely reduce the difficulty of production; failure to recognize them will result in inflated scores. Detection is based on two criteria: first, if the child's gaze shifts to the side rather than the task presentation interface at the moment of response, it indicates that they are receiving information from outside the interface; second, if a voice from the side is detected within the child's response time window. If either cues are detected, the trial is marked as having suspicious external cues; its score is not included in the gain, and its cue hierarchy record is not included in the independence index.
[0107] The disappearance of observations refers to the failure to collect valid speech or detect correct output in multiple consecutive trials, resulting in no output from the scoring channel. This phenomenon has at least three causes: the child's ability is insufficient to produce the task at the current difficulty level, the child has stopped participating and no longer responds, and the acquisition equipment itself is malfunctioning. The handling of these three causes differs significantly. If all are scored as failures, equipment malfunction would be attributed to ability degradation. Therefore, a two-stage detection approach is adopted to eliminate possibilities sequentially: First, a low-difficulty task that the child has stopped training on is presented. If the child responds normally, it indicates that the equipment is working properly and the child is still participating; the previous silence should be attributed to the current difficulty being too high. If there is still no response, a self-check of the acquisition channel is performed by playing a calibration signal and re-acquiring the data. If the re-acquisition signal is normal, it indicates that the equipment is not malfunctioning and the child has indeed stopped participating; if the re-acquisition signal is abnormal, it is judged as equipment malfunction, and data acquisition for this session is terminated.
[0108] As an optional implementation, when the output score of a generalized probe deviates from the expected range by more than a preset deviation threshold, attribution determination is performed in the following order: score drift verification, external contact determination, and confirmation of actual improvement. Based on the attribution results, the process of deducting score drift, removing terms, and proceeding along the phoneme learning order constraint graph is executed. The verification order of the three causes is arranged from highest to lowest exclusionability: score drift, directly measured from historical speech re-evaluation, can be immediately confirmed as being caused by drift, hence it is placed first; external contact can be determined from the performance characteristics of this output, requiring comparison of the performance range, hence it is placed in the middle; actual improvement is the remaining explanation after excluding the first two, requiring further verification from data in subsequent cycles, hence it is placed last. This order ensures that the most costly decision is only triggered after the first two causes have been excluded.
[0109] The following section provides a detailed explanation of how to generate the training task configuration parameters.
[0110] S5. Determine the generalization parameter based on the first gain and the second gain. Based at least on the generalization parameter and the recorded cue levels, generate training task configuration parameters including the target phonemes for the next training cycle, the training vocabulary, and the cue starting level. The words in the training vocabulary will be used as training items in the next training cycle.
[0111] The generalization parameter, which refers to the proportion of ability transfer relative to the effectiveness of direct training, is determined by the following formula: in, and These are the first gain and the second gain, respectively. A parameter close to 1 indicates that the training results have been almost completely transferred to the unpracticed material, while a parameter close to 0 indicates that the improvement in performance is basically limited to the few words that have been repeatedly practiced. Continuing the previous example, the first gain is 5.0 points, the second gain is 1.5 points, and the generalization parameter is 0.30. The training task configuration parameters include three items: the target phoneme for the next training cycle determines whether the training focus shifts forward, the training vocabulary determines which specific words are practiced, and the cue starting level determines the level of support to start with. These three items are jointly determined by the generalization parameter and the recorded cue level. The former answers "whether the ability has truly transferred," and the latter answers "how much support the child still needs." The absence of either will result in an inappropriate configuration. After the configuration is generated, it is fed back into S1 of the next cycle. The words in the training vocabulary are used as training items in the next training cycle, thus forming a closed-loop iteration of "presentation, scoring, gain calculation, configuration generation, and re-presentation." In addition to being constructed based on the ratio of gains, the generalization parameter can also be normalized based on the difference between the two types of gains relative to the first gain. The two are monotonically corresponding when the gains are both positive, and the ratio form is more intuitive to interpret.
[0112] In some embodiments, S5 includes sub-steps S51 to S52.
[0113] S51. Determine the comparison results of the generalization parameter with the preset generalization threshold, the retention parameter with the preset retention threshold, and the cue level at which the training item is first correctly produced with the preset level threshold. In this embodiment, the three thresholds are set to 0.35, 0.85, and level one, respectively: the generalization parameter must reach 0.35, indicating that transfer has indeed occurred and not just that the training words have been memorized; the retention parameter must reach 0.85, indicating that the mastered content can be maintained after training stops; the cue level at which the first correct production is produced must not be higher than level one, indicating that the child can produce independently with very little support. All three conditions must be met simultaneously to constitute the advancement condition. Meeting any single condition is insufficient to trigger advancement, because the three conditions correspond to three independent aspects: transfer, maintenance, and independence. The absence of any one condition means that the current phoneme has not been truly mastered. Continuing the previous example, if the generalization parameter of 0.30 is lower than 0.35, no advancement will be made even if the other two conditions are met.
[0114] S52. Based on the comparison results, the target phoneme for the next training cycle is determined along the topological order of the phoneme learning order constraint graph. The phoneme learning order constraint graph uses phonemes as nodes and directed edges based on the learning order dependencies between phonemes. When all three comparison results meet the criteria, the next target phoneme is selected from the direct downstream nodes of the current target phoneme whose in-degree is fully satisfied, meaning all its preceding dependent phonemes have met the criteria. If multiple candidates exist, the one with the most downstream dependencies is selected to remove the blockage on more subsequent phonemes as early as possible. Advancing according to the topological order rather than according to error frequency or trainee preference ensures that the movement patterns required for the new target phoneme are all met, preventing children from repeatedly failing on phonemes where the foundation is not yet established. In addition to the form where all three conditions are met simultaneously, the advancement criterion can also be a weighted sum of the three conditions and compared with a single threshold. The weighted form is more flexible when the three conditions are inversely related, but it allows a serious deficiency in one condition to be compensated by the other two. Therefore, this embodiment adopts the form where all three conditions are met simultaneously.
[0115] In S5, if the increase in the output score of any prospective probe exceeds the expected range determined based on the natural developmental baseline, the corresponding phoneme can be optionally marked as spontaneously acquired on the phoneme acquisition order constraint map. Phonemes marked as spontaneously acquired are then skipped when generating training task configuration parameters. The expected range is obtained by multiplying the natural developmental baseline by the elapsed time and adding its estimation error. A prospective probe score falling within this range indicates normal development; a significant deviation indicates that the phoneme has been mastered spontaneously outside of training. If such phonemes are still trained according to the constraint map order, the entire training cycle will be wasted on content the child already knows. Therefore, they are skipped during configuration generation, allocating the time to phonemes that truly require intervention.
[0116] In other embodiments, in S5, the training task configuration parameters for the next training cycle are generated to be the same as the current training task configuration parameters, and the number of generalized probes corresponding to the target phoneme in the next training cycle is reduced. When the absolute value of the first gain reaches the minimum detectable change threshold, the number of generalized probes is restored. When the current cycle is judged to have no significant change detected, any configuration adjustment lacks data basis, and maintaining the original configuration for another cycle is the only reasonable action; at the same time, since the current data is not sufficient to support the conclusion, continuing to consume one-time probe entries at the original rate is a pure waste, so the number of their presentation is reduced simultaneously. In this embodiment, the number of generalized probes in the undetected state is reduced from 6 per session to 3. When the first gain exceeds the threshold again, indicating that the training has once again produced detectable results, the original number of presentations is restored to support the subsequent progress judgment.
[0117] As an optional implementation, the trainee's pronunciation deviation is determined as either systematic or unstable based on the convergence time constant of the score drift as a function of the number of training sessions. If determined to be systematic, training task configuration parameters for comparative focused training targeting alternative directions are generated; if determined to be unstable, training task configuration parameters for consistent and stable training are generated. The change in score drift with the number of sessions reflects the amount of data required for the acoustic model to adapt to the child's pronunciation, and its decay process can be obtained by exponentially fitting the convergence time constant. Fast convergence means that although the child's pronunciation deviates from the standard, it is highly consistent, and the model can adapt within a few sessions, classifying it as systematic. Slow convergence means that each pronunciation is different, making it difficult for the model to stabilize, classifying it as unstable. In this embodiment, a convergence time constant less than 6 sessions is considered systematic; otherwise, it is considered unstable. The training configurations for the two types are opposite: children with systematic biases need to replace stable erroneous movement patterns with correct ones, thus generating contrastive focused training targeting their substitution direction, i.e., intensively presenting the target sound and its substitute sound as minimal opposing pairs; children with unstable biases need to first fix their movement patterns, thus generating consistent and stable training, i.e., highly repetitive consolidation of a single output form in the same context. The value of this classification lies in the fact that it is derived from the drift data already generated during system operation, without the need to collect any additional data for classification.
[0118] In one optional implementation, the external contact rate is defined as the proportion of generalized probes identified as having been externally contacted to the total number of generalized probes presented within the statistical window. When both the generalization parameter and the external contact rate are below a preset generalization threshold, the coverage of the training vocabulary is expanded. When both the generalization parameter and the external contact rate reach the preset contact threshold, the size of the training vocabulary is maintained, and the contextual variation dimension of the training task is adjusted. The phenomenon of a low generalization parameter actually corresponds to two different causes. If the external contact rate is also low, it indicates that the measurement is clean, and transfer has indeed not occurred. The problem lies in the narrow phonological environment covered by the training vocabulary; the child has only learned to pronounce the target sound correctly in specific syllable structures. In this case, the vocabulary coverage should be expanded. If the external contact rate is high, it indicates that a significant portion of the probe terms have been pre-practiced and eliminated by families. The remaining sample size is small and lacks representativeness, and the low parameter is likely due to measurement contamination. In this case, expanding the vocabulary is not the solution. Instead, the vocabulary size should be maintained while changing the contextual variation dimension of the training task, such as shifting from single-word output to phrase and sentence output, to improve the transfer requirements without adding new words. In this embodiment, the preset generalization threshold is set to 0.35, and the preset contact threshold is set to 0.15. For example, if 24 generalized probe terms are presented in the statistical window, and 2 are determined to have been externally contacted, the external contact rate is approximately 0.083, which is lower than 0.15. Combined with a generalization parameter of 0.30, the approach is to expand the coverage of the training vocabulary.
[0119] In some embodiments, a trial period is set after advancing the target phoneme along the phoneme acquisition sequence constraint graph. During the trial period, if the slope of the independence index of the new target phoneme is lower than a preset slope threshold, the system reverts to the target phoneme before advancement, and this revert event is included in the advancement tolerance parameter. The advancement tolerance parameter is used to adjust the preset thresholds on which the advancement decision is based. Even if all three conditions are met simultaneously, there is still a possibility of misjudgment in the advancement decision. The purpose of the trial period is to provide a revocable observation window for this decision, limiting the cost of misjudgment within this window rather than waiting until the child stagnates on the new phoneme for a long time before it is discovered. In this embodiment, the trial period is set as two training cycles. During this period, if the slope of the independence index of the new target phoneme is lower than 0.02 per session, it is judged as premature advancement, and the system reverts to the original target phoneme to continue training. Each revert is included in the advancement tolerance parameter, which in turn raises the three thresholds on which the advancement decision is based, making the system automatically more cautious after multiple misjudgments.
[0120] When the first gain is below a preset saturation threshold, the method for determining the generalization parameter can be optionally switched from being based on the gain ratio to being based on the arrival time ratio. The arrival time ratio is the ratio of the number of presentations required for the training item and the generalization probe item to reach the same output score level. Once the training item's score approaches full marks, its gain will inevitably approach zero. At this point, the denominator of the gain ratio is too small, causing the parameter value to fluctuate wildly and lose its discriminative power, making it impossible to judge the child's transfer status at higher levels. The arrival time ratio uses a different measurement method: instead of comparing the increase in score per unit time, it compares how many presentations are required for each type of term to reach the same score level. This method remains stable even when the score is close to its upper limit. In this embodiment, the switch is executed when the first gain is below 2.0. For example, after the switch, statistics show that the training item reaches 80 points on average after 6 presentations, while the generalization probe item requires 20 presentations to reach the same level, resulting in an arrival time ratio of 0.30, consistent with the parameter level before the switch. Besides switching the measurement caliper, the results can also be pushed back above the upper limit by increasing the difficulty of the task. However, changing the difficulty will interrupt the longitudinal comparability, so switching the caliper is the better option.
[0121] In other embodiments, when the output scores of prospective probes corresponding to multiple non-adjacent downstream phonemes rise synchronously, a developmental acceleration period is identified, and the progression along the phoneme acquisition order constraint graph is frozen until the synchronous rise subsides. "Non-adjacent" means that these phonemes are not connected by directed edges on the constraint graph, thus preventing the mastery of one phoneme from influencing another. The only reasonable explanation for the synchronous improvement of several unrelated phonemes within the same time period is that the child has entered an accelerated phase of speech development as a whole, rather than the effectiveness of the current training. If conventional criteria are followed at this point, the developmental leap would be mistakenly recorded as training success and used to accelerate scheduling, pushing the child to later phonemes without adequate preparation. Therefore, progression is frozen during the acceleration period, maintaining only the training of the current phoneme until the synchronous rise subsides and the changes in each phoneme return to an independent state before resuming normal judgment.
[0122] The following explains why this method uses children as trainees, including the differences between children and adults in speech rehabilitation training, the design of this method to address these differences, the differences between this method and assessment methods for adults, and the applicability of this method to adults.
[0123] In speech rehabilitation training, the differences between children and adults are mainly reflected in five aspects. First, the nature of the training differs. Children's speech disorders often occur during the speech acquisition process, where the target phoneme has never been correctly mastered before. Training helps children establish the ability to produce that phoneme for the first time, and the acquisition of each phoneme is sequential. Adult speech rehabilitation training mainly targets acquired speech disorders caused by stroke, traumatic brain injury, etc. The trainee had already mastered all the phonemes of their native language before the illness; the training aims to restore or compensate for the original ability, without a process of progressively advancing phonemes according to the acquisition order. Second, the sources of natural changes in performance differ. Children's speech abilities naturally improve with age during training, and even untrained phonemes will improve, with the rate of improvement varying from person to person. Adults' speech systems are already fixed, and there is no natural development with age. Third, the acoustic characteristics of speech differ. Children's fundamental frequency, formants, and speech rate differ significantly from adults and continue to change with growth. General acoustic models often use adult speech as corpus, resulting in significant biases in scoring children's speech. Adult acoustic characteristics are stable and consistent with the corpus of general acoustic models. Fourth, their level of cooperation and dependence on external support differ. Children have short attention spans and are easily affected by drowsiness, emotions, and physical discomfort. Their performance in training sessions at the same ability level fluctuates greatly, often requiring external cues to complete output, and they cannot accurately state whether they have tried their best. Adults generally understand and consistently cooperate with instructions and can describe their daily state. Fifth, their practice methods outside of training differ. Children's home practice is led by guardians, who may have the children practice the vocabulary used for measurement in advance. Adults arrange their own practice and can understand and adhere to the requirement not to practice measurement materials in advance.
[0124] To address the aforementioned differences, this method is designed accordingly. Regarding the first aspect, S1 sets up a generalization probe containing the target phoneme, which has not been used in training and is presented at most once. S2 presents the generalization probe and training probe in the same manner. S4 and S5 determine the first gain, second gain, and generalization parameters accordingly. S52 determines the target phoneme for the next training cycle according to the topological order of the phoneme learning sequence constraint graph. This combination tests whether the child has established the ability to produce the target phoneme and can use it for unpracticed words, rather than simply memorizing words from the training vocabulary. This ensures that the child is not pushed to the next phoneme before mastering the target phoneme, nor remains stagnant after mastering it. Regarding the second aspect, the natural developmental baseline is calibrated using the child's own output score on prospective probes of untrained downstream phonemes. The natural developmental gain is subtracted from the first and second gains, and spontaneously acquired phonemes are skipped. Progress is frozen during periods of accelerated development, ensuring that slower-developing children are not deemed to have ineffective training, and faster-developing children are not overestimated. Training cycles are not wasted on phonemes already mastered by the child. Regarding the third aspect, the adaptive acoustic model performs speaker-adaptive updates for each child. Updates use only the speech signal from the training items, and the score drift is obtained by re-scoring historical speech signals and subtracted from the gain, ensuring the score matches the child's acoustic characteristics. Simultaneously, the model's gradual adaptation to the child's speech is not counted as progress.
[0125] Regarding the fourth aspect, each task item is presented starting from the lowest cue level without prompting. For training items that are not produced correctly, prompts are applied level by level along the cue level sequence, and the cue level at the first correct production is recorded. Based on this, the independence index and the cue starting level for the next training cycle are determined, so that the progress of children who cannot produce independently but whose need for prompts has been reduced can be observed. External support for children is gradually withdrawn as their ability improves. The effectiveness weight is determined based on the conversation state offset determined by the anchor item and the posterior probability of the cooperation state determined by the speech signal and image signal, so that low scores obtained by children when their state is poor or their cooperation is insufficient are not counted as a decline in ability. The number of prospective exploration items presented and the number of cue attempts are limited to reduce the frustration of children due to repeated failures. Regarding the fifth aspect, visibility markers are set for entries, and reports for guardians are aggregated and output according to phonemes and skill levels without listing specific entries. The generalization probes are determined based on the child's own performance characteristics to determine whether they have been exposed to externally. The entries used for measurement are not invalidated by the guardian's prior practice, and this determination does not depend on the guardian's reporting.
[0126] Adult-oriented assessment methods typically use the same set of assessment materials to be administered repeatedly before and after training, with scoring by therapists or general acoustic models. The difference in scores before and after is directly taken as the training effect, assuming that the trainee's speech system is already established, acoustic characteristics are stable, and they can respond consistently to instructions. This method differs from this in three ways: First, the generalization probes used for measurement are presented at most once and are mixed with the training items in the same format, without repeating the same set of assessment materials; Second, before the change in scores is recognized as a training effect, the natural developmental gain and score drift determined by the child's own data are deducted, and observations of poor condition or insufficient cooperation are reduced according to validity weights; Third, the assessment output includes not only the score of training effect, but also the target phonemes, training vocabulary, and cue starting level for the next training cycle, with the target phonemes determined according to the phoneme acquisition order.
[0127] The above design is based on the characteristics of children in the speech development stage and is not applicable to speech rehabilitation training for adults for four reasons. First, adults do not have natural development with age, nor do they have a phoneme acquisition sequence to follow. A phoneme acquisition sequence constraint map cannot be established, and prospective probes cannot be selected. Therefore, there is no target for defining and deducting the natural developmental baseline, marking spontaneously acquired phonemes, determining the developmental acceleration period, and advancing target phonemes along the constraint map. Second, the performance changes in acquired speech disorders in adults include components such as spontaneous recovery after the illness, the source and time process of which differ from children's speech development. This method does not set corrections for these components. Third, adults' acoustic characteristics are stable and consistent with the corpus of a general acoustic model, eliminating the need for continuous speaker adaptive updates for individuals. Therefore, the measurement and deduction of score drift are unnecessary. Fourth, adults can generally cooperate stably and arrange their own practice. The sources of problems addressed by the determination of cooperation status, visibility markers, and external contact determination do not exist. Even if the same data processing is performed on adult speech signals, the components targeted by each correction do not exist, and the target phonemes generated for the next training cycle are not instructive.
[0128] The same or similar parts in the various embodiments of this application can be referenced to each other, and the technical features of each embodiment can be combined with each other to form new embodiments without conflict.
[0129] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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. Such 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 this application, and should all be included within the protection scope of this application.
Claims
1. An intelligent assessment method for the effectiveness of speech rehabilitation training in children, characterized in that, include: S1. Determine a set of task items including training items and generalization probes, wherein the generalization probes are words containing target phonemes that have not been used for training, and each of the generalization probes is presented at most once; S2. The training items and the generalization probe items are presented in the same presentation format. The speech signals produced by the trainee for each task item in the task item set are collected. The training items that are not produced correctly are prompted level by level along the prompting level sequence. The prompting level sequence includes the lowest prompting level that represents no prompting. The prompting level at which the training item is first produced correctly is recorded. S3. Force alignment is performed on the speech signal based on the target text of each task item to obtain the output score; S4. Determine the validity weight of each output score, and determine the change of the output score statistics after weighting the training item and the generalization probe item relative to the previous training cycle as the first gain and the second gain, respectively. S5. Determine the generalization parameter based on the first gain and the second gain, and generate training task configuration parameters including the target phoneme, training vocabulary, and cue start level for the next training cycle based at least on the generalization parameter and the recorded cue level, wherein the words in the training vocabulary are used as training items in the next training cycle.
2. The intelligent assessment method for the effectiveness of speech rehabilitation training in children according to claim 1, characterized in that, In step S4, when determining the first gain and the second gain, the natural developmental gain is subtracted. The natural developmental gain is determined based on the trainee's natural developmental baseline and the time interval between two adjacent training cycles. The natural developmental baseline characterizes the natural rate of change of the trainee's output score with the developmental process.
3. The intelligent assessment method for the effectiveness of children's speech rehabilitation training according to claim 2, characterized in that, The task set also includes prospective probes, which are words selected from the words corresponding to phonemes that are downstream of the target phoneme and have not yet entered training on the phoneme learning order constraint graph. The phoneme learning order constraint graph uses phonemes as nodes and directed edges based on the learning order dependencies between phonemes. In S2, the prospective probes are presented periodically at the lowest prompt level without providing correctness feedback. The number of presentations of the prospective probes in each training session is lower than the number of presentations of the generalization probes. The same prospective probe is presented at most a preset number of times before its corresponding phoneme enters training. In S4, the natural development baseline is determined based on the longitudinal change of the output scores of each prospective probe. In S5, in response to the increase in the output score of any prospective probe exceeding the expected range determined based on the natural development baseline, the corresponding phoneme is marked as spontaneously learned on the phoneme learning order constraint graph, and the phonemes marked as spontaneously learned are skipped when generating the training task configuration parameters.
4. The intelligent assessment method for the effectiveness of children's speech rehabilitation training according to claim 1, characterized in that, S3 includes the following sub-steps: S31. The output score of each task item is determined by an adaptive acoustic model, wherein the adaptive acoustic model performs speaker adaptive update based on the speech signal of the training item, and the speech signal of the generalized probe item is used only to determine the output score. S32. Store the collected speech signals, and periodically re-evaluate the stored historical speech signals while keeping the model version and feature extraction configuration of the adapted acoustic model unchanged. S33. The difference between two scores of the same historical speech signal is determined as the score drift; in S4, the score drift is also deducted when determining the first gain and the second gain.
5. The intelligent assessment method for the effectiveness of speech rehabilitation training in children according to claim 1, characterized in that, The task set also includes a retention probe, which is a historical training item that has stopped training for a preset retention interval; in S2, the retention probe is presented at the lowest prompt level; in S4, the retention degree parameter of the phoneme corresponding to the retention probe is determined based on the output score of the retention probe and the output score when the retention probe stops training.
6. The intelligent assessment method for the effectiveness of speech rehabilitation training in children according to claim 5, characterized in that, S5 includes the following sub-steps: S51. Determine the comparison results of the generalization parameter and the preset generalization threshold, the comparison results of the retention parameter and the preset retention threshold, and the comparison results of the prompt level at which the training item is first correctly produced and the preset level threshold. S52. Based on the comparison results, determine the target phoneme for the next training cycle along the topological order of the phoneme learning order constraint graph, wherein the phoneme learning order constraint graph has phonemes as nodes and directed edges with the learning order dependency between phonemes.
7. The intelligent assessment method for the effectiveness of speech rehabilitation training in children according to claim 5, characterized in that, Each training cycle includes at least one training session; the task set also includes anchor terms, which are words selected from historical training terms that have stopped training and meet the anchoring conditions. The anchoring conditions include the retention parameter reaching a preset anchoring threshold and the syllable duration variation coefficient being lower than a preset variation threshold. In S2, the anchor terms are presented before the training terms and the generalization probe terms in each training session. In S4, the median deviation of the output score of each anchor term from the anchoring baseline is determined as the session state offset. The anchoring baseline is determined based on the historical output scores of each anchor term, and the validity weight decreases as the session state offset increases.
8. The intelligent assessment method for the effectiveness of speech rehabilitation training in children according to claim 1, characterized in that, The intelligent evaluation method for the effect of speech rehabilitation training in children further includes: determining a minimum detectable change threshold based on the measurement standard error determined by weighting the effectiveness weights of each output score; in response to the absolute values of both the first gain and the second gain being lower than the minimum detectable change threshold, determining the target phoneme as having no significant change detected, postponing the evaluation of the target phoneme to the next training cycle, wherein in step S5, the training task configuration parameters for the next training cycle are generated to be the same as the current training task configuration parameters, and the presentation number of the generalized probe corresponding to the target phoneme in the next training cycle is reduced; in response to the absolute value of the first gain reaching the minimum detectable change threshold, restoring the presentation number of the generalized probe.
9. The intelligent assessment method for the effectiveness of speech rehabilitation training in children according to claim 1, characterized in that, In step S2, image signals of the trainee are also collected; in step S4, the posterior probability of the trainee being in each cooperation state in the current training session is determined based on the voice signal and the image signal, the cooperation state with the highest posterior probability is determined as the cooperation state of the trainee, and the effectiveness weight is determined based on the posterior probability of the determined cooperation state.
10. The intelligent assessment method for the effectiveness of speech rehabilitation training in children according to claim 1, characterized in that, S2 includes the following sub-steps: S21. Each task item in the task item set is presented starting from the lowest prompt level. For the training item that is not correctly produced, prompts are applied level by level along the prompt level sequence starting from the prompt starting level. The prompt starting level is not higher than the prompt level at which the training item was first correctly produced in the previous training session. S22. Determine the independence index based on the cue level at which the training item is first correctly produced. The value of the cue level in the cue level sequence increases with the increase of cue intensity. The lower the cue level at which the item is first correctly produced, the higher the independence index. S23. In response to the independence index of any training item corresponding to the target phoneme reaching a preset independence threshold, increase the proportion of the generalization probe item corresponding to the target phoneme in the mixed presentation.
