Abstract: A method includes training a neural network having parameters on training data, in which the neural network receives an input state and processes the input state to generate a respective score for each decision in a set of decisions. The method includes receiving training data including training text sequences and, for each training text sequence, a corresponding gold decision sequence. The method includes training the neural network on the training data to determine trained values of parameters of the neural network. Training the neural network includes for each training text sequence: maintaining a beam of candidate decision sequences for the training text sequence, updating each candidate decision sequence by adding one decision at a time, determining that a gold candidate decision sequence matching a prefix of the gold decision sequence has dropped out of the beam, and in response, performing an iteration of gradient descent to optimize an objective function.

Abstract: Embodiments relate to generation of timed text in web video. In an embodiment, a computer-implemented method generates timed text for online video. In the method, a request to play a timed text track of a video incorporated into a web video service is received from a client computing device. Prior to receipt of the request, audio of the video is processed to determine intermediate timed text data. The intermediate timed text data lacks a complete text transcription of the audio, but includes data to enable the complete text transcription to be generated when playing the video. In response to receipt of the request, a text transcription of the audio is determined using the intermediate data with an automated speech-to-text algorithm. Finally, the text transcription of the audio is sent to the client computing device for display along with the video.

Abstract: The subject matter of this specification can be implemented in, among other things, a computer-implemented method including receiving audio data and a transcript of the audio data. The method further includes generating a language model including a factor automaton that includes automaton states and arcs, each of the automaton arcs corresponding to a language element from the transcript. The method further includes receiving language elements recognized from the received audio data and times at which each of the recognized language elements occur in the audio data. The method further includes comparing the recognized language elements to one or more of the language elements from the factor automaton to identify times at which the one or more of the language elements from the transcript occur in the audio data. The method further includes aligning a portion of the transcript with a portion of the audio data using the identified times.

Abstract: Embodiments relate to generation of timed text in web video. In an embodiment, a computer-implemented method generates timed text for online video. In the method, a request to play a timed text track of a video incorporated into a web video service is received from a client computing device. Prior to receipt of the request, audio of the video is processed to determine intermediate timed text data. The intermediate timed text data lacks a complete text transcription of the audio, but includes data to enable the complete text transcription to be generated when playing the video. In response to receipt of the request, a text transcription of the audio is determined using the intermediate data with an automated speech-to-text algorithm. Finally, the text transcription of the audio is sent to the client computing device for display along with the video.

Abstract: The subject matter of this specification can be implemented in, among other things, a computer-implemented method including receiving audio data and a transcript of the audio data. The method further includes generating a language model including a factor automaton that includes automaton states and arcs, each of the automaton arcs corresponding to a language element from the transcript. The method further includes receiving language elements recognized from the received audio data and times at which each of the recognized language elements occur in the audio data. The method further includes comparing the recognized language elements to one or more of the language elements from the factor automaton to identify times at which the one or more of the language elements from the transcript occur in the audio data. The method further includes aligning a portion of the transcript with a portion of the audio data using the identified times.