Abstract: A tool for managing text-item recognition systems such as NER (Named Entity Recognition) systems. The tool applies the system to a text corpus containing instances of text items, such as named entities, to be recognized by the system, and selecting from the text corpus a set of instances of text items which the system recognized. The tool tokenizes the text corpus such that each instance in the aforementioned set is encoded as a single token and processing the tokenized text via a word embedding scheme to generate a word embedding matrix. The tool, responsive to selecting a seed token corresponding to an instance in the aforementioned set, performs a nearest-neighbor search of the embedding space to identify a set of neighboring tokens for the seed token, and identifies the text corresponding to each neighboring token as a potential instance of a text item to be annotated.

Abstract: Computer-based natural language understanding of input and output for a computer interlocutor is improved using a method of classifying conversation segments from transcribed conversations. The improvement includes one or more methods of splitting transcribed conversations into groups related to a conversation ontology using metadata; identifying dominant paths of conversational behavior by counting the frequency of occurrences of the behavior for a given path; creating a conversation model comprising conversation behaviors, metadata, and dominant paths; and using the conversation model to assign a probability score for a matched input to the computer interlocutor or a generated output from the computer interlocutor.

Abstract: One or more processing devices derive values indicative of various aspects of how a particular service in an information technology (IT) environment is performing at a point in time or for a period of time. The values are derived by a search query over machine data associated with the one or more entities that provide the service. The one or more processing devices define and apply time varying static thresholds in respect to the values. A user (e.g., IT manager) may be enabled to manipulate or define multiple sets of KPI thresholds that vary over time.

Abstract: Method starts with processing, by a processor, audio signal to generate audio caller utterance and transcribed caller utterance. Processor generates identified task based on transcribed caller utterance. Processor samples audio caller utterance to generate samples of audio caller utterance. Processor generates loudness result based on loudness values of samples using loudness neural network associated with identified task. Processor generates pitch result based on pitch values of samples using pitch neural network associated with identified task. Processor generates tone result for each word in transcribed caller utterance using tone neural network associated with identified task. Using task completion probability neural network associated with identified task, processor generates task completion probability result that is based on at least one of: loudness result, pitch result, or tone result. Other embodiments are disclosed herein.

Abstract: The present disclosure is generally related to a data processing system to selectively invoke applications for execution. A data processing system can receive an input audio signal and can parse the input audio signal to identify a command. The data processing system can identify a first functionality of a first digital assistant application hosted on the data processing system in the vehicle and a second functionality of a second digital assistant application accessible via a client device. The data processing system can determine that one of the first functionality or the second functionality supports the command. The data processing system can select one of the first digital assistant application or the second digital assistant application based on the determination. The data processing system invoke one of the first digital assistant application or the second digital assistant application based on the selection.

Abstract: The present disclosure relates to a method and device for generating a finite state automata for recognizing a chemical name in a text, and a recognition method. According to an embodiment of the present disclosure, the method comprises substituting representation constants of categories of character segments appearing in an organic compound name set into the organic compound name set to obtain a conversion name set; updating the conversion name set based on a conversion name segment which repeatedly appears in the conversion name set; and generating the finite state automata based on the updated conversion name set.

Abstract: An electronic apparatus configured to acquire information on a plurality of candidate texts corresponding to input speech of a user through a general speech recognition module, determine text corresponding to the input speech from among the plurality of candidate texts using a trained personal language model, and output the text as a result of speech recognition of the input speech.

Abstract: A speech recognition method includes receiving speech data, obtaining candidate texts corresponding to the speech data and respective scores of the candidate texts using a speech recognition model, adjusting the score of a current candidate text, from among the obtained candidate texts, in response to a text length of the current candidate text satisfying a condition determined based on text lengths of the obtained candidate texts, and determining a target text corresponding to the speech data, from among the obtained candidate texts and the current candidate texts.

Abstract: It is disclosed a voice pickup method and apparatus for an intelligent rearview mirror, an electronic device and a computer readable storage medium which relates to the technical field of vehicle-mounted equipment, and may be used in the field of automatic driving technologies. A voice pickup implementation of the intelligent rearview mirror according to some embodiments includes: acquiring an image of the interior of the vehicle; determining the position of a person in the vehicle with the image of the interior of the vehicle; and adjusting a beamforming direction of a microphone array according to the position of the person in the vehicle.

Abstract: A electronic apparatus includes a display, a voice receiver configured to receive a user voice input, and a processor to obtain a first text from the user voice input that is received through the voice receiver based on a function corresponding to a first voice recognition related to a first language, based on an entity name not being included in the first text using the function corresponding to the first voice recognition related to the first language, obtain a second text corresponding to the entity name from of the user voice input based on a function corresponding to a second voice recognition related to a second language, and control the display to display a voice recognition result corresponding to the user voice input based on the first text and the second text.

Abstract: Systems and methods for detecting and analyzing audio in a media presentation environment to determine whether to replay missed portions of media content are disclosed herein. In an embodiment, one or more computing devices detect audio in a media presentation environment. The one or more computing devices determine whether the audio relates to the media being presented. If the audio does not relate to the media being presented, the one or more computing devices cause replaying a portion of the media presentation corresponding to when the audio was being detected.

Abstract: Systems and methods of presenting interrupting content during human speech are disclosed. The proposed systems offer improved duplex communications in conversational AI platforms. In some embodiments, the system receives speech data and evaluates the data using linguistic models. If the linguistic models detect indications of linguistic irregularities such as mispronunciation, a smart feedback assistant can determine that the system should interrupt the speaker in near-real-time and provide feedback regarding their pronunciation. In addition, conversational irregularities may also be detected, causing the smart feedback assistant to interrupt with presentation of moderating guidance. In some cases, emotion models may also be utilized to detect emotional states based on the speaker's voice in order to offer near-immediate feedback. Users can also customize the manner and occasions in which they are interrupted.

Abstract: A semantic augmentation system includes a sensor with a computing system and a memory in communication with the computing system, the memory storing a plurality of endpoints. The computing system is configured to infer a first and a second semantic identity for an object, based on inputs from the sensor, project a coherent narrative and perform semantic augmentation towards a user. In further examples, the system infers a first narrative comprising two semantic identities and a second narrative wherein the system infers that a user observing view didn't infer the second semantic identity and further doesn't use the second semantic identity in the second narrative. It further, uses the corresponding narrative to remind the user to carry an item and/or credential in order to start an activity.

Abstract: A multi faceted graphic user interface with multiple shells or layers may be provided for interaction with a user to speech enable interaction with applications and processes that do not necessarily have native support for speech input. The shells may be components of an operating system or of a parent application which supports such shells. Each shell has multiple facets for displaying applications and processes, and typically speech and other input is directed the application or process in the facet which has focus within the active shell. These multiple shells lend themselves to grouping of input or grouping of related applications and processes. For example, input from a speech recognizer, a mouse and a keyboard may each be directed at different shells; or a user may group related windows within various shells, such that all documents are displayed in one shell and all windows of an instant messaging application are displayed in another, thereby enabling better organization of work and work flow.

Abstract: A method of presenting product information in a vending machine may include detecting audio information from a consumer and converting the audio information to a text string. The method may include identifying a keyword in the text string, and determining products from a product database associated with the keyword. The method may include returning a list of the products that correspond to the keyword.

Abstract: Systems and processes for providing user-specific acoustic models are provided. In accordance with one example, a method includes, at an electronic device having one or more processors, receiving a plurality of speech inputs, each of the speech inputs associated with a same user of the electronic device; providing each of the plurality of speech inputs to a user-independent acoustic model, the user-independent acoustic model providing a plurality of speech results based on the plurality of speech inputs; initiating a user-specific acoustic model on the electronic device; and adjusting the user-specific acoustic model based on the plurality of speech inputs and the plurality of speech results.

Abstract: Techniques are disclosed relating to scoring partial matches between words. In certain embodiments, a method may include receiving a request to determine a similarity between an input text data and a stored text data. The method also includes determining, based on comparing one or more words included in the input text data with one or more words included in the stored text data, a set of word pairs and a set of unpaired words. Further, in response to determining that the set of unpaired words passes elimination criteria, the method includes calculating a base similarity score between the input text data and the stored text data based on the set of word pairs. The method also includes determining a scoring penalty based on the set of unpaired words and generating a final similarity score between the input text data and the stored text data by modifying the base similarity score based on the scoring penalty.

Abstract: A method for managing data includes obtaining, by a first data node, a notification, wherein the first data node is associated with a first power zone group (PZG), and in response to the notification: selecting a second data node, wherein the second data node is not associated with the first PZG, sending a data processing request to the second data node, obtaining a response based on the data processing request, wherein the response specifies a confirmation by the second data node to service the data processing request, storing a ledger entry in a ledger service that indicates the confirmation, and initiating a data transfer based on the data processing request, wherein the first data node is associated with the PZG based on a primary power source of the first data node.

Abstract: To provide more useful additional information to the user at a more effective timing. Provided is an information processing apparatus including an output control unit that controls an output of additional information related to a product purchased by a user. The output control unit controls an output of the additional information on the basis of delivery information about the product. In addition, provided is an information processing method that is executed by a processor, the information processing method including controlling an output of additional information related to a product purchased by a user. The controlling of an output includes controlling an output of the additional information on the basis of delivery information about the product.

Abstract: The present invention provides a pronunciation error detection apparatus capable of following a text without the need for a correct sentence even when erroneous recognition such as a reading error occurs.

Abstract: Some implementations are directed to selective invocation of a particular third-party (3P) agent by an automated assistant to achieve an intended action determined by the automated assistant during a dynamic dialog between the automated assistant and a user. In some of those implementations, the particular 3P agent is invoked with value(s) for parameter(s) that are determined during the dynamic dialog; and/or the particular 3P agent is selected, from a plurality of candidate 3P agents, for invocation based on the determined value(s) for the parameter(s) and/or based on other criteria. In some of those implementations, the automated assistant invokes the particular 3P agent by transmitting, to the particular 3P agent, a 3P invocation request that includes the determined value(s) for the parameter(s).

Abstract: Examples described herein provide for associating a network device to a network management system (NMS). Examples herein include determining, by a network orchestrator, a set of embeddings indicative of characteristics of the network device and each of a plurality of instances of the NMS. Examples herein include determining, by the network orchestrator for each of the plurality of instances, a probability score based on the set of embeddings, wherein the probability score is indicative of a likelihood of the network device to be associated with the instance. Examples herein further include, based on the probability score for each of the plurality of instances, selecting, by the network orchestrator, a first instance of the plurality of instances to associate with the network device. Examples herein include associating, by the network orchestrator, the network device to the first instance.

Abstract: An acoustic system and method is disclosed for providing spatial and temporal classification of a range of different types of sound producing targets in a geographical area. The system includes an optical signal transmitter arrangement for repeatedly transmitting, at multiple instants, interrogating optical signals into each of one or more optical fibres distributed across the geographical area and forming at least part of an installed fibre-optic communications network. An optical signal detector arrangement receives, during an observation period following each of the multiple instants, returning optical signals scattered in a distributed manner over distance along the one or more of optical fibres.

Abstract: The present invention allows for the capture and sentiment analysis of text the customer inputs into a chat, but never actually sends to the customer service representative (ghost text). The system captures this ghost text with a ghost capture system (GCS) software module. The GCS module analyzes the ghost text to generate metadata. The ghost text and metadata are used by a sentiment analysis engine to apply appropriate sentiment to the ghost text. The sentiment and ghost text are routed to a customer service representative (CSR). This provides the customer service agent with additional detail and information about a customer's emotions during a text chat conversation, allowing the CSR to determine a court of interaction not only based on the customer's response, but also based on the ghost text and the sentiment from the ghost text.

Abstract: A method includes receiving audio data encoding an utterance, processing, using a speech recognition model, the audio data to generate speech recognition scores for speech elements, and determining context scores for the speech elements based on context data indicating a context for the utterance. The method also includes executing, using the speech recognition scores and the context scores, a beam search decoding process to determine one or more candidate transcriptions for the utterance. The method also includes selecting a transcription for the utterance from the one or more candidate transcriptions.

Abstract: A programmable device such as a smart phone allows a user an opportunity to make final corrections to textual data in a message after the user has instructed the device to send the message, but before transmittal of the message. The opportunity is temporary, to avoid impeding the flow of communication, and the textual data is transmitted unmodified if the opportunity to modify it is not accepted. Modifications made during the opportunity period may be used to adapt an autocorrect functionality of the programmable device.

Abstract: Embodiments described herein are generally directed towards systems and methods relating to a crowd-sourced digital assistant and system. In particular, embodiments facilitate the intuitive creation and distribution of action datasets that include computing events or tasks that can be reproduced when an associated command, stored in an action dataset, is determined received by a digital assistant device. The digital assistant device described herein can generate new action datasets, on-board new action datasets, and receive new action datasets or updates to existing action datasets. Each digital assistant device in the described system can participate in the building of action datasets, so as to crowd-source a dialect that can be understood by a digital assistant device.

Abstract: A server for supporting speech recognition of a device and an operation method of the server. The server and method identify a plurality of estimated character strings from the first character string and obtain a second character string, based on the plurality of estimated character strings, and transmit the second character string to the device. The first character string is output from a speech signal input to the device, via speech recognition.

Abstract: Systems for optimized forecasting are provided. In some examples, data associated with strategy of one or more business units may be received. The strategy data may include identification of projects or goals. In some examples, industry trend data may be received and may include data associated with in-demand job skills and the like. An instruction to capture user data may be transmitted to one or more user devices of an employee user. The instruction may cause activation of one or more sensors or data capture devices. The captured user data may be received and analyzed to determine a competency of the user. Based on the strategy data, industry data and determined competency, one or more deficiencies between the resources needed to meet the business unit strategy data and the available resources may be identified. Based on the identified deficiency, one or more actions for execution may be identified and executed.

Abstract: Aspects of the subject disclosure may include, for example, applying first data associated with a first content item to a model to generate first classification characteristics, analyzing the first classification characteristics to generate a first marker, wherein the first marker delineates a first location of inventory within the first content item, selecting a first creative to populate a portion of the inventory, and populating, based on the selecting, the portion of the inventory with the first creative. Other embodiments are disclosed.

Abstract: Systems and processes for providing a virtual assistant service are provided. In accordance with one or more examples, a method includes receiving, from an accessory device communicatively coupled to the first electronic device, a representation of a speech input representing a user request. The method further includes detecting a second electronic device and transmitting, from the first electronic device, a representation of the user request and data associated with the detected second electronic device to a third electronic device. The method further includes receiving, from the third electronic device, a determination of whether a task is to be performed by the second electronic device in accordance with the user request; and in accordance with a determination that a task is to be performed by the second electronic device, requesting the second electronic device to performed the task in accordance with the user request.

Abstract: In some embodiments, recognition results produced by a speech processing system (which may include two or more recognition results, including a top recognition result and one or more alternative recognition results) based on an analysis of a speech input, are evaluated for indications of potential errors. In some embodiments, the indications of potential errors may include discrepancies between recognition results that are meaningful for a domain, such as medically-meaningful discrepancies. The evaluation of the recognition results may be carried out using any suitable criteria, including one or more criteria that differ from criteria used by an ASR system in determining the top recognition result and the alternative recognition results from the speech input. In some embodiments, a recognition result may additionally or alternatively be processed to determine whether the recognition result includes a word or phrase that is unlikely to appear in a domain to which speech input relates.

Abstract: Disclosed is a regional-features-based speech recognition method, including learning speech features by region using speech data classified by region category, and recognizing input speech using an acoustic model and a language model generated through classification of a region category for the input speech and the learning. A user may use a dialect recognition service that is improved using learning based on artificial intelligence (AI) and enhanced mobile broadband (eMBB), ultra-reliable and low latency communications (URLLC), and massive machine-type communications (mMTC) techniques of 5G mobile communication.

Abstract: A method and apparatus for determining a unique wake word for devices within an incident. One system includes an electronic computing device comprising a transceiver and an electronic processor communicatively coupled to the transceiver. The electronic processor is configured to receive a notification indicative of an occurrence of an incident and one or more communication devices present at the incident, determine contextual information associated with the incident and the one or more communication devices, and identify one or more wake words based on the contextual information. The electronic processor is further configured to determine a phonetic distance for each pair of wake words included in the one or more wake words, and select a unique wake word from the one or more wake words for each communication device of the one or more communication devices based on the determined phonetic distance.

Abstract: This application discloses a method and apparatus for processing audio information, a storage medium, and an electronic apparatus. The method includes: detecting that a segment of audio information is being received on a client, a first portion of audio information in the segment of audio information having been currently received on the client; obtaining first information, second information, and third information based on the first portion of audio information that has been currently received, the first information including text information corresponding to the first portion of audio information, the second information including information that meets a target condition and that corresponds to the first information, and the third information including information to be pushed to the client, which is obtained based on a keyword in the first information; and displaying the first information, the second information, and the third information on the client.

Abstract: In one aspect, a computerized method useful for implementing a language neutral virtual assistant including the step of providing a language detector. The language detector comprises one or more trained language classifiers. With language detector identifying a language of an incoming message from a user to an artificially intelligent (AI) personal assistant. The method includes the step of receiving an incoming message to the AI personal assistant. The method includes the step of normalizing the incoming message, wherein the normalizing the incoming message comprises a set of spelling corrections and a set of grammar corrections. The method includes the step of translating the incoming message to a specified language with a specified encoding process and a specified decoding process. The method includes the step of providing an AI personal assistant engine that comprise an artificial intelligence which conducts a conversation via auditory or textual methods.

Abstract: A robot capable of conversation with another robot and a method of controlling the same are disclosed. The robot includes a main body having a first region corresponding to a human face and rotatable in left-right direction directions, a signal generator generating a first data signal to be transmitted to a listener robot and a first robot voice signal corresponding to the first data signal, a communication unit transmitting the first data signal to an external server, a speaker outputting the first robot voice signal, and a controller controlling a rotation direction of the main body such that the first region is directed toward the listener robot at a time point adjacent to a transmission time of the first data signal and controlling the speaker to output the first robot voice signal after the rotation direction of the robot is controlled, wherein the listener robot receives the first data signal transmitted from the external server and is controlled to operate based on the first data signal.

Abstract: A method and system of reducing noise associated with telephony-based activities occurring in shared workspaces is provided. An end-user may lower their own voice to a whisper or other less audible or intelligible utterances and submit such low-quality audio signals to an automated speech recognition system via a microphone. The words identified by the automated speech recognition system are provided to a speech synthesizer, and a synthesized audio signal is created artificially that carries the content of the original human-produced utterances. The synthesized audio signal is significantly more audible and intelligible than the original audio signal. The method allows customer support agents to speak at barely audible levels yet be heard clearly by their customers.

Abstract: An electronic device is disclosed. The electronic device comprises: a voice input unit; a storage unit for storing a first text according to a first transcript format and at least one second text obtained by transcribing the first text in a second transcript format; and a processor for, when a voice text converted from a user voice input through the voice input unit corresponds to a preset instruction, executing a function according to the preset instruction. The processor executes a function according to a preset instruction when the preset instruction includes a first text and a voice text is a text in which the first text of the preset instruction has been transcribed into a second text of a second transcript format.

Abstract: Quantization matrix can be used to adjust quantization of transform coefficients at different frequencies. In one embodiment, a single fixed parametric model, such as a polynomial is used to represent a quantization matrix. Modulation of bit cost and complexity is achieved by specifying only the n first polynomial coefficients, the remaining ones being implicitly set to zero or other default values. One form of the single fixed polynomial is a fully developed polynomial in (x, y), where x, y indicate the coordinates of a given coefficient in a quantization matrix, with terms ordered by increasing exponent. Since higher exponents are the last ones, reducing the number of polynomial coefficients reduces the degree of the polynomial, hence its complexity. The polynomial coefficients can be symmetrical in x and y, and thus reducing the number of polynomial coefficients that need to be signaled in the bitstream.

Abstract: [Problem] Provided is a presentation support system that makes it possible to give effective presentations, for both presentations by machines and normal presenters. [Solution] The presentation support system included: a display unit 3; a material storage unit 5 that stores a presentation material and a plurality of keywords; an audio storage unit 7; an audio analysis unit 9 that analyzes a term contained in a presentation; a keyword order adjustment unit 11 that analyzes an order of appearance of a plurality of keywords contained in the audio analyzed by the audio analysis unit and changes the order of the plurality of keywords on the basis of the order of appearance; and a display control unit 13 that controls content displayed in the display unit 3.

Abstract: Systems and processes for operating an intelligent automated assistant are provided. In one example process, an event associated with an audio input is detected with a first process. In accordance with a detection of the event, a delay value associated with an electronic device is determined. The delay value corresponds to a time required to determine, with a second process, whether the audio input includes a spoken trigger. In accordance with a determination that the delay value exceeds a threshold, the delay value is broadcast during a first advertising session, and determination is made, during a second advertising session, whether the electronic device is to respond to the audio input. In accordance with a determination that the threshold is not exceeded, a determination is made, during the first advertising session, whether the electronic device is to respond to the audio input or wait for the second advertising session.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for contextual hotwords are disclosed. In one aspect, a method, during a boot process of a computing device, includes the actions of determining, by a computing device, a context associated with the computing device. The actions further include, based on the context associated with the computing device, determining a hotword. The actions further include, after determining the hotword, receiving audio data that corresponds to an utterance. The actions further include determining that the audio data includes the hotword. The actions further include, in response to determining that the audio data includes the hotword, performing an operation associated with the hotword.

Abstract: A method for generating an output for controlling a vehicular function of a vehicle includes providing (i) a vehicular sensing device having at least one illumination source operable to backlight a plurality of icons, each icon representative of a respective vehicle function, and (ii) a plurality of sensors, each sensor having a respective field of sensing associated with a respective icon of the plurality of icons. With the vehicular sensing device disposed at a vehicle, and with the at least one illumination source activated to backlight the plurality of icons, the backlit icons are viewable at an exterior portion of the vehicle, and the sensors sense movement of a person's hand or foot in a field of sensing of one of the sensors, and a controller generates an output to control the vehicular function that is represented by the respective backlit icon associated with that sensor.

Abstract: An electronic device is disclosed. The electronic device comprises: a voice input unit; a storage unit for storing a first text according to a first transcript format and at least one second text obtained by transcribing the first text in a second transcript format; and a processor for, when a voice text converted from a user voice input through the voice input unit corresponds to a preset instruction, executing a function according to the preset instruction. The processor executes a function according to a preset instruction when the preset instruction includes a first text and a voice text is a text in which the first text of the preset instruction has been transcribed into a second text of a second transcript format.

Abstract: The present disclosure provides a method for controlling a terminal, including the following operations: obtaining recognition results corresponding to control signals after receiving the control signals, and determining whether control instructions corresponding to the recognition results conflict, each control signal comprising at least one of a voice signal or a gesture signal; determining a credibility of each control instruction in response to a determination that there exists conflict among control instructions; and sending the control instruction with highest credibility to a control terminal. The present disclosure further provides a device for controlling a terminal and a computer readable storage medium. When control instructions are received and there exists conflict among control instructions, the control instruction with the highest credibility is sent to the control terminal after the credibility of each control instructions is determined, thereby avoiding settings from conflict.

Abstract: A device for changing a speech recognition sensitivity for speech recognition can include a memory and a processor configured to obtain a first plurality of speech data input at different times, apply a pre-trained speech recognition model to the first plurality of speech data at a plurality of different speech recognition sensitivities, obtain a first speech recognition sensitivity from among the plurality of different speech recognition sensitivities based on the pre-trained speech recognition model and the plurality of different speech recognition sensitivities, the first speech recognition sensitivity corresponding to an optimal speech recognition sensitivity at which a speech recognition success rate of the speech recognition model satisfies a set first recognition success rate criterion, and change a setting of the speech recognition sensitivity based on the first speech recognition sensitivity obtained from among the plurality of different speech recognition sensitivities.

Abstract: A speaker identification system (“system”) automatically assigns a speaker to voiced segments in a call. The system identifies one or more speakers in a call using one or more speaker-identification parameters. The system processes the call to determine one or more speaker-identification parameters, such as a transcript of the call, a facial image of the speaker, a scene image, which is an image of a scene in which the speaker is located during the call, or textual data associated with the call such as names of the speaker or an organization that are retrieved from the scene images or video data of the call. The system analyzes one or more of the speaker-identification parameters and determines the identity of the speaker. The system then identifies the voice segments associated with the identified speaker and marks the voice segments with the identity of the speaker.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for voice recognition. In one aspect, a method includes the actions of receiving a voice input; determining a transcription for the voice input, wherein determining the transcription for the voice input includes, for a plurality of segments of the voice input: obtaining a first candidate transcription for a first segment of the voice input; determining one or more contexts associated with the first candidate transcription; adjusting a respective weight for each of the one or more contexts; and determining a second candidate transcription for a second segment of the voice input based in part on the adjusted weights; and providing the transcription of the plurality of segments of the voice input for output.

Abstract: This disclosure describes techniques and architectures for evaluating conversations. In some instances, conversations with users, virtual assistants, and others may be analyzed to identify potential risks within a language model that is employed by the virtual assistants and other entities. The potential risks may be evaluated by administrators, users, systems, and others to identify potential issues with the language model that need to be addressed. This may allow the language model to be improved and enhance user experience with the virtual assistants and others that employ the language model.