A score integration unit 7 obtains a new score Score (l.sub.1:n.sup.b, c) that integrates a score Score (l.sub.1:n.sup.b, c) and a score Score (w.sub.1:o.sup.b, c). This new score Score (l.sub.1:n.sup.b, c) becomes a score Score (l.sub.1:n.sup.b) in a hypothesis selection unit 8. Thus, the score Score (l.sub.1:n.sup.b) can be said to take into account the score Score (w.sub.1:o.sup.b, c). In a speech recognition apparatus, first information is extracted on the basis of the score Score (l.sub.1:n.sup.b) taking into account the score Score (w.sub.1:o.sup.b, c). Thus, speech recognition with higher performance than that in the related art can be achieved.

A score integration unit 7 obtains a new score Score (l.sub.1:n.sup.b, c) that integrates a score Score (l.sub.1:n.sup.b, c) and a score Score (w.sub.1:o.sup.b, c). This new score Score (l.sub.1:n.sup.b, c) becomes a score Score (l.sub.1:n.sup.b) in a hypothesis selection unit 8. Thus, the score Score (l.sub.1:n.sup.b) can be said to take into account the score Score (w.sub.1:o.sup.b, c). In a speech recognition apparatus, first information is extracted on the basis of the score Score (l.sub.1:n.sup.b) taking into account the score Score (w.sub.1:o.sup.b, c). Thus, speech recognition with higher performance than that in the related art can be achieved.

Disclosed are a method for training speech recognition model, a method and a system for speech recognition. The disclosure relates to field of speech recognition and includes: inputting an audio training sample into the acoustic encoder to represent acoustic features of the audio training sample in an encoded way and determine an acoustic encoded state vector; inputting a preset vocabulary into the language predictor to determine text prediction vector; inputting the text prediction vector into the text mapping layer to obtain a text output probability distribution; calculating a first loss function according to a target text sequence corresponding to the audio training sample and the text output probability distribution; inputting the text prediction vector and the acoustic encoded state vector into the joint network to calculate a second loss function, and performing iterative optimization according to the first loss function and the second loss function.

Disclosed are a method for training speech recognition model, a method and a system for speech recognition. The disclosure relates to field of speech recognition and includes: inputting an audio training sample into the acoustic encoder to represent acoustic features of the audio training sample in an encoded way and determine an acoustic encoded state vector; inputting a preset vocabulary into the language predictor to determine text prediction vector; inputting the text prediction vector into the text mapping layer to obtain a text output probability distribution; calculating a first loss function according to a target text sequence corresponding to the audio training sample and the text output probability distribution; inputting the text prediction vector and the acoustic encoded state vector into the joint network to calculate a second loss function, and performing iterative optimization according to the first loss function and the second loss function.

An audio stream is detected during a communication session with a user. Natural language processing on the audio stream is performed to update a set of attributes by supplementing the set of attributes based on attributes derived from the audio stream. A set of filter values is updated based on the updated set of attributes. The updated set of filter values is used to query a set of databases to obtain datasets. A probabilistic program is executed during the communication session by determining a set of probability parameters characterizing a probability of an anomaly occurring based on the datasets and the set of attributes. A determination is made if whether the probability satisfies a threshold. In response to a determination that the probability satisfies the threshold, a record is updated to identify the communication session to indicate that the threshold is satisfied.

An audio stream is detected during a communication session with a user. Natural language processing on the audio stream is performed to update a set of attributes by supplementing the set of attributes based on attributes derived from the audio stream. A set of filter values is updated based on the updated set of attributes. The updated set of filter values is used to query a set of databases to obtain datasets. A probabilistic program is executed during the communication session by determining a set of probability parameters characterizing a probability of an anomaly occurring based on the datasets and the set of attributes. A determination is made if whether the probability satisfies a threshold. In response to a determination that the probability satisfies the threshold, a record is updated to identify the communication session to indicate that the threshold is satisfied.

A method includes receiving a training example for a listen-attend-spell (LAS) decoder of a two-pass streaming neural network model and determining whether the training example corresponds to a supervised audio-text pair or an unpaired text sequence. When the training example corresponds to an unpaired text sequence, the method also includes determining a cross entropy loss based on a log probability associated with a context vector of the training example. The method also includes updating the LAS decoder and the context vector based on the determined cross entropy loss.

A method includes receiving a training example for a listen-attend-spell (LAS) decoder of a two-pass streaming neural network model and determining whether the training example corresponds to a supervised audio-text pair or an unpaired text sequence. When the training example corresponds to an unpaired text sequence, the method also includes determining a cross entropy loss based on a log probability associated with a context vector of the training example. The method also includes updating the LAS decoder and the context vector based on the determined cross entropy loss.

A method, performed by an electronic device, of providing a conversational service includes: receiving an utterance input; identifying a temporal expression representing a time in a text obtained from the utterance input; determining a time point related to the utterance input based on the temporal expression; selecting a database corresponding to the determined time point from among a plurality of databases storing information about a conversation history of a user using the conversational service; interpreting the text based on information about the conversation history of the user, the conversation history information being acquired from the selected database; generating a response message to the utterance input based on a result of the interpreting; and outputting the generated response message.

A method, performed by an electronic device, of providing a conversational service includes: receiving an utterance input; identifying a temporal expression representing a time in a text obtained from the utterance input; determining a time point related to the utterance input based on the temporal expression; selecting a database corresponding to the determined time point from among a plurality of databases storing information about a conversation history of a user using the conversational service; interpreting the text based on information about the conversation history of the user, the conversation history information being acquired from the selected database; generating a response message to the utterance input based on a result of the interpreting; and outputting the generated response message.