Sensitive word dynamic monitoring method and device in Tibetan complex environment speech recognition

Abstract

The invention provides a sensitive word dynamic monitoring method and device in Tibetan complex environment speech recognition, and the method comprises the steps: carrying out the segment segmentation of a to-be-monitored Tibetan speech input stream in a complex environment, and obtaining the framed speech signals of the Tibetan speech in the complex environment; performing multi-dimensional perceptual fusion on the phoneme-level score, the semantic-level score, the scene-level score and the emotion-level score of each candidate sensitive word in the framing voice signals to obtain perceptual fusion features of the context corresponding to each candidate sensitive word; determining the decoding confidence coefficient of the Tibetan syllable sequence through the path stability of a decoding path in speech recognition and the posterior probability of each Tibetan syllable in the Tibetan syllable sequence; and determining the context sensitivity of each candidate sensitive word based on the decoding confidence and each perceptual fusion feature, and carrying out graded early warning on the Tibetan speech input stream in the complex environment through the context sensitivity. Based on the scheme, multi-dimensional fusion scoring of sensitive words in Tibetan complex environment speech recognition can be realized.

Description

Technical Field

[0001] This application relates to the field of Tibetan speech recognition technology. More specifically, this application relates to a method and device for dynamically monitoring sensitive words in Tibetan speech recognition in complex environments. Background Art

[0002] Tibetan, as a language with rich dialects and complex phonetic structures, requires comprehensive consideration of the diversity of speech and the complexity of semantics in the recognition of sensitive words. In Tibetan speech recognition, the recognition of sensitive words is an important technical challenge. At present, certain progress has been made in Tibetan speech recognition technology, but the recognition of sensitive words remains an urgent problem to be solved.

[0003] The pronunciation differences among the three major Tibetan dialects pose severe challenges to the monitoring of sensitive words. Due to the lack of adaptability to different dialects, traditional single-phoneme matching methods are prone to misjudging these legal dialect variants as sensitive words, resulting in a high false alarm rate. At the same time, the special acoustic environment on the plateau further exacerbates the recognition difficulty: the reverberation effect in religious places will cause the syllable boundaries to be blurred, and legal dialects may be misrecognized as sensitive words under echo interference, while continuous wind noise will cover up the features of voiceless consonants, making it difficult to distinguish minimal pairs and causing missed detections. Therefore, how to achieve multi-dimensional fusion scoring of sensitive words in Tibetan speech recognition in complex environments has become a difficult problem faced by the industry. Summary of the Invention

[0004] This application provides a method and device for dynamically monitoring sensitive words in Tibetan speech recognition in complex environments, which can achieve multi-dimensional fusion scoring of sensitive words in Tibetan speech recognition in complex environments.

[0005] In a first aspect, this application provides a method for dynamically monitoring sensitive words in Tibetan speech recognition in complex environments, including: Obtain the input stream of Tibetan speech to be monitored in a complex environment; Segment the input stream of Tibetan speech to obtain the framed speech signals of Tibetan speech in the complex environment, and then extract the multi-dimensional features of the speech syllables in the context of the framed speech signals; Perform a preliminary match between the framed speech signals and the sensitive word library of Tibetan speech to obtain multiple candidate sensitive words in the input stream of Tibetan speech. Then, perform multi-dimensional perceptual fusion on the phoneme-level score, semantic-level score, scene-level score, and sentiment-level score of each candidate sensitive word in the multi-dimensional features to obtain the perceptual fusion features corresponding to the context of each candidate sensitive word; Use a Tibetan acoustic model to perform speech recognition on the framed speech signal, generate a sequence of Tibetan syllables, and then determine the path stability of the decoding path in speech recognition and the posterior probability of each Tibetan syllable in the sequence of Tibetan syllables. Determine the decoding confidence of the sequence of Tibetan syllables based on the path stability and each posterior probability; Based on the decoding confidence and each perceptual fusion feature, perform a fusion evaluation on each candidate sensitive word to obtain the context sensitivity of each candidate sensitive word, and then perform a hierarchical warning on the Tibetan speech input stream in a complex environment based on the context sensitivity.

[0006] In some embodiments, extracting the multi-dimensional features of speech syllables in the context of the framed speech signal specifically includes: Set the window radius and window time interval of the time context window; For each speech syllable word in the framed speech signal, match the speech syllable word with a standard sensitive word according to the Tibetan pinyin fuzzy matching algorithm to obtain the phoneme-level score of the speech syllable word; Evaluate the semantic rationality of the speech syllable word in the current context based on the Tibetan semantic model to obtain the semantic-level score of the speech syllable word; Adjust the sensitivity threshold through the acoustic scene classification label and the speaker's dialect feature to obtain the scene-level score of the speech syllable word; Evaluate the emotional polarity intensity of the candidate sensitive word through the Tibetan emotion dictionary to obtain the emotion-level score of the speech syllable word; Furthermore, obtain the phoneme-level score, semantic-level score, scene-level score, and emotion-level score of each speech syllable word in the framed speech signal, and determine the multi-dimensional features of the speech syllables in the context of the framed speech signal through all the phoneme-level scores, semantic-level scores, scene-level scores, and emotion-level scores.

[0007] In some embodiments, initially matching the framed speech signal with a sensitive word library of Tibetan speech to obtain multiple candidate sensitive words in the Tibetan speech input stream specifically includes: Obtain the sensitive word library of Tibetan speech and the candidate sensitive threshold; Perform an edit distance match between each speech syllable word in the framed speech signal and the sensitive word library to obtain the Tibetan pinyin edit distance of each speech syllable word; Screen out multiple candidate sensitive words in the Tibetan speech input stream from the framed speech signal through each Tibetan pinyin edit distance and the candidate sensitive threshold.

[0008] In some embodiments, the multi-dimensional perceptual fusion of the phoneme-level score, semantic-level score, scenario-level score, and sentiment-level score of each candidate sensitive word in the multi-dimensional feature to obtain the perceptual fusion feature corresponding to the context of each candidate sensitive word specifically includes: For each candidate sensitive word, determine the fusion information of the candidate sensitive word in the multi-dimensional feature; Based on the fusion information, fuse the phoneme-level score, semantic-level score, scenario-level score, and sentiment-level score of the candidate sensitive word into the perceptual fusion feature corresponding to the context of the candidate sensitive word, thereby obtaining the perceptual fusion feature corresponding to the context of each candidate sensitive word.

[0009] In some embodiments, determining the path stability of the decoding path in speech recognition and the posterior probability of each Tibetan syllable in the Tibetan syllable sequence specifically includes: Determine the path stability interval of the decoding path in speech recognition; Based on the path stability interval, statistically calculate the proportion of stable frames in the framed speech signal; Determine the path stability of the decoding path in speech recognition through the proportion of stable frames; Obtain the posterior probability of each Tibetan syllable in the Tibetan syllable sequence from the Tibetan acoustic model.

[0010] In some embodiments, determining the decoding confidence of the Tibetan syllable sequence through the path stability and each posterior probability specifically includes: Initialize a linear probability model based on dynamic weight fusion; Use the path stability as the time continuity feature in the linear probability model; Use each posterior probability as the instantaneous reliability feature in the linear probability model; Use the linear probability model after feature input to perform weighted scoring on the multi-frame decoding results of the Tibetan syllable sequence to obtain the decoding confidence of the Tibetan syllable sequence.

[0011] In some embodiments, based on the decoding confidence and each perceptual fusion feature, perform fusion evaluation on each candidate sensitive word to obtain the context sensitivity of each candidate sensitive word specifically includes: For each candidate sensitive word, obtain the influence weight of the context on the candidate sensitive word; Based on the influence weight, fuse the perceptual fusion feature of the candidate sensitive word and the decoding confidence to obtain the context sensitivity of the candidate sensitive word, thereby obtaining the context sensitivity of each candidate sensitive word.

[0012] In a second aspect, the present application provides a sensitive word dynamic monitoring device in Tibetan speech recognition in a complex environment, including: An acquisition module, configured to acquire a Tibetan speech input stream to be monitored in a complex environment; A processing module, configured to segment the Tibetan speech input stream to obtain framed speech signals of Tibetan speech in a complex environment, and further extract multi-dimensional features of speech syllables in the context of the framed speech signals; The processing module is further configured to preliminarily match the framed speech signals with a sensitive word library of Tibetan speech to obtain multiple candidate sensitive words in the Tibetan speech input stream, and further perform multi-dimensional perception fusion on the phoneme-level score, semantic-level score, scene-level score, and emotion-level score of each candidate sensitive word in the multi-dimensional features to obtain the perception fusion features of the context corresponding to each candidate sensitive word; The processing module is further configured to perform speech recognition on the framed speech signals using a Tibetan acoustic model to generate a Tibetan syllable sequence, and further determine the path stability of the decoding path in the speech recognition and the posterior probability of each Tibetan syllable in the Tibetan syllable sequence, and determine the decoding confidence of the Tibetan syllable sequence through the path stability and each posterior probability; An execution module, configured to perform fusion evaluation on each candidate sensitive word based on the decoding confidence and each perception fusion feature to obtain the context sensitivity of each candidate sensitive word, and further perform hierarchical early warning on the Tibetan speech input stream in the complex environment through the context sensitivity.

[0013] In a third aspect, the present application provides a computer device, which includes a memory and a processor. The memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the computer device executes the sensitive word dynamic monitoring method in the above-mentioned Tibetan complex environment speech recognition.

[0014] In a fourth aspect, the present application provides a computer-readable storage medium, in which instructions or codes are stored. When the instructions or codes are run on a computer, the computer is caused to execute the sensitive word dynamic monitoring method in the above-mentioned Tibetan complex environment speech recognition.

[0015] The technical solutions provided by the disclosed embodiments of the present application have the following beneficial effects: In a method and device for dynamically monitoring sensitive words in Tibetan speech recognition in a complex environment provided by the present application, a Tibetan speech input stream to be monitored in the complex environment is obtained; the Tibetan speech input stream is segmented into segments to obtain framed speech signals of Tibetan speech in the complex environment, and then multi-dimensional features of speech syllables in the context of the framed speech signals are extracted; the framed speech signals are preliminarily matched with a sensitive word library of Tibetan speech to obtain multiple candidate sensitive words in the Tibetan speech input stream, and then the phoneme-level score, semantic-level score, scene-level score, and emotion-level score of each candidate sensitive word in the multi-dimensional features are subjected to multi-dimensional perceptual fusion to obtain the perceptual fusion features of the context corresponding to each candidate sensitive word; a Tibetan acoustic model is used to perform speech recognition on the framed speech signals to generate a Tibetan syllable sequence, and then the path stability of the decoding path in the speech recognition and the posterior probability of each Tibetan syllable in the Tibetan syllable sequence are determined, and the decoding confidence of the Tibetan syllable sequence is determined through the path stability and each posterior probability; based on the decoding confidence and each perceptual fusion feature, a fusion evaluation is performed on each candidate sensitive word to obtain the context sensitivity of each candidate sensitive word, and then a hierarchical early warning is performed on the Tibetan speech input stream in the complex environment through the context sensitivity.

[0016] It can be seen that in this application, based on the decoding confidence and each perceptual fusion feature, each candidate sensitive word is fused and evaluated to obtain the context sensitivity of each candidate sensitive word, and then the Tibetan speech input stream in a complex environment is classified and warned through the context sensitivity. First, determining the perceptual fusion feature can obtain a multi-dimensional context representation, which can significantly improve the accuracy and adaptability of sensitive word discrimination. The construction of the perceptual fusion feature provides a comprehensive context representation for Tibetan sensitive word recognition by integrating scores from four dimensions: phoneme level, semantic level, scene level, and emotion level. The phoneme-level score is based on the Tibetan pinyin fuzzy matching algorithm, which can capture the acoustic differences between dialect variants and standard pronunciations and solve the misjudgment problem of single phoneme matching. The semantic-level score uses a bidirectional long short-term memory network (Bi-LSTM) model to analyze the dependency relationship of syllable sequences and effectively distinguish the pronunciation similarity between homophones and sensitive words. The scene-level score combines acoustic environment classification to dynamically adjust the threshold for sensitive word discrimination. The emotion-level score quantifies the emotional polarity of words through a Tibetan emotion dictionary. Multi-dimensional fusion can improve the F1 value of sensitive word recognition, especially significantly reducing the false alarm rate in scenarios with mixed dialects. Then, determining the decoding confidence can obtain a reliability measure of the speech recognition result, thereby enhancing the robustness and dynamic decision-making ability of sensitive word monitoring. The decoding confidence is calculated through the decoding path stability and syllable posterior probability, which reflects the credibility of the speech recognition result. The decoding path stability measures the volatility of the output path of the acoustic model and can identify low-quality speech segments. The syllable posterior probability quantifies the certainty of each syllable in the joint decoding of the acoustic language model. Combining the decoding path stability and syllable posterior probability can automatically filter out low-confidence recognition results caused by environmental noise or dialect differences, thereby reducing the missed detection rate in a noisy environment. At the same time, the detection priority of key sensitive words is ensured through dynamic weight allocation. The decoding confidence can also trigger a hierarchical response mechanism: high-confidence sensitive words are warned in real time, and low-confidence results are transferred to the manual review process to achieve optimal resource allocation. In summary, based on the above solution, multi-dimensional fusion scoring of sensitive words in Tibetan complex environment speech recognition can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present application. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0018] Figure 1 is an exemplary flowchart of a method for dynamically monitoring sensitive words in Tibetan complex environment speech recognition according to some embodiments of the present application; Figure 2It is a schematic flowchart of Tibetan speech recognition shown in some embodiments of the present application; Figure 3 It is a schematic flowchart of determining path stability and posterior probability shown in some embodiments of the present application; Figure 4 It is a schematic structural diagram of a sensitive word dynamic monitoring device in Tibetan speech recognition in complex environments shown in some embodiments of the present application; Figure 5 It is a schematic structural diagram of a computer device for implementing a method for dynamically monitoring sensitive words in Tibetan speech recognition in complex environments shown in some embodiments of the present application. Detailed implementation manners

[0019] To better understand the technical solutions of the present application, the technical solutions of the present application will be described in detail below in conjunction with the accompanying drawings of the specification and specific implementation manners.

[0020] Refer to Figure 1 , this figure is an exemplary flowchart of a method for dynamically monitoring sensitive words in Tibetan speech recognition in complex environments shown in some embodiments of the present application. The method for dynamically monitoring sensitive words in Tibetan speech recognition in complex environments mainly includes the following steps: In step 101, obtain a Tibetan speech input stream to be monitored in a complex environment.

[0021] It should be noted that in the present application, the Tibetan speech input stream refers to continuously collected real-time Tibetan speech signals; the complex environment refers to a speech acquisition scenario that includes background noise (such as wind noise, human voices, mechanical sounds), overlapping of multiple speakers, dialect variations, or emotional intonation changes, which directly affects the robustness of speech recognition; specifically, when implemented, a microphone array is used to collect continuously collected Tibetan speech signals in a complex environment as the Tibetan speech input stream to be monitored.

[0022] In some embodiments, refer to Figure 2 as described, this figure is a schematic flowchart of Tibetan speech recognition shown in some embodiments of the present application. This figure describes the overall processing flow of Tibetan speech recognition, mainly including the integration and optimization processing of an acoustic model, a pronunciation dictionary, and a language model. After constructing a search space, a Viterbi Beam search is performed based on the input speech file to output the optimal recognition result. Specifically, the system first integrates the acoustic model, the pronunciation dictionary, and the language model, and sequentially performs the determinization and minimization steps to optimize the search space structure. Then, the input speech file and the optimized search space are input into the Viterbi Beam search module, and the initialization, score judgment, path pruning, and backtracking operations are completed in sequence, so as to obtain the final optimal recognition result. The entire process realizes the efficient modeling and accurate recognition of Tibetan speech signals.

[0023] In step 102, the Tibetan speech input stream is segmented into segments to obtain framed speech signals of Tibetan speech in a complex environment, and then multi-dimensional features of speech syllables in the context of the framed speech signals are extracted.

[0024] In some embodiments, the segmentation of the Tibetan speech input stream to obtain the framed speech signals of Tibetan speech in a complex environment can be implemented by the following steps: Normalize the Tibetan speech input stream to obtain a continuous speech stream to be monitored; Use a fixed duration to segment the continuous speech stream into multiple short-time signal segments; Determine the framed speech signals of Tibetan speech in a complex environment through all the short-time signal segments.

[0025] It should be noted that in this application, there are framed speech signals, continuous speech streams, and short-time signal segments.

[0026] When specifically implemented, first, the normalization of the Tibetan speech input stream to obtain a continuous speech stream to be monitored can be implemented in the following manner, that is: unify the sampling rate and normalize the amplitude of the Tibetan speech input stream, use an anti-aliasing filter to ensure signal quality, uniformly convert the speech from different sources to a sampling rate of 16 kHz, adjust the signal amplitude through automatic gain control to avoid numerical overflow or quantization error in subsequent processing, thereby completing the normalization adjustment of the Tibetan speech input stream, and use the normalized Tibetan speech input stream as the continuous speech stream to be monitored; then, the use of a fixed duration to segment the continuous speech stream into multiple short-time signal segments can be implemented in the following manner, that is: use an analysis window with a frame length of 25 ms and a frame shift of 10 ms to frame the continuous speech stream, each analysis frame is weighted by a Hamming window function to effectively reduce spectral leakage, and to maintain continuity, a 1 ms overlapping region is reserved between adjacent frames, and multiple short-time signal segments can be obtained; finally, the determination of the framed speech signals of Tibetan speech in a complex environment through all the short-time signal segments can be implemented in the following manner, that is: use the set of all short-time signal segments as the framed speech signals of Tibetan speech in a complex environment.

[0027] In some embodiments, the extraction of multi-dimensional features of speech syllables in the context of the framed speech signals can be implemented by the following steps: Set the window radius and window time interval of the time context window; For each speech syllable word in the framed speech signals, match the speech syllable word with a standard sensitive word according to the Tibetan pinyin fuzzy matching algorithm to obtain the phoneme-level score of the speech syllable word; Evaluate the semantic rationality of the speech syllable vocabulary in the current context based on the Tibetan semantic model to obtain the semantic-level score of the speech syllable vocabulary; Adjust the sensitivity threshold through the acoustic scene classification label and the speaker's dialect feature to obtain the scene-level score of the speech syllable vocabulary; Evaluate the emotional polarity intensity of the candidate sensitive word through the Tibetan emotion dictionary to obtain the emotion-level score of the speech syllable vocabulary; Furthermore, obtain the phoneme-level score, semantic-level score, scene-level score, and emotion-level score of each speech syllable vocabulary in the framed speech signal, and determine the multi-dimensional features of the speech syllable in the context of the framed speech signal through all the phoneme-level scores, semantic-level scores, scene-level scores, and emotion-level scores.

[0028] It should be noted that in this application, the multi-dimensional features; the time context window represents a fixed duration range before and after the current syllable, used to capture the semantic association between syllables; the phoneme-level score is a parameter quantifying the similarity in pronunciation between the current syllable and the standard sensitive word; the semantic-level score is an index reflecting the reasonable degree of the vocabulary appearance in the current context; the scene-level score is an environmental adaptation score affecting the final sensitivity judgment; the emotion-level score is a parameter quantifying the emotional intensity carried by the vocabulary.

[0029] In specific implementation, first, the window radius and window time interval of the time context window can be set in the following way, that is: set the window radius of the time context window to the size of the analysis window in the continuous speech stream, and obtain the interval from the start time to the end time of each time context window as the window time interval; second, for each speech syllable vocabulary in the framed speech signal, match the speech syllable vocabulary with the standard sensitive word according to the Tibetan pinyin fuzzy matching algorithm, and the phoneme-level score of the speech syllable vocabulary can be obtained in the following way, that is: for each speech syllable vocabulary in the framed speech signal, match the speech syllable vocabulary with the standard sensitive word according to the Tibetan pinyin fuzzy matching algorithm, and use the matching result as the phoneme-level score of the speech syllable vocabulary. This Tibetan pinyin fuzzy matching algorithm is to count the number of pinyin modifications required for the speech syllable vocabulary to be converted into the standard sensitive word as the matching result; then, evaluate the semantic rationality of the speech syllable vocabulary in the current context based on the Tibetan semantic model to obtain the semantic-level score of the speech syllable vocabulary in the following way, that is: calculate the co-occurrence probability of the candidate sensitive word in the front and back windows based on the pre-trained Tibetan bidirectional long short-term memory network model, calculate the semantic deviation degree in combination with the Tibetan grammar rules, and output a rationality score within the range of 0-1.

[0030] Furthermore, in specific implementation, the sensitivity threshold is adjusted by the acoustic scene classification label and the speaker's dialect feature. The scene-level scoring of the speech syllable vocabulary can be implemented in the following way, that is: match the preset sensitivity coefficient according to the acoustic scene classification label, and superimpose the threshold offset corresponding to the speaker's dialect feature. Thus, the result of dividing the product of the sensitivity coefficient and the threshold offset by the sensitivity threshold is used as the scene-level scoring of the speech syllable vocabulary; then, the emotional polarity intensity of the candidate sensitive words is evaluated through the Tibetan emotion dictionary. The emotional-level scoring of the speech syllable vocabulary can be implemented in the following way, that is: query the Tibetan emotion dictionary to obtain the basic emotional polarity and intensity value of the candidate sensitive words, and combine the emotional word density within the time context window for weighting, and finally output the normalized emotional tendency scoring. Among them, the emotional word density can be obtained by counting the number of occurrences of the candidate sensitive words within the time context window; finally, furthermore, the phoneme-level scoring, semantic-level scoring, scene-level scoring, and emotional-level scoring of each speech syllable vocabulary in the framed speech signal are obtained. Determining the multi-dimensional features of the speech syllables in the context of the framed speech signal through all the phoneme-level scoring, semantic-level scoring, scene-level scoring, and emotional-level scoring can be implemented in the following way, that is: the phoneme-level scoring, semantic-level scoring, scene-level scoring, and emotional-level scoring of each speech syllable vocabulary in the framed speech signal can be obtained through the above method. Thus, the set of the phoneme-level scoring, semantic-level scoring, scene-level scoring, and emotional-level scoring is used as the multi-dimensional features of the speech syllables in the context of the framed speech signal.

[0031] It should be noted that in this application, the Tibetan bidirectional long short-term memory network model is a deep learning architecture designed specifically for Tibetan sequence modeling. It captures the context semantic association through bidirectional time series analysis. The forward and backward long short-term memory networks are used to process the speech sequence respectively. The forward network analyzes the syllable stream in chronological order to learn the relationship between the current word and the previous text; the backward network processes in reverse order to capture the association between the current word and the following text. After the hidden layer outputs of the two networks are concatenated, the co-occurrence probability of each candidate sensitive word within the specified window is calculated through the fully connected layer. The Tibetan grammar rules are learned in the pre-training stage. During actual inference, the predicted probability is compared with the grammar rule library, and the score of the candidate words that violate the conventional grammar structure is reduced. Finally, the semantic rationality scoring of 0-1 is output through normalization. The higher the semantic rationality scoring, the greater the possibility that the word appears in the context and conforms to the Tibetan grammar specification.

[0032] In step 103, the framed speech signal is initially matched with the sensitive word library of the Tibetan speech to obtain multiple candidate sensitive words in the Tibetan speech input stream. Furthermore, the phoneme-level scoring, semantic-level scoring, scene-level scoring, and emotional-level scoring of each candidate sensitive word in the multi-dimensional features are subjected to multi-dimensional perceptual fusion to obtain the perceptual fusion features corresponding to the context of each candidate sensitive word.

[0033] In some embodiments, the preliminary matching of the framed speech signal and the sensitive word library of Tibetan speech to obtain multiple candidate sensitive words in the Tibetan speech input stream can be implemented by the following steps: Obtain the sensitive word library of Tibetan speech and the candidate sensitive threshold; Perform an edit distance matching between each speech syllable word in the framed speech signal and the sensitive word library to obtain the Tibetan phonetic edit distance of each speech syllable word; Screen out multiple candidate sensitive words in the Tibetan speech input stream from the framed speech signal based on each Tibetan phonetic edit distance and the candidate sensitive threshold.

[0034] It should be noted that in this application, a candidate sensitive word is a syllable word that is preliminarily matched successfully in the Tibetan speech input stream; the sensitive word library refers to a predefined set of Tibetan sensitive words, which includes standard spellings, common variants, and pinyin transcriptions; the candidate sensitive threshold is a similarity critical value for determining whether a syllable is a sensitive word, and a syllable with a value lower than this candidate sensitive threshold is regarded as a non-sensitive word; the Tibetan phonetic edit distance represents the degree of difference between a syllable word and the pinyin of a sensitive word, and is calculated by the number of operations of adding, deleting, or modifying characters. The smaller the Tibetan phonetic edit distance, the more similar they are.

[0035] Specifically, when implementing, first, obtaining the sensitive word library of Tibetan speech and the candidate sensitive threshold can be achieved in the following way, that is: obtain the sensitive word library of Tibetan speech and the candidate sensitive threshold from the console of the Tibetan recognition device; then, performing an edit distance matching between each speech syllable word in the framed speech signal and the sensitive word library to obtain the Tibetan phonetic edit distance of each speech syllable word can be achieved in the following way, that is: for each speech syllable word in the framed speech signal, perform pinyin conversion on the speech syllable word, and use the dynamic programming algorithm to calculate the edit distance between its pinyin and each word in the sensitive word library one by one. When calculating, it can be optimized according to the characteristics of Tibetan, setting a higher weight for the difference in initials, a lower weight for the difference in finals, and the lowest weight for the difference in tones, and record the minimum value in the pinyin edit distance as the Tibetan phonetic edit distance of the speech syllable word. Through the above method, the Tibetan phonetic edit distance of each speech syllable word can be obtained; finally, screening out multiple candidate sensitive words in the Tibetan speech input stream from the framed speech signal based on each Tibetan phonetic edit distance and the candidate sensitive threshold can be achieved in the following way, that is: regard the speech syllable words in the framed speech signal with a Tibetan phonetic edit distance less than the candidate sensitive threshold as candidate sensitive words, and then multiple candidate sensitive words in the Tibetan speech input stream can be obtained.

[0036] In some embodiments, the multi-dimensional perception fusion of the phoneme-level score, semantic-level score, scenario-level score, and sentiment-level score of each candidate sensitive word in the multi-dimensional features to obtain the perception fusion feature corresponding to the context of each candidate sensitive word can be implemented by the following steps: For each candidate sensitive word, determine the fusion information of the candidate sensitive word in the multi-dimensional features; Based on the fusion information, fuse the phoneme-level score, semantic-level score, scenario-level score, and sentiment-level score of the candidate sensitive word into the perception fusion feature corresponding to the context of the candidate sensitive word, thereby obtaining the perception fusion feature corresponding to the context of each candidate sensitive word.

[0037] It should be noted that in this application, the perception fusion feature is an index reflecting the overall sensitivity degree of the candidate sensitive word in the context; specifically, when implementing, first, for each candidate sensitive word, the fusion information of the candidate sensitive word in the multi-dimensional features can be implemented in the following manner, that is: obtain the dimension weights of the candidate sensitive word in each dimension from the console of the Tibetan recognition device, where the phoneme dimension weight is defaulted to 0.4, the semantic dimension weight is defaulted to 0.3, the scenario dimension weight is defaulted to 0.2, and the sentiment dimension weight is defaulted to 0.1, so as to use the set of all dimension weights as the fusion information of the candidate sensitive word in the multi-dimensional features, and this fusion information represents the influence degree of each dimension on the overall sensitivity degree of the candidate sensitive word in the context; then, based on the fusion information, fusing the phoneme-level score, semantic-level score, scenario-level score, and sentiment-level score of the candidate sensitive word into the perception fusion feature corresponding to the context of the candidate sensitive word, and thereby obtaining the perception fusion feature corresponding to the context of each candidate sensitive word can be implemented in the following manner, that is: use the dimension weights of each dimension in the fusion information to calculate the weighted sum of the phoneme-level score, semantic-level score, scenario-level score, and sentiment-level score as the perception fusion feature corresponding to the context of the candidate sensitive word, and through the above method, the perception fusion feature corresponding to the context of each candidate sensitive word can be obtained.

[0038] In step 104, use the Tibetan acoustic model to perform speech recognition on the framed speech signal to generate a Tibetan syllable sequence, and then determine the path stability of the decoding path in the speech recognition and the posterior probability of each Tibetan syllable in the Tibetan syllable sequence, and determine the decoding confidence of the Tibetan syllable sequence through the path stability and each posterior probability.

[0039] In some embodiments, a Tibetan acoustic model is used to perform speech recognition on the framed speech signal to generate a sequence of Tibetan syllables. It should be noted that in this application, the Tibetan acoustic model is a statistical model specifically trained for the characteristics of Tibetan speech. Its core function is to map the acoustic features of the speech signal to the corresponding Tibetan phoneme or syllable sequence. This Tibetan acoustic model adopts a hybrid modeling method that combines a deep neural network and a Hidden Markov Model (HMM). The neural network is responsible for learning the common acoustic patterns of the three major Tibetan dialects (U-Tsang, Amdo, and Kham), and extracting discriminative feature representations through multi-layer non-linear transformations. The HMM, on the other hand, models the temporal dynamic characteristics of Tibetan phonemes. Each phoneme corresponds to an HMM state, and the temporal relationship between phonemes is described by the state transition probability. During training, the maximum likelihood criterion is adopted, and the network parameters are optimized using the labeled Tibetan speech data through the backpropagation algorithm, so that the Tibetan acoustic model can accurately output the posterior probability of each speech frame belonging to each Tibetan phoneme. In the decoding stage, this Tibetan acoustic model works in cooperation with the Tibetan language model, and combines the Viterbi algorithm to search for the optimal syllable path to achieve robust recognition of complex acoustic environments (such as: plateau noise, reverberation).

[0040] In some embodiments, the path stability of the decoding path in speech recognition and the posterior probability of each Tibetan syllable in the sequence of Tibetan syllables are determined with reference to Figure 3 As described, this figure is a schematic flowchart of determining path stability and posterior probability in some embodiments of this application. In this embodiment, the determination of path stability and posterior probability can be implemented by the following steps: In step 1041, the path stability interval of the decoding path in speech recognition is determined; In step 1042, based on the path stability interval, the proportion of stable frames in the framed speech signal is statistically calculated; In step 1043, the path stability of the decoding path in speech recognition is determined through the proportion of stable frames; In step 1044, the posterior probability of each Tibetan syllable in the sequence of Tibetan syllables is obtained from the Tibetan acoustic model.

[0041] It should be noted that in this application, the posterior probability represents the probability value of the current speech frame corresponding to a specified Tibetan syllable, and this posterior probability reflects the recognition confidence of a single frame; the path stability is an index comprehensively measuring the consistency of recognition results in the decoding path, and the higher the path stability, the more reliable the recognition result; the path stability interval; the proportion of stable frames refers to the ratio of the number of stable frames to the total number of frames, and this proportion of stable frames can quantify the recognition stability of the entire framed speech signal.

[0042] In specific implementation, first, the path stability interval of the decoding path in speech recognition can be implemented in the following way: set the minimum stability interval length to 3 frames. When the same syllable is recognized in 3 consecutive frames in speech recognition, it is marked as a stability interval. The interval boundary can be dynamically adjusted by means of a sliding window to ensure that all stable segments can be captured. For each detected stability interval, record its start frame, end frame, and the corresponding syllable label, and then the path stability interval of the decoding path in speech recognition can be obtained. Second, the proportion of stable frames in the framed speech signal can be statistically calculated based on the path stability interval in the following way: count the number of frames of the framed speech signal in the path stability interval, and calculate the ratio of this number of frames to the total number of frames of the framed speech signal as the proportion of stable frames in the framed speech signal. Then, the path stability of the decoding path in speech recognition can be determined by the proportion of stable frames in the following way: use a normalization algorithm (e.g., min-max normalization) to normalize the proportion of stable frames and use it as the path stability of the decoding path in speech recognition. Finally, the posterior probability of each Tibetan syllable in the Tibetan syllable sequence can be obtained from the Tibetan acoustic model in the following way: for each Tibetan syllable in the Tibetan syllable sequence, obtain the posterior probability value of the Tibetan syllable from the Tibetan acoustic model. For the frames within the path stability interval, calculate the mean value of the posterior probability values of all frames in the path stability interval where the Tibetan syllable is located as the posterior probability of the Tibetan syllable; for the frames not within the path stability interval, directly use the single-frame probability value as the posterior probability of the Tibetan syllable. Through the above method, the posterior probability of each Tibetan syllable in the Tibetan syllable sequence can be obtained.

[0043] In some embodiments, the decoding confidence of the Tibetan syllable sequence can be determined by the path stability and each posterior probability through the following steps: Initialize a linear probability model based on dynamic weight fusion; Use the path stability as the time continuity feature in the linear probability model; Use each posterior probability as the instantaneous reliability feature in the linear probability model; Use the linear probability model after feature input to perform weighted scoring on the multi-frame decoding results of the Tibetan syllable sequence to obtain the decoding confidence of the Tibetan syllable sequence.

[0044] It should be noted that in this application, the decoding confidence is a metric for evaluating the reliability of Tibetan syllable sequence recognition. Specifically, when the Tibetan acoustic model decodes the framed speech signal, it is a reliability metric value calculated through the dynamic weight fusion model based on the stability of the decoding path and the posterior probability of the Tibetan syllable sequence. In this application, the linear probability model is a mathematical model based on weighted summation, which is used to fuse multi-dimensional features and output a probabilistic score. In the weighted scoring, the linear probability model takes the time continuity feature and the instantaneous reliability feature as two key input dimensions, and performs a linear combination through a preset weight coefficient. First, the two features are normalized respectively to eliminate the dimensional difference; then, according to the characteristics of Tibetan speech, a fixed weight of 0.4 is assigned to the time continuity, and a weight of 0.6 is assigned to the instantaneous reliability to reflect the dominant role of the single-frame recognition confidence; finally, the decoding confidence is calculated through the weighted summation formula, and the result is mapped to the [0,1] interval through normalization processing to form an output with probabilistic interpretability. The linear characteristic of this linear probability model ensures the calculation efficiency and is suitable for real-time processing scenarios. At the same time, the weight coefficient can be dynamically adjusted according to different dialect environments, taking into account both flexibility and accuracy.

[0045] In step 105, based on the decoding confidence and each perceptual fusion feature, each candidate sensitive word is fused and evaluated to obtain the context sensitivity of each candidate sensitive word, and then the Tibetan speech input stream in a complex environment is graded and warned through the context sensitivity.

[0046] In some embodiments, the fusion evaluation of each candidate sensitive word based on the decoding confidence and each perceptual fusion feature to obtain the context sensitivity of each candidate sensitive word can be implemented by the following steps: For each candidate sensitive word, obtain the influence weight of the context on the candidate sensitive word; Based on the influence weight, fuse the perceptual fusion feature of the candidate sensitive word and the decoding confidence to obtain the context sensitivity of the candidate sensitive word, and then obtain the context sensitivity of each candidate sensitive word.

[0047] In specific implementation, first, for each candidate sensitive word, obtaining the influence weight of the context on the candidate sensitive word can be achieved in the following manner, that is: for each candidate sensitive word, obtain the influence weight of the context on the candidate sensitive word from the console of the Tibetan recognition device; then, based on the influence weight, fuse the perceptual fusion feature of the candidate sensitive word and the decoding confidence to obtain the context sensitivity of the candidate sensitive word. Furthermore, obtaining the context sensitivity of each candidate sensitive word can be achieved in the following manner, that is: use the fusion formula for feature fusion, that is: context sensitivity = perceptual fusion feature * influence weight + decoding confidence * (1 - influence weight), and the context sensitivity of the candidate sensitive word can be obtained. Through the above method, the context sensitivity of each candidate sensitive word can be obtained.

[0048] In some embodiments, grading and warning the Tibetan speech input stream in a complex environment through the context sensitivity can be achieved in the following manner, that is: obtain the warning level mapping table from the console of the Tibetan recognition device, and obtain the warning level corresponding to the context sensitivity from this warning level mapping table, so as to use this warning level as the warning level of the Tibetan speech input stream in the complex environment.

[0049] In addition, on the other hand of the present application, in some embodiments, the present application provides a sensitive word dynamic monitoring device in Tibetan speech recognition in a complex environment. Refer to Figure 4 , this figure is a schematic structural diagram of the sensitive word dynamic monitoring device in Tibetan speech recognition in a complex environment according to some embodiments of the present application. The sensitive word dynamic monitoring device in Tibetan speech recognition in a complex environment includes: an acquisition module 201, a processing module 202, and an execution module 203, which are described as follows: The acquisition module 201, in the present application, the acquisition module 201 is mainly used to acquire the Tibetan speech input stream to be monitored in a complex environment; The processing module 202, in the present application, the processing module 202 is used to segment the Tibetan speech input stream into segments to obtain the framed speech signals of the Tibetan speech in the complex environment, and further extract the multi-dimensional features of the speech syllables in the context of the framed speech signals; It should be noted that the processing module 202 is further used to preliminarily match the framed speech signals with the sensitive word library of the Tibetan speech to obtain multiple candidate sensitive words in the Tibetan speech input stream, and further perform multi-dimensional perceptual fusion on the phoneme-level score, semantic-level score, scene-level score, and emotion-level score of each candidate sensitive word in the multi-dimensional features to obtain the perceptual fusion feature corresponding to the context of each candidate sensitive word; In addition, the processing module 202 is further configured to perform speech recognition on the framed speech signal using a Tibetan acoustic model to generate a Tibetan syllable sequence, and further determine the path stability of the decoding path in the speech recognition and the posterior probability of each Tibetan syllable in the Tibetan syllable sequence, and determine the decoding confidence of the Tibetan syllable sequence based on the path stability and each posterior probability; An execution module 203. In this application, the execution module 203 is mainly configured to perform fusion evaluation on each candidate sensitive word based on the decoding confidence and each perceptual fusion feature to obtain the context sensitivity of each candidate sensitive word, and further perform hierarchical early warning on the Tibetan speech input stream in a complex environment based on the context sensitivity.

[0050] The above has introduced in detail the examples of the method and device for dynamically monitoring sensitive words in Tibetan speech recognition in a complex environment provided by the embodiments of the present application. It can be understood that, in order to implement the above functions, the corresponding device includes the corresponding hardware structure and / or software module for executing each function. Those skilled in the art should easily realize that, combining the units and algorithm steps of each example described in the embodiments disclosed in this article, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a certain function is executed in the way of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0051] In some embodiments, the present application further provides a computer device, which includes a memory and a processor. The memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the computer device executes the method for dynamically monitoring sensitive words in Tibetan speech recognition in a complex environment described above.

[0052] In some embodiments, refer to Figure 5 , the dotted line in this figure indicates that the unit or module is optional. This figure is a schematic structural diagram of a computer device for implementing the method for dynamically monitoring sensitive words in Tibetan speech recognition in a complex environment provided by the embodiments of the present application. The method for dynamically monitoring sensitive words in Tibetan speech recognition in a complex environment described in the above embodiments can be implemented by Figure 5 the computer device shown, which includes at least one processor 301, a memory 302, and at least one communication unit 305. The computer device can be a terminal device, a server, or a chip.

[0053] The processor 301 can be a general-purpose processor or a special-purpose processor. For example, the processor 301 can be a central processing unit (CPU), and the CPU can be used to control a computer device, execute software programs, and process data of the software programs. The computer device can also include a communication unit 305 for realizing signal input (reception) and output (transmission).

[0054] For example, the computer device can be a chip, and the communication unit 305 can be the input and / or output circuit of the chip, or the communication unit 305 can be the communication interface of the chip. The chip can be a component of a terminal device, a network device, or other devices.

[0055] For another example, the computer device can be a terminal device or a server, and the communication unit 305 can be the transceiver of the terminal device or the server, or the communication unit 305 can be the transceiver circuit of the terminal device or the server.

[0056] The computer device can include one or more memories 302 on which a program 304 is stored. The program 304 can be run by the processor 301 to generate instructions 303, enabling the processor 301 to execute the methods described in the above method embodiments according to the instructions 303. Optionally, data (such as a target audit model) can also be stored in the memory 302. Optionally, the processor 301 can also read the data stored in the memory 302. The data can be stored at the same storage address as the program 304, or it can be stored at a different storage address from the program 304.

[0057] The processor 301 and the memory 302 can be set separately or integrated together. For example, they can be integrated on a system on chip (SOC) of a terminal device.

[0058] It should be understood that the steps of the above method embodiments can be completed by the logic circuit in the form of hardware or the instructions in the form of software in the processor 301. The processor 301 can be a CPU, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices. For example, discrete gate, transistor logic devices, or discrete hardware components.

[0059] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0060] For example, in some embodiments, the present application further provides a computer-readable storage medium, in which instructions or codes are stored. When the instructions or codes are run on a computer, the computer is caused to implement the above-mentioned sensitive word dynamic monitoring method in Tibetan complex environment speech recognition.

[0061] Although the preferred embodiments of the present application have been described, those skilled in the art can make additional changes and modifications to these embodiments once they learn the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications that fall within the scope of the present application.

[0062] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalent technologies, the present application is also intended to include these modifications and variations.

Claims

1. A method for dynamically monitoring sensitive words in Tibetan speech recognition in complex environments, characterized in that, It includes the following steps: Obtain the Tibetan speech input stream to be monitored in a complex environment; Segment the Tibetan speech input stream to obtain the framed speech signals of Tibetan speech in a complex environment, and then extract the multi-dimensional features of speech syllables in the context of the framed speech signals; Perform a preliminary match between the framed speech signals and the sensitive word library of Tibetan speech to obtain multiple candidate sensitive words in the Tibetan speech input stream, and then perform multi-dimensional perceptual fusion on the phoneme-level score, semantic-level score, scene-level score, and sentiment-level score of each candidate sensitive word in the multi-dimensional features to obtain the perceptual fusion features corresponding to the context of each candidate sensitive word; Use the Tibetan acoustic model to perform speech recognition on the framed speech signals to generate a sequence of Tibetan syllables, and then determine the path stability of the decoding path in speech recognition and the posterior probability of each Tibetan syllable in the sequence of Tibetan syllables, and determine the decoding confidence of the sequence of Tibetan syllables through the path stability and each posterior probability; Based on the decoding confidence and each perceptual fusion feature, perform a fusion evaluation on each candidate sensitive word to obtain the context sensitivity of each candidate sensitive word, and then perform a hierarchical warning on the Tibetan speech input stream in a complex environment through the context sensitivity; 2. The method according to claim 1, characterized in that, Specifically, extracting the multi-dimensional features of speech syllables in the context of the framed speech signals includes: Set the window radius and window time interval of the time context window; For each speech syllable word in the framed speech signals, match the speech syllable word with the standard sensitive word according to the Tibetan pinyin fuzzy matching algorithm to obtain the phoneme-level score of the speech syllable word; Evaluate the semantic rationality of the speech syllable word in the current context based on the Tibetan semantic model to obtain the semantic-level score of the speech syllable word; Adjust the sensitivity threshold through the acoustic scene classification label and the speaker's dialect features to obtain the scene-level score of the speech syllable word; Evaluate the emotional polarity intensity of the candidate sensitive word through the Tibetan emotion dictionary to obtain the sentiment-level score of the speech syllable word; Furthermore, obtain the phoneme-level score, semantic-level score, scene-level score, and sentiment-level score of each speech syllable word in the framed speech signals, and determine the multi-dimensional features of speech syllables in the context of the framed speech signals through all the phoneme-level scores, semantic-level scores, scene-level scores, and sentiment-level scores; 3. The method according to claim 1, wherein Specifically, performing a preliminary match between the framed speech signals and the sensitive word library of Tibetan speech to obtain multiple candidate sensitive words in the Tibetan speech input stream includes: Obtain the sensitive word library of Tibetan speech and the candidate sensitive threshold; Perform an edit distance match between each speech syllable word in the framed speech signals and the sensitive word library to obtain the Tibetan pinyin edit distance of each speech syllable word; Select multiple candidate sensitive words in the Tibetan speech input stream from the framed speech signals through each Tibetan pinyin edit distance and the candidate sensitive threshold; 4. The method according to claim 1, wherein Specifically, performing multi-dimensional perceptual fusion on the phoneme-level score, semantic-level score, scene-level score, and sentiment-level score of each candidate sensitive word in the multi-dimensional features to obtain the perceptual fusion features corresponding to the context of each candidate sensitive word includes: For each candidate sensitive word, determine the fusion information of the candidate sensitive word in the multi-dimensional features; Based on the fusion information, fuse the phoneme-level score, semantic-level score, scenario-level score, and sentiment-level score of the candidate sensitive word into the perceptual fusion feature corresponding to the context of the candidate sensitive word, thereby obtaining the perceptual fusion feature corresponding to the context of each candidate sensitive word.

5. The method according to claim 1, characterized in that, Determining the path stability of the decoding path in speech recognition and the posterior probability of each Tibetan syllable in the Tibetan syllable sequence specifically includes: Determine the path stability interval of the decoding path in speech recognition; Based on the path stability interval, count the proportion of stable frames in the framed speech signal; Determine the path stability of the decoding path in speech recognition through the proportion of stable frames; Obtain the posterior probability of each Tibetan syllable in the Tibetan syllable sequence from the Tibetan acoustic model.

6. The method according to claim 1, wherein Determining the decoding confidence of the Tibetan syllable sequence through the path stability and each posterior probability specifically includes: Initialize a linear probability model based on dynamic weight fusion; Use the path stability as the time continuity feature in the linear probability model; Use each posterior probability as the instantaneous reliability feature in the linear probability model; Use the linear probability model after feature input to perform weighted scoring on the multi-frame decoding results of the Tibetan syllable sequence to obtain the decoding confidence of the Tibetan syllable sequence.

7. The method according to claim 1, characterized in that, Based on the decoding confidence and each perceptual fusion feature, perform a fusion evaluation on each candidate sensitive word to obtain the context sensitivity of each candidate sensitive word, specifically including: For each candidate sensitive word, obtain the influence weight of the context on the candidate sensitive word; Based on the influence weight, fuse the perceptual fusion feature of the candidate sensitive word and the decoding confidence to obtain the context sensitivity of the candidate sensitive word, thereby obtaining the context sensitivity of each candidate sensitive word.

8. A sensitive word dynamic monitoring device in Tibetan speech recognition in complex environments, characterized in that, Including: An acquisition module for acquiring the Tibetan speech input stream to be monitored in a complex environment; A processing module for segmenting the Tibetan speech input stream to obtain the framed speech signal of the Tibetan speech in the complex environment, and further extracting the multi-dimensional features of the speech syllables in the context of the framed speech signal; The processing module is further configured to preliminarily match the framed speech signal with the sensitive word library of the Tibetan speech to obtain multiple candidate sensitive words in the Tibetan speech input stream, and further perform multi-dimensional perceptual fusion on the phoneme-level score, semantic-level score, scenario-level score, and sentiment-level score of each candidate sensitive word in the multi-dimensional features to obtain the perceptual fusion feature corresponding to the context of each candidate sensitive word; The processing module is further configured to use the Tibetan acoustic model to perform speech recognition on the framed speech signal to generate a Tibetan syllable sequence, and further determine the path stability of the decoding path in speech recognition and the posterior probability of each Tibetan syllable in the Tibetan syllable sequence, and determine the decoding confidence of the Tibetan syllable sequence through the path stability and each posterior probability; An execution module for performing a fusion evaluation on each candidate sensitive word based on the decoding confidence and each perceptual fusion feature to obtain the context sensitivity of each candidate sensitive word, and further performing a hierarchical warning on the Tibetan speech input stream in the complex environment through the context sensitivity.

9. A computer device, characterized in that, The computer device includes a memory and a processor. The memory is used to store a computer program, and the processor is used to call and run the computer program from the memory, so that the computer device executes the sensitive word dynamic monitoring method in Tibetan complex environment speech recognition according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Instructions or code are stored in the computer-readable storage medium. When the instructions or code run on a computer, the computer is caused to implement the sensitive word dynamic monitoring method in Tibetan complex environment speech recognition according to any one of claims 1 to 7.

Images