Systems and methods are disclosed for customizing a neural network for a custom dataset, when the neural network has been trained on data from a general dataset. The neural network may comprise an output layer including one or more nodes corresponding to candidate outputs. The values of the nodes in the output layer may correspond to a probability that the candidate output is the correct output for an input. The values of the nodes in the output layer may be adjusted for higher performance when the neural network is used to process data from a custom dataset.

Systems and methods are disclosed for end-to-end neural networks for speech recognition and classification and additional machine learning techniques that may be used in conjunction or separately. Some embodiments comprise multiple neural networks, directly connected to each other to form an end-to-end neural network. One embodiment comprises a convolutional network, a first fully-connected network, a recurrent network, a second fully-connected network, and an output network. Some embodiments are related to generating speech transcriptions, and some embodiments relate to classifying speech into a number of classifications.

Systems and methods are disclosed for generating internal state representations of a neural network during processing and using the internal state representations for classification or search. In some embodiments, the internal state representations are generated from the output activation functions of a subset of nodes of the neural network. The internal state representations may be used for classification by training a classification model using internal state representations and corresponding classifications. The internal state representations may be used for search, by producing a search feature from an search input and comparing the search feature with one or more feature representations to find the feature representation with the highest degree of similarity.

Systems and methods are disclosed for customizing a neural network for a custom dataset, when the neural network has been trained on data from a general dataset. The neural network may comprise an output layer including one or more nodes corresponding to candidate outputs. The values of the nodes in the output layer may correspond to a probability that the candidate output is the correct output for an input. The values of the nodes in the output layer may be adjusted for higher performance when the neural network is used to process data from a custom dataset.

A speech recognition method is provided. The method includes: obtaining a voice signal; processing the voice signal according to a speech recognition algorithm to obtain n candidate recognition results, the candidate recognition results including text information corresponding to the voice signal; identifying a target result from among the n candidate recognition results according to a selection rule selected from among m selection rules, the selection rule having an execution sequence of j, the target result being a candidate recognition result that has a highest matching degree with the voice signal in the n candidate recognition results, an initial value of j being 1; and identifying the target result from among the n candidate recognition results according to a selection rule having an execution sequence of j+1 based on the target result not being identified according to the selection rule having the execution sequence of j.

One embodiment of the present invention sets forth a technique for generating a summary of a recording. The technique includes generating an index associated with the recording, wherein the index identifies a set of terms included in the recording and, for each term in the set of terms, a corresponding location of the term in the recording. The technique also includes determining categories of predefined terms to be identified in the index and identifying a first subset of the terms in the index that match a first portion of the predefined terms in the categories. The technique further includes outputting a summary of the recording comprising the locations of the first subset of terms in the recording and listings of the first subset of terms under one or more corresponding categories.

A method and system for identifying the beginning and ending of songs via a machine learning analysis. A machine learning model analyzes streaming audio (such as a radio broadcast) in overlapping, 3-second samples. Each sample is labeled into groups such as song, talk, commercial and transition. Based on the location of the transition samples, an exact second a given song begins and ends in the audio stream is derivable. The model further identifies when two songs shift between one another.

A method of performing speech recognition using a two-pass deliberation architecture includes receiving a first-pass hypothesis and an encoded acoustic frame and encoding the first-pass hypothesis at a hypothesis encoder. The first-pass hypothesis is generated by a recurrent neural network (RNN) decoder model for the encoded acoustic frame. The method also includes generating, using a first attention mechanism attending to the encoded acoustic frame, a first context vector, and generating, using a second attention mechanism attending to the encoded first-pass hypothesis, a second context vector. The method also includes decoding the first context vector and the second context vector at a context vector decoder to form a second-pass hypothesis.

A novel method provides an AI language virtual agent having self-improvement features and which uses language modeling and tree search techniques. The AI language virtual agent exchanges textual discussion with users and other simulated agents. The method includes receiving a current situational description depicting natural language user input, temperament qualities and textual tendencies of the virtual agent, and indicia regarding subject matter context of a present conversation. The indicia regarding subject matter context include textual logs from recent conversational exchanges. The current situational description includes audio, visual, and tactile inputs collected proximate to the virtual agent. The method preferably utilizes an MCTS tree search in combination with self-moving modules, one or more language models, tree search techniques outputting textual responses to the current situation description, and the virtual agent responding with textual expression to verbal input in combination with the audio, visual, tactile, and other sensory inputs.

The present invention provides a wake-on-voice method and device. The method includes: obtaining a voice inputted by a user; processing data frames of the voice with a frame skipping strategy and performing a voice activity detection on the data frames by a time-domain energy algorithm; extracting an acoustic feature of the voice and performing a voice recognition on the acoustic feature according to a preset recognition network and an acoustic model; and performing an operation corresponding to the voice if the voice is a preset wake-up word in the preset recognition network.