Abstract: An overall endpointing measure can be generated based on an audio-based endpointing measure and (1) an accelerometer-based endpointing measure and/or (2) a gaze-based endpointing measure. The overall endpointing measure can be used in determining whether a candidate endpoint is an actual endpoint. Various implementations include generating the audio-based endpointing measure by processing an audio data stream, capturing a spoken utterance of a user, using an audio model. Various implementations additionally or alternatively include generating the accelerometer-based endpointing measure by processing a stream of accelerometer data using an accelerometer model. Various implementations additionally or alternatively include processing an image data stream using a gaze model to generate the gaze-based endpointing measure.

Abstract: Systems, methods, and non-transitory computer-readable media including instructions for detecting and utilizing facial skin micromovements are disclosed. In some non-limiting embodiments, the detection of the facial skin micromovements occurs using a speech detection system that may include a wearable housing, a light source (either a coherent light source or a non-coherent light source), a light detector, and at least one processor. One or more processors may be configured to analyze light reflections received from a facial region to determine the facial skin micromovements, and extract meaning from the determined facial skin micromovements.

Abstract: Systems and methods used in a cognitive assistant for detecting human emotions from speech audio signals is described. The system obtains audio signals from an audio receiver and extracts human speech samples. Subsequently, it runs a machine learning based classifier to analyze the human speech signal and classify the emotion observed in it. The user is then notified, based on their preferences, with a summary of the emotion detected. Notifications can also be sent to other systems that have been configured to receive them. Optionally, the system may include the ability to store the speech sample and emotion classification detected for future analysis. The system's machine learning classifier is periodically re-trained based on labelled audio speech data and updated.

Abstract: Systems, methods, and computer program products are disclosed for initiating content interpretation operations prior to vocalization of content to be interpreted. Initiating content interpretation operations prior to vocalization of content to be interpreted includes receiving signals representing facial skin micromovements; determining from the signals at least one word to be spoken prior to vocalization of the at least one word in an origin language; prior to the vocalization of the at least one word, instituting an interpretation of the at least one word; and causing the interpretation of the at least one word to be presented as the at least one word is spoken.

Abstract: Systems are provided for multilingual speech data processing. A language identification module is configured to analyze spoken utterances in an audio stream and to detect at least one language corresponding to the spoken language utterances. The language identification module detects that a first language corresponds to the first portion of the audio stream. A first transcription of the first portion of the audio stream in the first language is generated and stored in a cache. A second transcription of a second portion of the audio stream in the first language is also generated and stored. When the second portion of the audio stream corresponds to a second language, a third transcription is generated in the second language using a second speech recognition engine configured to transcribe spoken language utterances in the second language. Then, the second transcription is replaced with the third transcription in the cache and any displayed instances.

Abstract: A plurality of audio datasets associated with captured audio are provided to a plurality of automatic speech recognition engines, wherein each of the automatic speech recognition engines is configured to recognize speech of a first language. Word error rate estimates that comprise at least one word error rate estimate for each of the plurality of audio datasets are determined from outputs of the plurality of automatic speech recognition engines. From the word error rate estimates, audio in the plurality of audio datasets is determined to include speech in a second language.

Abstract: A method includes receiving a sequence of acoustic frames extracted from audio data corresponding to an utterance. During a first pass, the method includes processing the sequence of acoustic frames to generate N candidate hypotheses for the utterance. During a second pass, and for each candidate hypothesis, the method includes: generating a respective un-normalized likelihood score; generating a respective external language model score; generating a standalone score that models prior statistics of the corresponding candidate hypothesis; and generating a respective overall score for the candidate hypothesis based on the un-normalized likelihood score, the external language model score, and the standalone score. The method also includes selecting the candidate hypothesis having the highest respective overall score from among the N candidate hypotheses as a final transcription of the utterance.

Abstract: A method including receiving at least one of touch data or force data representing a touch input received at the controller, determining one or more model(s), generating image data using the one or more models, the image data representing at least a hand gesture corresponding to the touch input received at the controller, and transmitting the image data to a virtual reality (VR) environment for display.

Abstract: A system has multiple audio-enabled devices that communicate with one another over an open microphone mode of communication. When a user says a trigger word, the nearest device validates the trigger word and opens a communication channel with another device. As the user talks, the device receives the speech and generates an audio signal representation that includes the user speech and may additionally include other background or interfering sound from the environment. The device transmits the audio signal to the other device as part of a conversation, while continually analyzing the audio signal to detect when the user stops talking. This analysis may include watching for a lack of speech in the audio signal for a period of time, or an abrupt change in context of the speech (indicating the speech is from another source), or canceling noise or other interfering sound to isolate whether the user is still speaking.

Abstract: A document creation assistance apparatus includes a tree structure generation unit configured to analyze a learning document for system development and generate a tree structure representing separate sections of the learning document, a frequency calculation unit configured to calculate, per leaf node of the tree structure, a frequency vector of a word that appears, a question extraction unit configured to extract, according to the frequency vector, a word about which a user is to be questioned, a question presentation unit configured to present a question about the extracted word to the user and receive an answer, and a document generation unit configured to generate a document with the extracted word and the answer set in a section of the separate sections of the leaf node according to the separate sections of the tree structure.

Abstract: Methods, computer systems and computer program product are provided for retrieving contextually relevant documents in near real time. When text data it's received from an application, the text data is processed through a text segmentation model to generate a set of documents. Each document corresponds to a segment of the text data. A first vector representation is generated for a first document of the set of documents. A machine learning process compares the first vector representation and a set of vector representations for a set of documents within a data repository to determine a subset of the documents. A composite rank is generated for each respective document of the subset. The subset of documents is then presented through an interface, sorted according to the respective composite ranks.

Abstract: Various embodiments described herein relate to determining and providing user-specific feedback based on an analysis of audible input sessions performed by a user. In this regard, a set of term recognition structures that each comprise a plurality of term data objects and a respective confidence score for each term data object are generated. For at least one pairing of term data objects of a predefined term glossary, a correlation coefficient value for the respective pairing is determined. In accordance with determining that the correlation coefficient value for the at least one pairing satisfies a predefined threshold a generate a visualization is generated and displayed that includes an indication of the term data objects of the at least one pairing.

Abstract: A cognitive function evaluation device includes: an instruction unit that instructs quick pronunciation of pseudoword in which a predetermined syllable is repeated; an obtainment unit that obtains voice data indicating a voice of an evaluatee who has received an instruction; a calculation unit that calculates a feature from the voice data obtained by the obtainment unit; an evaluation unit that evaluates a cognitive function of the evaluatee from the feature calculated by the calculation unit; and an output unit that outputs a result of the evaluation by the evaluation unit.

Abstract: The disclosed technologies include methods for generating a calibration model using data that is selected to match the conditions of a particular trial that involves an automated comparison of data samples, such as a comparison-based trial performed by an audio-based recognition, identification, or detection system. The disclosed technologies also include improved methods for selecting candidate data used to build the calibration model. The disclosed technologies further include methods for evaluating the performance of the calibration model and for rejecting a trial when not enough matched candidate data is available to build the calibration model. The disclosed technologies additionally include the use of regularization and automated data generation techniques to further improve the robustness of the calibration model.

Abstract: Certain aspects of the present disclosure provide techniques for vectorization of transactions including: receiving electronic transaction information of one or more transactions of a user; for each transaction of the one or more transactions: segmenting the electronic transaction information of the transaction into one or more transaction words; generating a second transaction description related to the transaction; and identifying a category of the transaction; generating, based on the corresponding identified categories of the one or more transactions, a set of transaction history data of the user; providing the set of transaction history data of the user as an input to a machine learned model trained to output a set of word embedding vectors; determining, based on an output of the machine learned model comprising a set of word embedding vectors, a set of similar merchants; and providing the set of similar merchants for display to the user.

Abstract: A system configured to perform utterance detection using data processing techniques that are similar to those used for object detection is provided. For example, the system may treat utterances within audio data as analogous to an object represented within an image and employ techniques to separate and identify individual utterances. The system may include one or more trained models that are trained to perform utterance detection. For example, the system may include a first module to process input audio data and identify whether speech is represented in the input audio data, a second module to apply convolution filters, and a third module configured to determine a boundary identifying a beginning and ending of a portion of the input audio data along with an utterance score indicating how closely the portion of the input audio data represents an utterance.

Abstract: Methods and apparatuses for feature-driven communications are described. A set of features describing an observed subject is transmitted by a transmitting electronic device (ED) to a base station (BS). The BS translates the received features to another set of transmission features to be transmitted to a receiving ED. The receiving ED recovers information about the subject from the features received from the BS.

Abstract: A method comprises receiving a natural language search query, identifying a first set of semantic concepts in the query, creating a vector representation of the first set of semantic concepts, identifying a second set of semantic concepts having a vector representation within a predetermined threshold of similarity to the first set of semantic concepts, performing a search of documents based on the first set of semantic concepts, presenting a result set of documents and the first, second, and third sets of semantic concepts to a user, receiving input from the user, performing a second search of the documents based on the input from the user to obtain a second result set of documents, identifying a fourth set of semantic concepts based on the second result set of documents, and presenting the second result set of documents and the fourth set of semantic concepts to the user.

Abstract: An approach is provided for improving a named entity recognition. Using a multi-label classification in a neural network, a sub-entity is identified in an original sentence. First and second labels are determined indicating first and second candidate types of the sub-entity. First and second replacement sentences are generated. The first replacement sentence replaces the sub-entity in the original sentence with a first sub-entity of the first candidate type. The second replacement sentence replaces the sub-entity in the original sentence with a second sub-entity of the second candidate type. First and second confidence scores for the first and second replacement sentences are determined. Based on the first confidence score exceeding the second confidence score by more than a threshold amount, the neural network is retrained by selecting the first instead of the second candidate type as the sub-entity type.

Abstract: It discloses a speech feature reuse-based storing and calculating compression method for a keyword-spotting CNN, and belongs to the technical filed of calculating, reckoning or counting. If the updated row number of input data is equal to a convolution step size, every time new input data arrive, an input layer of a neural network replaces the earliest part of the input data with the new input data and meanwhile adjusts an addressing sequence of the input data, thereby performing an operation on the input data and corresponding convolution kernels in an arrival sequence of the input data, and an operation result is stored in an intermediate data memory of the neural network to update corresponding data.

Abstract: Techniques performed by a data processing system for analyzing training data for a machine learning model and identifying outliers in the training data herein include obtaining training data for the model from a memory of the data processing system; analyzing the training data using a Siamese Neural Network to determine within-label similarities and cross-label similarities associated with a plurality of data elements within the training data, the within-label representing similarities between a respective data element and a first set of data elements similarly labeled in the training data, the cross-label similarities representing similarities between the respective data element and a second set of data elements dissimilarly labeled in the training data; identifying outlier data elements in the plurality of data elements based on the within-label and cross-label similarities; and processing the training data comprising the outlier data elements.

Abstract: In some implementations, a determination that a first party has spoken a query for a voice-enabled virtual assistant during a voice call between the first party and a second party is made, in response to the determination that the first party has spoken the query for the voice-enabled virtual assistant during the voice call between the first party and the second party, the voice call between the first party and the second party is placed on hold, a determination that the voice-enabled virtual assistant has resolved the query is made, and in response to the determination that the voice-enabled virtual assistant has handled the query, the voice call between the first party and the second party is resumed from hold.

Abstract: An exemplary method for generating a test vector to activate a Trojan triggering condition includes the operations of obtaining a design graph representation of an electronic circuit; constructing a satisfiability graph from the design graph representation, wherein the satisfiability graph includes a set of vertices representing rare signals of the electronic circuit and satisfiability connections between the vertices; finding a plurality of maximal satisfiable cliques in the satisfiability graph, wherein a maximal satisfiable clique corresponds to a triggering condition for a payload of the electronic circuit; generating a test vector for each of the maximal satisfiable cliques; and performing a test for the presence of a hardware Trojan circuit in the electronic circuit using the generated test vectors as input signals.

Abstract: At least one processor may obtain a document comprising text tokens. The at least one processor may determine, based on a pre-trained language model, word embeddings corresponding to the text tokens. The at least one processor may determine, based on the word embeddings, named entities corresponding to the text tokens; and one or more accuracy predictions corresponding to the named entities. The at least one processor may compare the one or more accuracy predictions with at least one threshold. The at least one processor may associate, based on the comparing, the named entities with one or more confidence levels. The at last one processor may deliver the named entities and the one or more confidence levels.

Abstract: In some implementations, a determination that a first party has spoken a query for a voice-enabled virtual assistant during a voice call between the first party and a second party is made, in response to the determination that the first party has spoken the query for the voice-enabled virtual assistant during the voice call between the first party and the second party, the voice call between the first party and the second party is placed on hold, a determination that the voice-enabled virtual assistant has resolved the query is made, and in response to the determination that the voice-enabled virtual assistant has handled the query, the voice call between the first party and the second party is resumed from hold.

Abstract: Systems and methods to group terms based on context to facilitate determining intent of a command are disclosed. Exemplary implementations to train a model: obtain a set of writings within a particular knowledge domain; obtain a vector generation model that generates vectors for individual instances of the terms in the set of writings; generate a first set of vectors that represent the instances of a first term and other vectors that represent instances of the other terms of the set of writings; train the vector generation model to group the vectors of a similar context in a space of a vector space; obtain a transcript include a new term generated from user audio dictation; generate a new vector that represent the instance of the new term; obtain the space; compare the new vector with the space; utilize the new term as the first term.

Abstract: A system and method performs wakeword detection using a feedforward neural network model. A first output of the model indicates when the wakeword appears on a right side of a first window of input audio data. A second output of the model indicates when the wakeword appears in the center of a second window of input audio data. A third output of the model indicates when the wakeword appears on a left side of a third window of input audio data. Using these outputs, the system and method determine a beginpoint and endpoint of the wakeword.

Abstract: Techniques are described herein for training and evaluating machine learning (ML) models for document processing computing applications based on in-domain and out-of-domain characteristics. In some embodiments, an ML system is configured to form feature vectors by mapping unknown tokens to known tokens within a domain based, at least in part, on out-of-domain characteristics. In other embodiments, the ML system is configured to map the unknown tokens to an aggregate vector representation based on the out-of-domain characteristics. The ML system may use the feature vectors to train ML models and/or estimate unknown labels for the new documents.

Abstract: Techniques are described herein for training and evaluating machine learning (ML) models for document processing computing applications based on in-domain and out-of-domain characteristics. In some embodiments, an ML system is configured to form feature vectors by mapping unknown tokens to known tokens within a domain based, at least in part, on out-of-domain characteristics. In other embodiments, the ML system is configured to map the unknown tokens to an aggregate vector representation based on the out-of-domain characteristics. The ML system may use the feature vectors to train ML models and/or estimate unknown labels for the new documents.

Abstract: Efficient and effective communications with customers is a cornerstone of many businesses. Automation, such as in the form of automated agents that can engage in a communication with a customer, furthers those efficient and effective communications. However, textual messages are a series of messages that are often limited to factual statements and direct questions, leaving many customers, such as those that prefer or require high-context communications, may have the impression that the organization with which they are communicating is cold and uncaring or unconcerned about them. By selectively altering the series of messages, an appropriate degree of concern or empathy may be conveyed to facilitate a better relationship and more effective and efficient communications.

Abstract: An utterance support apparatus includes: a processor configured to execute a program; and a storage device configured to store the program, wherein the processor is configured to execute: calculation processing of calculating an accumulated value of utterance periods of each of a plurality of speakers, and clearing the accumulated value of a speaker having the accumulated value that has reached a predetermined value; and display processing of displaying a first graphic in a display region, which is included in a group of display regions each assigned to each of the plurality of speakers, and which is assigned to the speaker having the accumulated value that has reached the predetermined value.

Abstract: Disclosed are a speech recognition device and a speech recognition method which perform speech recognition by executing an artificial intelligence (AI) algorithms and/or a machine learning algorithm installed thereon, to communicate with other electronic devices and an external server in a 5G communication environment. The speech recognition method according to an embodiment of the present disclosure may include converting a series of spoken utterance signals to a text item, extracting a discordant named-entity that is discordant with a parent domain inferred form the text, calculating probabilities of candidate words associated with the discordant named-entity based on calculated distances between a term representing the parent domain and each candidate word associated with the discordant named-entity, and based on the calculated probabilities, modifying the discordant named-entity in the text to one of the candidate words associated with the discordant named-entity.

Abstract: In one embodiment, an embedding is determined for each entity in a set of entities that is selected from a plurality of entities. Each embedding corresponds to a point in an embedding space, which includes points corresponding to embeddings of entities. The embeddings of the entities are determined using a deep-learning model. Embeddings are determined for each entity attribute in a set of entity attributes. Each of the entity attributes in the set is of an entity-attribute type and is associated with at least one entity. The entity-attribute embeddings are refined using the deep-learning model. The embeddings of the entities in the set of entities are modified based on the entity-attribute embeddings that are associated with the respective entity to obtain updated embeddings for each entity in the set. The updated embeddings include information regarding the entity attributes that are associated with the respective entities.

Abstract: Apparatuses, systems, methods, and computer program products are disclosed for medical assessment based on voice. A query module is configured to audibly question a user from a speaker of a mobile computing device with one or more open ended questions. A response module is configured to receive a conversational verbal response of a user from a microphone of a mobile computing device in response to one or more open ended questions. A detection module is configured to provide a machine learning assessment for a user of a medical condition based on a machine learning analysis of a received conversational verbal response of the user.

Abstract: A system and method provides for multibeam keyword detection. A composite audio signal may include sound components. The system and method groups the sound components into subsets based on the angles of arrival of sound components. Keyword detectors evaluate each subset and determine whether a keyword is present.

Abstract: A natural language processing apparatus includes: a first calculation unit configured to calculate a distributed vector of a word included in a plurality of sentences based on a database that manages the plurality of sentences associated with a classification word; a second calculation unit configured to calculate a distributed vector of the sentence based on the distributed vector of the word included in each sentence; and a third calculation unit configured to calculate a distributed vector of the classification word based on the distributed vector of each sentence associated with the same classification word.

Abstract: A table organized into a set of batch units is accessed. A set of N-grams are generated for a data value in the source table. The set of N-grams include a first N-gram of a first length and a second N-gram of a second length where the first N-gram corresponds to a prefix of the second N-gram. A set of fingerprints are generated for the data value based on the set of N-grams. The set of fingerprints include a first fingerprint generated based on the first N-gram and a second fingerprint generated based on the second N-gram and the first fingerprint. A pruning index that indexes distinct values in each column of the source table is generated based on the set of fingerprints and stored in a database with an association with the source table.

Abstract: A method for operating an electronic device to configure a subject device, the method comprising steps of: receiving an intent from a subject device, wherein the received intent comprises an action identifier identifying an action the subject device wishes to perform; receiving action data about the received intent from an intent store, wherein the action data comprises an action associated with each action identifier, and at least one constraint associated with the action; and generating invocation data to perform the action, wherein the invocation data comprises the action identifier, and zero or more parameters.

Abstract: A system for incremental natural language understanding includes a media module, a memory storing a software code, and a hardware processor communicatively coupled to the media module. The hardware processor is configured to execute the software code to receive an audio stream including a first utterance, and generate a first and second incremental speech recognition outputs based on first and second portions of the first utterance. In addition, the hardware processor is configured to execute the software code to determine, prior to generating the second incremental speech recognition output, a first intent of the first utterance based on the first incremental speech recognition output. The hardware processor is further configured to execute the software code to retrieve a first resource based on the determined first intent, and incorporate the first resource in the media content to be played by the media module.

Abstract: Techniques herein improve computational efficiency for wildcard searches by using numeric string hashes. In an embodiment, a plurality of query K-gram tokens for a term in a query are generated. Using a first index, an intersection of hash tokens is determined, wherein said first index indexes each query K-gram token of said K-gram tokens to a respective subset of hash tokens of a plurality of hash tokens, each of hash token of said plurality of hash tokens corresponding to a term found in one or more documents of a corpus of documents. The intersection of hash tokens comprises only hash tokens indexed to all of said plurality of query K-gram tokens by said first index. Using a second index, documents of said corpus of documents that contain said term are determined, said second index indexing said hash tokens to a plurality of terms in said corpus of documents and for each term of said plurality of terms, a respective subset of documents of corpus of documents that contain said each term.

Abstract: A method, system and computer program product for improving collaboration among participants in a collaboration system. In one embodiment of the present invention, a system, referred to herein as the “integration system,” connected to a collaboration system monitors for comments or updates pertaining to a work item involving a customer problem to be resolved by different participants of the collaboration system. These comments or updates for completing the work item are analyzed. After analyzing the comments or updates, strategies are derived for completing the work item based on the analysis of the comments or updates as well as based on stored data of previously resolved work items. The derived strategies are then presented to the appropriate participant(s) to resolve the work item, such as based on the roles of the participant(s) that would most effectively and efficiently perform the strategy.

Abstract: A computer-implemented method includes: comparing features extracted from a first document that include a sound to features extracted from acoustic files related to the sound; designating the sound in a document of the plurality of documents as a true; designating the sound in the first document as a false negative; generating a first sound vector for the sound in the first document in response to the sound in the first document being designated a false negative; generating a sound vector for each of the documents designated as a true positive; creating a centroid vector for the sound vectors of the documents designated as a true positive; and redesignating the sound in the first document from a false negative to a true positive in response to the first sound vector and the centroid vector being a Banach space.

Abstract: According to an embodiment, a voice recognition device includes one or more processors. The one or more processors are configured to: recognize a voice signal representing a voice uttered by an object speaker, to generate text and meta information representing information that is not included in the text and included in the voice signal; generate an object presentation vector including a plurality of parameters representing a feature of a presentation uttered by the object speaker; calculate a similarity between the object presentation vector and a reference presentation vector including a plurality of parameters representing a feature of a presentation uttered by a reference speaker; and output the text. The one or more processors are further configured to determine whether to output the meta information based on the similarity, and upon determining to output the meta information, add the meta information to the text and output the meta information.

Abstract: A method is described comprising receiving a conversational transcript of a conversational interaction among a plurality of participants, wherein each participant contributes a sequence of contributions to the conversational interaction. The method includes projecting contributions of the plurality of participants into a semantic space using a natural language vectorization, wherein the semantic space describes semantic relationships among words of the conversational interaction. The method includes computing interaction process measures using information of the conversational transcript, the conversational interaction, and the natural language vectorization.

Abstract: Example methods, apparatus, systems and articles of manufacture are disclosed to determine tags for unknown media using multiple media features. Disclosed examples extract features from audio and image portions of the unknown media. Disclosed examples weight the features based at least partially on respective recognition technologies used to extract the features to determine corresponding weighted features, wherein disclosed examples assign a first weight to a first feature extracted by an image-based recognition technology, and assign a second weight, different from the first weight, to a second feature extracted by an audio-based recognition technology. Disclosed examples search a database of pre-tagged media with a combination of the weighted features to generate a list of suggested tags for the unknown media. Disclosed examples assign one or more tags from the list of suggested tags to the unknown media.

Abstract: Embodiments of the present disclosure relate to a method and apparatus for outputting information. The method includes: outputting a to-be-read audio in response to receiving a reading instruction from a user; acquiring an actually read audio obtained by reading the to-be-read audio by the user; performing speech recognition on the actually read audio to obtain a recognition result; calculating a similarity between the actually read audio and the to-be-read audio based on a character string corresponding to the recognition result and a character string corresponding to the to-be-read audio; determining, from a predetermined set of similarity intervals, a similarity interval to which the calculated similarity belongs; and outputting a reading evaluation corresponding to the determined similarity interval. The embodiment may help a reader to improve the learning efficiency and learning interest, thereby improving the rate of a user using a device.

Abstract: There is provided an information processing apparatus including: a processing unit configured to perform a summarization process of summarizing content of speech indicated by voice information based on speech of a user on a basis of acquired information indicating a weight related to a summary.

Abstract: A query directed at a source table organized into a set of batch units is received. The query includes a pattern matching predicate that specifies a search pattern. A set of N-grams are generated based on the search pattern. A pruning index associated with the source table is accessed. The pruning index comprises a set of filters that index distinct N-grams in each column of the source table. The pruning index is used to identify a subset of batch units to scan for matching data based on the set of N-grams generated for the search pattern. The query is processed by scanning the subset of batch units.

Abstract: A computer-implemented method and supporting system transcribes spoken words being monitored from a telephonic interaction among two or more individuals. Telephonic interactions among the individuals are monitored, and at least two of the individuals are each assigned to a separate channel. While still being monitored, each of the channels is assigned a context-based speech recognition models, and in substantially real-time, the monitored telephonic interaction is transcribed from speech to text based on the different assigned models.

Abstract: A speech recognition method and apparatus are disclosed. The speech recognition method includes determining a first score of candidate texts based on an input speech, determining a weight for an output of a language model based on the input speech, applying the weight to a second score of the candidate texts output from the language model to obtain a weighted second score, selecting a target candidate text from among the candidate texts based on the first score and the weighted second score corresponding to the target candidate text, and determining the target candidate text to correspond to a portion of the input speech.