194 results about "Word model" patented technology

The error rate of a pronunciation guesser that guesses the phonetic spelling of words used in speech recognition is improved by causing its training to weigh letter-to-phoneme mappings used as data in such training as a function of the frequency of the words in which such mappings occur. Preferably the ratio of the weight to word frequency increases as word frequencies decreases. Acoustic phoneme models for use in speech recognition with phonetic spellings generated by a pronunciation guesser that makes errors are trained against word models whose phonetic spellings have been generated by a pronunciation guesser that makes similar errors. As a result, the acoustic models represent blends of phoneme sounds that reflect the spelling errors made by the pronunciation guessers. Speech recognition enabled systems are made by storing in them both a pronunciation guesser and a corresponding set of such blended acoustic models.

Out-of-vocabulary word models for a speech recognizer vocabulary are generated by forming phonemic transcriptions (phonetic baseforms) of user's utterances in terms of existing reference phonemes by using a speech recognition algorithm to match input sub-word feature sample sequences to suitably-constrained allowable sequences of existing reference phoneme features. The resultant new-vocabulary-word phonetic baseform models are stored for subsequent speech recognition using the same recognition algorithm.

A system and method enabling acoustic barge-in during a voice prompt in a communication system. An acoustic prompt model is trained to represent the system prompt using the specific speech signal of the prompt. The acoustic prompt model is utilized in a speech recognizer in parallel with the recognizer's active vocabulary words to suppress the echo of the prompt within the recognizer. The speech recognizer may also use a silence model and traditional garbage models such as noise models and out-of-vocabulary word models to reduce the likelihood that noises and out-of-vocabulary words in the user utterance will be mapped erroneously onto active vocabulary words.

Methods, systems, and apparatuses are described for isolated word training and detection. Isolated word training devices and systems are provided in which a user may provide a wake-up phrase from 1 to 3 times to train the device or system. A concatenated phoneme model of the user-provided wake-up phrase may be generated based on the provided wake-up phrase and a pre-trained phoneme model database. A word model of the wake-up phrase may be subsequently generated from the concatenated phoneme model and the provided wake-up phrase. Once trained, the user-provided wake-up phrase may be used to unlock the device or system and / or to wake up the device or system from a standby mode of operation. The word model of the user-provided wake-up phrase may be further adapted based on additional provisioning of the wake-up phrase.

The invention discloses an intelligent equipment awakening method and device, a server and intelligent equipment. The method comprises the following steps of: obtaining a voice signal sent by the intelligent equipment; recognizing the voice signal so as to determine a recognition result corresponding to the voice signal; and distributing an awakening instruction to the intelligent equipment when determining that a matching degree between the recognition result and any awakening word model corresponding to the intelligent equipment is greater than a threshold value. According to the method, voice signals are recognized by utilizing the server and awakening control is carried out on the intelligent equipment, so that the consumption of hardware performance in the awakening process is reduced, the users can awaken the intelligent equipment by utilizing self-defined awakening words, the awakening words are enriched, the awakening flexibility is improved, individualized demands of the userscan be better satisfied and the user experience is improved.

A Statistical Language Model (SLM) that can be used in an ASR for Interactive Voice Response (IVR) systems in general and Natural Language Speech Applications (NLSAs) in particular can be created by first manually producing a brief description in text for each task that can be performed in an NLSA. These brief descriptions are then analyzed, in one embodiment, to generate spontaneous speech utterances based pre-filler patterns and a skeletal set of content words. The pre-filler patterns are in turn used with Part-of-Speech (POS) tagged conversations from a spontaneous speech corpus to generate a set of pre-filler phrases. The skeletal set of content words is used with an electronic lexico-semantic database and with a thesaurus-based content word extraction process to generate a more extensive list of content words. The pre-filler phrases and content words set, thus generated, are combined into utterances using a lexico-semantic resource based process. In one embodiment, a lexico-semantic statistical validation process is used to correct and / or add the automatically generated utterances to the database of expected utterances. The system requires a minimum amount of human intervention and no prior knowledge regarding the expected user utterances, and the WWW is used to validate the word models. The system requires a minimum amount of human intervention and no prior knowledge regarding the expected user utterances in response to a particular prompt.

The invention discloses a commodity target word oriented emotional tendency analysis method, which belongs to the field of the analysis processing of online shopping commodity reviews. The method comprises the following four steps that: 1: corpus preprocessing: carrying out word segmentation on a dataset, and converting a category label into a vector form according to a category number; 2: word vector training: training review data subjected to the word segmentation through a CBOW (Continuous Bag-of-Words Model) to obtain a word vector; 3: adopting a neural network structure, and using an LSTM(Long Short Term Memory) network model structure to enable the network to pay attention to whole-sentence contents; and 4: review sentence emotion classification: taking the output of the neural network as the input of a Softmax function to obtain a final result. By use of the method, semantic description in a semantic space is more accurate, the data is trained through the neural network so as to optimize the weight and the offset parameter in the neural network, parameters trained after continuous iteration make a loss value minimum, at the time, the trained parameters are used for traininga test set, and therefore, higher accuracy can be obtained.

The invention relates to a knowledge graph completion method based on entity description and relationship path, which comprises the steps of S1, establishing a continuous bag-of-words model based on entity text description, and performing vector representation on entity description of entities in a knowledge graph by utilizing the continuous bag-of-words model to obtain a description-based vector;S2, according to a conversion-based model between the entity vector and the relation vector and between the entity vector and the relation path, obtaining a conversion-based model, establishing a score function of the relation triad (h, r, t) and the path triad (h, p, t) and a loss function of the score function, learning vector representation of an entity, a relation and a path by minimizing theloss function, and learning vector representation based on a structure; and S3, adopting the learned entity vector representation, obtaining representation results in the vector space in different tasks, and complementing the knowledge graph or mining the potential relationship.

The invention provides a service intelligence navigation method, including the steps of: carrying out grammar matching based on metadata for word segmentation result of query proposed by a user; finding out a matched word model mode according to the word segmentation result and acquiring the service classification to which the matched word model mode belongs; carrying out matching search to the word segmentation result and descriptive information or business information and obtaining a first candidate service classification list; acquiring service from the only candidate service classification of the first candidate service classification list and returning the service to the user; if the candidate service classifications are two or more than two, carrying out fuzzy reading of non-login string in the query and obtaining a second candidate service classification list; carrying out fusion operation according to the first candidate service classification list and the second candidate service classification list and obtaining common candidate service classification; if the common candidate service classification is only one, returning the service to the user; if the common candidate service classification does not exist or is more than one, then showing the navigation is a failure.

The invention discloses a voice wake-up method and system. The method comprises the steps of step 1) acquiring input real-time audio stream data, detecting the input real-time audio stream data by using a voice detection algorithm, if voice data is found, starting a wake-up word detection algorithm, otherwise, stopping the wake-up word detection algorithm, and continuing to detect; step 2) using the wake-up word detection algorithm to analyze and detect an input voice data stream, and calculating the similarity between the voice data and a given wake-up word model; if the similarity is greaterthan a given threshold T1, considering that wake-up words are detected, starting a wake-up word confirmation algorithm, introducing the voice data of assumed wake-up words, and otherwise, continuously detecting the wake-up words until the voice data stream ends; step 3) using the wake-up word confirmation algorithm to analyze and calculate the obtained voice data stream which is judged as the wake-up words by the wake-up word detection algorithm, and calculating the confidence coefficient of the wake-up words on a voice segment.

The invention discloses an emotion analysis method for Chinese texts based on computer information processing technology. Comments on Chinese products are subjected to word segmentation. By utilizing a bag-of-words model, vector representations of product comments are generated. The vector of every comment is inputted to a visible unit of a limited Boltzmann machine (RBM) in deep learning. Sentimental characteristics of Chinese texts are extracted by the RBM and the extracted emotional characteristics are inputted to a SVM for text emotion classification. The emotion analysis method for the Chinese texts based on computer information processing technology is capable of improving relevance of emotional semantics of characteristics while the SVM is capable of improving accuracy of emotion classification of comments on Chinese products.

The invention provides a novel prosodic structure boundary division forming method based on prosodic phrases. The method combines machine learning with rules to greatly improve the accuracy of the prediction of Chinese text prosodic structure boundary. Prosodic phrase boundaries are firstly identified on the premise that input files goes through word segmentation and part of speech tagging, then prosodic word boundaries are formed by combining prosodic phrase boundary information, and finally a plurality rules are artificially added to carry out integral modification. In prosodic phrase and prosodic word boundary identification, characteristics are respectively designed and selected for establishing a characteristic template, and a prosodic phrase model and a prosodic word model are established by utilizing the maximum entropy algorithm for respectively identifying prosodic boundaries of two stages. In addition, aiming at the errors in identification of a maximum entropy model, an optimal rule is selected by utilizing an error-driven rule learning method to further improve the accuracy. Based on the method, the prosodic structure boundary division forming method based on prosodic phrases is provided, and the method can effectively improve the accuracy of prosodic structure prediction and the naturalness of speed synthesis.

Disclosed is a speech recognition method which is based on a dynamic extension of the word models in combination with an evaluation of the pronunciation variations.

A method and device for correcting mispronunciations of a user, the method comprising the following steps: providing a database comprising a plurality of records each of which comprising at least a textual and a vocal representation of a specific word; training a speech recognition module to recognize spoken utterances of said user comprising the words represented by said records; generating word models for each recognized spoken word; associating each word model with a respective database record; after training said speech recognition module with sufficient words receiving spoken utterance from said user; extracting a sequence of words from said spoken utterance and generating a word model for each extracted word; comparing said word models to the word models associated with said database records; constructing an audible output comprising vocal representations obtained from records which their word models matched word models generated for said extracted word, wherein said word models comprises features extracted from data of the words spoken by said user.

The invention discloses an intelligent operation and maintenance statement similarity matching method based on a natural language processing technology. The method mainly comprises two parts of data processing in knowledge base construction and sentence similarity matching based on deep learning. Compared with the prior art, the method has the advantages that (1) the operation and maintenance management knowledge is subjected to word segmentation by utilizing the specific word library and the HMM to find the new word model, so that the text word segmentation accuracy is improved, and the moreperfect text word library is established; (2) word vectors are trained through a deep learning method, so that the phenomenon of'dimensionality disaster 'represented by the word vectors can be avoided, information of vocabulary contexts can be fully mined, and relations between words can be obtained; And (3) on the basis of the sentence vectors configured with the weights, not only can the importance measure of each word be obtained, but also the information of the sentence vectors can be richer through the combination of the word vectors, and the accuracy of matching on the basis of forming the sentence vectors can be guaranteed through a cosine similarity matching algorithm.

A new word model is trained from synthetic word samples derived by Monte Carlo techniques from one or more prior word models. The prior word model can be a phonetic word model and the new word model can be a non-phonetic, whole-word, word model. The prior word model can be trained from data that has undergone a first channel normalization and the synthesized word samples from which the new word model is trained can undergo a different channel normalization similar to that to be used in a given speech recognition context. The prior word model can have a first model structure and the new word model can have a second, different, model structure. These differences in model structure can include, for example, differences of model topology; differences of model complexity; and differences in the type of basis function used in a description of such probability distributions.

The invention relates to a remote sensing image land utilization scene classification method based on two-dimension wavelet decomposition and a visual sense bag-of-word model. The method comprises the steps that a remote sensing image land utilization scene classification training set is built; scene images in the training set are converted to grayscale images, and two-dimension decomposition is conducted on the grayscale images; regular-grid sampling and SIFT extracting are conducted on the converted grayscale images and sub-images formed after two-dimension decomposition, and universal visual word lists of the converted grayscale images and the sub-images are independently generated through clustering; visual word mapping is conducted on each image in the training set to obtain bag-of-word characteristics; the bag-of-word characteristics of each image in the training set and corresponding scene category serial numbers serve as training data for generating a classification model through an SVM algorithm; images of each scene are classified according to the classification model. The remote sensing image land utilization scene classification method well solves the problems that remote sensing image texture information is not sufficiently considered through an existing scene classification method based on a visual sense bag-of-word model, and can effectively improve scene classification precision.

The invention discloses a speaker independent isolated words model training method, an identifying system and an identifying method, wherein, the speaker independent isolated words model training method comprises getting a plurality of training modes procedures of the speaker independent isolated words using the dynamic time wrapping algorithm, and a centre training model getting procedure, combining a plurality of training models of the speaker independent isolated words into a centre training model of the speaker independent isolated words. The invention decreases the calculated amount when matching the operation by combining a plurality of training models of the speaker independent isolated words into a centre training mode. At the same time, the invention directly identifies the identifying result smaller than the rejection threshold as out-of-vocabulary words when in identification and process, and effectively improves the rejecting identification capability of out-of-vocabulary words. In addition, by limiting the searching area coverage and loosing the matching jumping-off, the invention increases the recognition rate and decreases the calculated amount.

The invention provides a multi-example keyword detection method based on a multi-task neural network, which belongs to the technical field of voice signal processing. The method comprises the steps of: training a bottleneck deep neural network on a multi-language data set; extracting fbank features from target data set audio frame by frame and extracting bottleneck features of the target data set;based on a training set, training an HMM model for each keyword by using the bottleneck features of the keywords, acquiring a frame-level station tag thereof and training a filler word model by usingthe bottleneck features of all non-keywords; performing multi-task DNN acoustic model training by using the bottleneck features; acquiring the acoustic score of test set audio and obtaining keyword detection results through Viterbi decoding. The multi-task technology can effectively improve low resource conditions and remarkably improve multi-example keyword detection performance.

A speech recognition method and apparatus is disclosed in which outputs a confidence score indicative of the posterior probability of an utterance being correctly matched to a word model. The confidence score for the matching of an utterance to a word model is determined directly from the generated values indicative of the goodness of match between the utterance and stored word models utilizing the following equation:confidence=exp⁡(-2σ⁢S⁡(x|w))∑words⁢exp⁡(2σ⁢S⁡(x|w))where S(x|w) is the match score for the correlation between a signal x and word w and σ is an experimentally determined constant.

The invention provides a voice awakening method and a voice awakening device based on artificial intelligence, wherein the method comprises the following steps: clustering various phonemes so as to select out garbage phonemes for expressing all phonemes; in accordance with an awakening word which is set in advance, constructing a counter-awakening word similar to the awakening word; and constructing a decoding network in accordance with the garbage phonemes, the counter-awakening word and the awakening word, so that voice awakening can be implemented by virtue of the decoding network. Since the garbage phoneme is obviously less than a garbage word in data volume, the problem of the prior art that a garbage word model occupies an excessively high data volume is solved; meanwhile, since each word is constituted by a finite number of phonemes, the garbage phoneme, in comparison with the garbage word, can cover all words more comprehensively, so that awakening precision is enhanced and a probability of incorrect awakening is reduced.

A speech recognition apparatus and speech recognition method are provided for reducing such events as erroneous recognition and disabled recognition and improving a recognition efficiency. The speech recognition apparatus generates a word model based on a dictionary memory and a sub-word sound model, and matches the word model with a speech input signal in accordance with a predetermined algorithm to perform a speech recognition for the speech input signal, wherein the apparatus comprises main matching means, operative when matching the word model with the speech input signal along a processing path indicated by the algorithm, for limiting the processing path based on a course command to select the word model most approximate to the speech input signal, local template storing means for previously typifying local sound features of spoken speeches for storage as local templates; and local matching means for matching each of component sections of the speech input signal with the local templates stored in the local template storing means to definitely determine a sound feature for each of the component sections, and generating the course command in accordance with the result of the definite determination.

Techniques for performing speech recognition in a communication device with a voice dialing function is provided. Upon receipt of a voice input in a speech recognition mode, input feature vectors are generated from the voice input. Also, a likelihood vector sequence is calculated from the input feature vectors indicating the likelihood in time of an utterance of phonetic units. In a warping operation, the likelihood vector sequence is compared to phonetic word models and word model match likelihoods are calculated for that word models. After determination of a best-matching word model, the corresponding number to the name synthesized from the best-matching word model is dialed in a dialing operation.

The invention discloses a multi-channel topic model-based editable garment image search method. The method comprises the steps of firstly, finding a main body region of a garment commodity in a picture by using an object detection method, extracting multiple descriptors from the main body region and quantizing the descriptors into vectors in bag-of-word forms by using a bag-of-word model; secondly, according to search conditions, editing and modifying weights of visual words, fusing the vectors in the bag-of-word forms into retrieval features capable of describing high-level semanteme of the garment commodity by using a pre-trained multi-channel topic model, and establishing indexes; and during online detection, calculating vector similarity of a to-be-queried commodity sample image and animage of a database, and taking the commodity with the highest similarity as a search result. The information of vision, commodity text attributes and the like of the to-be-queried commodity can be re-edited; the demand of a user on the commodity is described more accurately; and the user-expected commodity is searched for through the multi-channel topic model.

This invention discloses one file index method based on language meanings small world, which comprises the following steps: firstly using latent meanings index to extract file property vector to maintain file property to lower its dimensions and to reduce the information memory volume; then using supportive vector machine to sort all common files to form sort information to mark the sort interest proportion; finally using social network small world with small linkage point with high proportion interest of certain file sort to form network topological structure with small property.

A word spotting system includes a semi-continuous hidden Markov model configured to model a handwritten word of interest. A writing segments extractor is configured to extract writing segments generally comprising images of handwritten character strings from a received document image. A word model adaptation processor is configured to adjust a shared pool of Gaussians of the semi-continuous hidden Markov model respective to the extracted writing segments. A modeler is configured to model extracted writing segments using the semi-continuous hidden Markov model with the adjusted shared pool of Gaussians to identify whether each modeled writing segment matches the handwritten word of interest.

A document (or multiple documents) is analyzed to identify entities of interest within that document. This is accomplished by constructing n-gram or bi-gram models that correspond to different kinds of text entities, such as chemistry-related words and generic English words. The models can be constructed from training text selected to reflect a particular kind of text entity. The document is tokenized, and the tokens are run against the models to determine, for each token, which kind of text entity is most likely to be associated with that token. The entities of interest in the document can then be annotated accordingly.

The invention discloses a phonetic segmentation-based isolate word recognition method. In the method, a continuous HMM model obtained by voice training takes syllable or semi-syllable as a unit set, trained unit set models are spliced into whole word models according to the syllables or semi-syllables of isolate words in a word list, and a Viterbi algorithm is adopted in recognition. The method has the advantages of improving recognition performance because each segment of HMM model in recognition results (accurately recognized isolate words) can be better matched with each segment of characteristic parameters of voices to be recognized, and shortening recognition response time because the recognition of the isolate words with relatively smaller probability values can be directly refused every time recognition operation is performed on the tail state of one syllable or semi-syllable.