Abstract: Deep Neural Networks (DNNs) with many hidden layers and many units per layer are very flexible models with a very large number of parameters. As such, DNNs are challenging to optimize. To achieve real-time computation, embodiments disclosed herein enable fast DNN feature transformation via optimized memory bandwidth utilization. To optimize memory bandwidth utilization, a rate of accessing memory may be reduced based on a batch setting. A memory, corresponding to a selected given output neuron of a current layer of the DNN, may be updated with an incremental output value computed for the selected given output neuron as a function of input values of a selected few non-zero input neurons of a previous layer of the DNN in combination with weights between the selected few non-zero input neurons and the selected given output neuron, wherein a number of the selected few corresponds to the batch setting.

Abstract: In some aspects, a method of recognizing speech that comprises natural language and at least one word specified in at least one domain-specific vocabulary is provided. The method comprises performing a first speech processing pass comprising identifying, in the speech, a first portion including the natural language and a second portion including the at least one word specified in the at least one domain-specific vocabulary, and recognizing the first portion including the natural language. The method further comprises performing a second speech processing pass comprising recognizing the second portion including the at least one word specified in the at least one domain-specific vocabulary.

Abstract: Methods of incrementally modifying a word-level finite state transducer (FST) are described for adding and removing sentences. A prefix subset of states and arcs in the FST is determined that matches a prefix portion of the sentence. A suffix subset of states and arcs in the FST is determined that matches a suffix portion of the sentence. A new sentence can then be added to the FST by appending a new sequence of states and arcs to the FST corresponding to a remainder of the sentence between the prefix and suffix. An existing sentence can be removed from the FST by removing any arcs and states between the prefix subset and the suffix subset. The resulting modified FST is locally efficient but does not satisfy global optimization criteria such as minimization.

Abstract: Computing the feature Maximum Mutual Information (fMMI) method requires multiplication of vectors with a huge matrix. The huge matrix is subdivided into block sub-matrices. The sub-matrices are quantized into different values and compressed by replacing the quantized element values with 1 or 2 bit indices. Fast multiplication with those compressed matrices with far fewer multiply/accumulate operations compared to standard matrix computation is enabled and additionally obviates a de-compression method for decompressing the sub-matrices before use.

Abstract: Techniques are described for updating an automatic speech recognition (ASR) system that, prior to the update, is configured to perform ASR using a first finite-state transducer (FST) comprising a first set of paths representing recognizable speech sequences. A second FST may be accessed, comprising a second set of paths representing speech sequences to be recognized by the updated ASR system. By analyzing the second FST together with the first FST, a patch may be extracted and provided to the ASR system as an update, capable of being applied non-destructively to the first FST at the ASR system to cause the ASR system using the first FST with the patch to recognize speech using the second set of paths from the second FST. In some embodiments, the patch may be configured such that destructively applying the patch to the first FST creates a modified FST that is globally minimized.

Abstract: Techniques are described for updating an automatic speech recognition (ASR) system that, prior to the update, is configured to perform ASR using a first finite-state transducer (FST) comprising a first set of paths representing recognizable speech sequences. A second FST may be accessed, comprising a second set of paths representing speech sequences to be recognized by the updated ASR system. By analyzing the second FST together with the first FST, a patch may be extracted and provided to the ASR system as an update, capable of being applied non-destructively to the first FST at the ASR system to cause the ASR system using the first FST with the patch to recognize speech using the second set of paths from the second FST. In some embodiments, the patch may be configured such that destructively applying the patch to the first FST creates a modified FST that is globally minimized.

Abstract: Computing the feature Maximum Mutual Information (fMMI) method requires multiplication of vectors with a huge matrix. The huge matrix is subdivided into block sub-matrices. The sub-matrices are quantized into different values and compressed by replacing the quantized element values with 1 or 2 bit indices. Fast multiplication with those compressed matrices with far fewer multiply/accumulate operations compared to standard matrix computation is enabled and additionally obviates a de-compression method for decompressing the sub-matrices before use.

Abstract: Deep Neural Networks (DNNs) with many hidden layers and many units per layer are very flexible models with a very large number of parameters. As such, DNNs are challenging to optimize. To achieve real-time computation, embodiments disclosed herein enable fast DNN feature transformation via optimized memory bandwidth utilization. To optimize memory bandwidth utilization, a rate of accessing memory may be reduced based on a batch setting. A memory, corresponding to a selected given output neuron of a current layer of the DNN, may be updated with an incremental output value computed for the selected given output neuron as a function of input values of a selected few non-zero input neurons of a previous layer of the DNN in combination with weights between the selected few non-zero input neurons and the selected given output neuron, wherein a number of the selected few corresponds to the batch setting.

Abstract: Disclosed are systems, methods, and computer-readable media for performing translations from a source language to a target language. The method comprises receiving a source phrase, generating a target bag of words based on a global lexical selection of words that loosely couples the source words/phrases and target words/phrases, and reconstructing a target phrase or sentence by considering all permutations of words with a conditional probability greater than a threshold.

Abstract: A system, method and computer readable medium that provides an automated web transcription service is disclosed. The method may include receiving input speech from a user using a communications network, recognizing the received input speech, understanding the recognized speech, transcribing the understood speech to text, storing the transcribed text in a database, receiving a request via a web page to display the transcribed text, retrieving transcribed text from the database, and displaying the transcribed text to the requester using the web page.

Abstract: In some aspects, a method of recognizing speech that comprises natural language and at least one word specified in at least one domain-specific vocabulary is provided. The method comprises performing a first speech processing pass comprising identifying, in the speech, a first portion including the natural language and a second portion including the at least one word specified in the at least one domain-specific vocabulary, and recognizing the first portion including the natural language. The method further comprises performing a second speech processing pass comprising recognizing the second portion including the at least one word specified in the at least one domain-specific vocabulary.

Abstract: Disclosed are systems, methods, and computer-readable media for performing translations from a source language to a target language. The method comprises receiving a source phrase, generating a target bag of words based on a global lexical selection of words that loosely couples the source words/phrases and target words/phrases, and reconstructing a target phrase or sentence by considering all permutations of words with a conditional probability greater than a threshold.

Abstract: A system, method and computer readable medium that provides an automated web transcription service is disclosed. The method may include receiving input speech from a user using a communications network, recognizing the received input speech, understanding the recognized speech, transcribing the understood speech to text, storing the transcribed text in a database, receiving a request via a web page to display the transcribed text, retrieving transcribed text from the database, and displaying the transcribed text to the requester using the web page.

Abstract: Methods of incrementally modifying a word-level finite state transducer (FST) are described for adding and removing sentences. A prefix subset of states and arcs in the FST is determined that matches a prefix portion of the sentence. A suffix subset of states and arcs in the FST is determined that matches a suffix portion of the sentence. A new sentence can then be added to the FST by appending a new sequence of states and arcs to the FST corresponding to a remainder of the sentence between the prefix and suffix. An existing sentence can be removed from the FST by removing any arcs and states between the prefix subset and the suffix subset. The resulting modified FST is locally efficient but does not satisfy global optimization criteria such as minimization.

Abstract: Disclosed are systems, methods, and computer-readable media for performing translations from a source language to a target language. The method comprises receiving a source phrase, generating a target bag of words based on a global lexical selection of words that loosely couples the source words/phrases and target words/phrases, and reconstructing a target phrase or sentence by considering all permutations of words with a conditional probability greater than a threshold.

Abstract: A system, method and computer readable medium that provides an automated web transcription service is disclosed. The method may include receiving input speech from a user using a communications network, recognizing the received input speech, understanding the recognized speech, transcribing the understood speech to text, storing the transcribed text in a database, receiving a request via a web page to display the transcribed text, retrieving transcribed text from the database, and displaying the transcribed text to the requester using the web page.

Abstract: A system, method and computer readable medium that provides an automated web transcription service is disclosed. The method may include receiving input speech from a user using a communications network, recognizing the received input speech, understanding the recognized speech, transcribing the understood speech to text, storing the transcribed text in a database, receiving a request via a web page to display the transcribed text, retrieving transcribed text from the database, and displaying the transcribed text to the requester using the web page.

Abstract: A system, method and computer readable medium that provides an automated web transcription service is disclosed. The method may include receiving input speech from a user using a communications network, recognizing the received input speech, understanding the recognized speech, transcribing the understood speech to text, storing the transcribed text in a database, receiving a request via a web page to display the transcribed text, retrieving transcribed text from the database, and displaying the transcribed text to the requester using the web page.

Abstract: Classification of sequences, such as the translation of natural language sentences, is carried out using an independence assumption. The independence assumption is an assumption that the probability of a correct translation of a source sentence word into a particular target sentence word is independent of the translation of other words in the sentence. Although this assumption is not a correct one, a high level of word translation accuracy is nonetheless achieved. In particular, discriminative training is used to develop models for each target vocabulary word based on a set of features of the corresponding source word in training sentences, with at least one of those features relating to the context of the source word. Each model comprises a weight vector for the corresponding target vocabulary word.

Abstract: Classification of sequences, such as the translation of natural language sentences, is carried out using an independence assumption. The independence assumption is an assumption that the probability of a correct translation of a source sentence word into a particular target sentence word is independent of the translation of other words in the sentence. Although this assumption is not a correct one, a high level of word translation accuracy is nonetheless achieved. In particular, discriminative training is used to develop models for each target vocabulary word based on a set of features of the corresponding source word in training sentences, with at least one of those features relating to the context of the source word. Each model comprises a weight vector for the corresponding target vocabulary word.