58 results about "Cache language model" patented technology

A speech understanding system includes a language model comprising a combination of an N-gram language model and a context-free grammar language model. The language model stores information related to words and semantic information to be recognized. A module is adapted to receive input from a user and capture the input for processing. The module is further adapted to receive SALT application program interfaces pertaining to recognition of the input. The module is configured to process the SALT application program interfaces and the input to ascertain semantic information pertaining to a first portion of the input and output a semantic object comprising text and semantic information for the first portion by accessing the language model, wherein performing recognition and outputting the semantic object are performed while capturing continues for subsequent portions of the input.

Certain disclosed methods and systems perform multiple different types of message recognition using a shared language model. Message recognition of a first type is performed responsive to a first type of message input (e.g., speech), to provide text data in accordance with both the shared language model and a first model specific to the first type of message recognition (e.g., an acoustic model). Message recognition of a second type is performed responsive to a second type of message input (e.g., handwriting), to provide text data in accordance with both the shared language model and a second model specific to the second type of message recognition (e.g., a model that determines basic units of handwriting conveyed by freehand input). Accuracy of both such message recognizers can be improved by user correction of misrecognition by either one of them. Numerous other methods and systems are also disclosed.

Assigning meanings to spoken utterances in a speech recognition system. A plurality of speech rules is generated, each of the of speech rules comprising a language model and an expression associated with the language model. At one interval (e.g. upon the detection of speech in the system), a current language model is generated from each language model in the speech rules for use by a recognizer. When a sequence of words is received from the recognizer, a set of speech rules which match the sequence of words received from the recognizer is determined. Each expression associated with the language model in each of the set of speech rules is evaluated, and actions are performed in the system according to the expressions associated with each language model in the set of speech rules.

To provide a speech recognition apparatus which appropriately performs speech recognition by generating, in real time, language models adapted to a new topic even in the case where topics are changed. The speech recognition apparatus includes: a word specification unit for obtaining and specifying a word; a language model information storage unit for storing language models for recognizing speech and the respectively corresponding pieces of tag information; a combination coefficient calculation unit for calculating the weights of the respective language models, as combination coefficients, according to the word obtained by the word specification unit, based on the relevance degree between the word obtained by the word specification unit and the tag information of each language model; a language probability calculation unit for calculating the probabilities of word appearance by combining the respective language models according to the calculated combination coefficients; and a speech recognition unit for recognizing speech using the calculated probabilities of word appearance.

Methods and systems for automatic speech recognition and methods and systems for training acoustic language models are disclosed. One system for automatic speech recognition includes a dialect recognition unit and a controller. The dialect recognition unit is configured to analyze acoustic input data to identify portions of the acoustic input data that conform to a general language and to identify portions of the acoustic input data that conform to at least one dialect of the general language. In addition, the controller is configured to apply a general language model and at least one dialect language model to the input data to perform speech recognition by dynamically selecting between the models in accordance with each of the identified portions.

A system and method is provided that randomly generates text with a given structure. The structure is taken from a number of learning examples. The structure of training examples is captured by word classification and the definition of the relationships between word classes in a given language. The text generated with this procedure is intended to replicate the information given by the original learning examples. The resulting text may be used to better model the structure of a language in a stochastic language model.

A language model structure for use in a speech recognition system employs a tree-structured network model. The language model is structured such that identifiers associated with each word and contained therein are arranged such that each node of the network model with which the language model is associated spans a continuous range of identifiers. A method of transferring tokens through a tree-structured network in a speech recognition process is also provided.

The invention a voice recognition method based on statistic language model, combining grammar rules, and oriented to language speaking recognition application. The invention comprises an acoustic model training, a language model training, a front-end processing, and a recognizing decoding. Based on the N element grammar statistic model, and grammar rule network, the language model of the invention is capable of effectively processing phenomenon such as out-of-vocabulary words, slogan, blurred pronunciation, fast switching between sentences, so that higher recognition rate is achieved with the proviso that naturalness of voice recognition is ensured.

A speech-based processing system includes a database of PIM data of a user, a set of language models, a learning unit, a recognition server, and a speech application. The learning unit uses a language model learning algorithm to provide language models based on the PIM data. The recognition server recognizes an utterance of the user by using one of the language models. The speech application identifies and accesses a subset of the PIM data specified by the utterance by using the recognition result. The language model learning algorithm may use grammar induction and/or or may train statistical language models based on the PIM data. The language model learning algorithm may be applied to generate language models periodically or on-the-fly during a session with the user.

Determination of a word input on a reduced keypad, such as a numeric keypad, by entering a key sequence ambiguously corresponding to the word, by taking into account the context of the word via a machine learning approach, is disclosed. Either the left context, the right context, or the double-sided context of the number sequence can be used to determine the intended word. The machine learning approach can use a statistical language model, such as an n-gram language model. The compression of a language model for use with small devices, such as mobile phones and other types of small devices, is also disclosed.

The invention provides a voice recognition method based on a layered circulation neural network language model. The method mainly comprises steps of character-level language modeling using RNN, expansion of an RNN structure by use of an external clock and a reset signal, character-level language modeling with graded RNN and voice recognition. According to the invention, the traditional single-clock RNN character-level language model is replaced by layered circulation neural network-based language model, so quite high recognition precision is achieved; quantity of parameters is reduced; vocabulary of the language model is huge; required storage space is quite small; and a layered language model can be expanded to process information of quite long period, such as sentences, topics or other contexts.

The invention discloses a joint modeling method for a spoken language understanding model and a language model, which comprises the steps of sampling a text sequence from a sample library, and transforming the text sequence into a corresponding training vector sequence; inputting the training vector sequence into a bidirectional long-short term memory network; performing joint training of the spoken language understanding model and the language model by adopting network output of the bidirectional long-short term memory network; and extracting text feature information from the training vectorsequence by adopting the bidirectional long-short term memory network for joint training of the spoken language understanding model and the language model, thereby realizing the sharing of semantic and grammatical feature information between the spoken language understanding model and the language model.

A speech recognition apparatus can include a transformation processor configured to transform at least one phoneme sequence included in speech into at least one word sequence, and to provide the word sequence with an appearance probability indicating that the phoneme sequence originally represented the word sequence. A renewal processor can renew the appearance probability based on a renewed numerical value indicated by language models corresponding to the word sequence provided by the transformation processor. A recognition processor can select one of the word sequences for which the renewed appearance probability is the highest to indicate that the phoneme sequence originally represented the selected word sequence. The renewal processor can calculate the renewed numerical value using a first language model prepared for expressions unique to spontaneous speech, and a second language model different from the first which employs the renewed numerical value to renew the appearance probability.

The invention relates to an unsupervised text similarity calculation method, which comprises the following steps of: 1, pre-training an embedded layer model, and pre-training all words in a problem set to generate word vectors meeting the requirements of the model; 2, mining semantic information of sentences through a coding layer network; step 3, performing model improvement based on TFIDF fusion; the method comprises the steps that when each question sentence is input into a neural network, TFIDF calculation is conducted on each input question sentence, calculated weights are input into theneural network, final sentence vector representation is controlled, a normalized TFIDF calculation method is adopted, and the final sentence vector representation is fused into a coding layer and a representation layer. According to the method, the deep neural network model (Bi-LSTM) is used for unsupervised training of the corpus to obtain the language model, and the information of the large-scale corpus can be fully utilized in an unsupervised training mode, so that the text matching accuracy is improved, and the information retrieval precision is improved.

The invention disclosed herein concerns a method of converting speech to text using a hierarchy of contextual models. The hierarchy of contextual models can be statistically smoothed into a language model. The method can include processing text with a plurality of contextual models. Each one of the plurality of contextual models can correspond to a node in a hierarchy of the plurality of contextual models. Also included can be identifying at least one of the contextual models relating to the text and processing subsequent user spoken utterances with the identified at least one contextual model.

The present disclosure provides a language model training method and device, including: obtaining a universal language model in an offline training mode, and clipping the universal language model to obtain a clipped language model; obtaining a log language model of logs within a preset time period in an online training mode; fusing the clipped language model with the log language model to obtain a first fusion language model used for carrying out first time decoding; and fusing the universal language model with the log language model to obtain a second fusion language model used for carrying out second time decoding. The method is used for solving the problem that a language model obtained offline in the prior art has poor coverage on new corpora, resulting in a reduced language recognition rate.