Abstract: Techniques are provided for calculating reset parameters for recurrent neural networks (RNN). A methodology implementing the techniques according to an embodiment includes generating a sequence of statistics. The calculation of each statistic is based on outputs of an RNN that is periodically re-initialized at a selected RNN reset time such that each of the calculated statistics is associated with a unique RNN reset time selected from a pre-determined range of reset times. The method further includes analyzing the sequence to identify a maximum interval during which the sequence remains relatively constant. The method further includes selecting a reset time parameter and reset context duration parameter, for re-initialization of the RNN during operation. The reset time parameter is based on the duration of the identified maximum interval and the reset context duration parameter is based on a time associated with the starting point of the identified maximum interval.

Abstract: A computer-implemented method of speech recognition comprises forming a weighted finite state transducer (WFST) having nodes associated with states and interconnected by arcs, and to identify at least one word or word sequence hypothesis, identifying multiple sub-graphs on the WFST, each sub-graph having the same arrangement of multiple states and at least one arc, and propagating tokens in parallel through the sub-graphs, where each sub-graph is stored as a supertoken each having an array of tokens.

Abstract: Technologies for improved keyword spotting are disclosed. A compute device may capture speech data from a user of the compute device, and perform automatic speech recognition on the captured speech data. The automatic speech recognition algorithm is configured to both spot keywords as well as provide a full transcription of the captured speech data. The automatic speech recognition algorithm may preferentially match the keywords compared to similar words. The recognized keywords may be used to improve parsing of the transcribed speech data or to improve an assistive agent in holding a dialog with a user of the compute device.

Abstract: Techniques are provided for wake-on-voice (WOV) key-phrase enrollment. A methodology implementing the techniques according to an embodiment includes generating a WOV key-phrase model based on identification of the sequence of sub-phonetic units of a user-provided key-phrase. The WOV key-phrase model is employed by a WOV processor for detection of the user spoken key-phrase and triggering operation of an automatic speech recognition (ASR) processor in response to the detection. The method further includes updating an ASR language model based on the user-provided key-phrase. The update includes one of embedding the WOV key-phrase model into the ASR language model, converting sub-phonetic units of the WOV key-phrase model and embedding the converted WOV key-phrase model into the ASR language model, or generating an ASR key-phrase model by applying a phoneme-syllable based statistical language model to the user-provided key-phrase and embedding the generated ASR key-phrase model into the ASR language model.

Abstract: Techniques are provided for reducing the latency of automatic speech recognition using hypothesis score trend analysis. A methodology implementing the techniques according to an embodiment includes generating complete-phrase hypotheses and partial-phrase hypotheses, along with associated likelihood scores, based on a segment of speech. The method also includes selecting the complete-phrase hypothesis associated with the highest of the complete-phrase hypotheses likelihood scores, and selecting the partial-phrase hypothesis associated with the highest of the partial-phrase hypotheses likelihood scores. The method further includes calculating a relative likelihood score based on a ratio of the likelihood score associated with the selected complete-phrase hypothesis to the likelihood score associated with the selected partial-phrase hypothesis.

Abstract: Techniques are provided for calculating reset parameters for recurrent neural networks (RNN). A methodology implementing the techniques according to an embodiment includes generating a sequence of statistics. The calculation of each statistic is based on outputs of an RNN that is periodically re-initialized at a selected RNN reset time such that each of the calculated statistics is associated with a unique RNN reset time selected from a pre-determined range of reset times. The method further includes analyzing the sequence to identify a maximum interval during which the sequence remains relatively constant. The method further includes selecting a reset time parameter and reset context duration parameter, for re-initialization of the RNN during operation. The reset time parameter is based on the duration of the identified maximum interval and the reset context duration parameter is based on a time associated with the starting point of the identified maximum interval.

Abstract: Techniques are provided for rejecting out-of-domain (OD) queries in a language understanding system. A methodology implementing the techniques according to an embodiment includes generating a plurality of in-domain (ID) utterances based on variations of provided ID sentences, and generating a plurality of OD utterances based on variations of provided OD sentences. The method may further include training an ID language model based on the generated ID utterances and training an OD language model based on the generated OD utterances. The ID language model is configured to generate an ID dataset based on calculated probabilities associated with the generated ID utterances. The OD language model is configured to generate an OD dataset based on calculated probabilities associated with the generated OD utterances. The method further includes training a classifier to detect OD queries from a plurality of received queries, the training based on the ID dataset and the OD dataset.

Abstract: A method for context-aware query recognition in an electronic device includes receiving user speech from an input device. A speech signal is generated from the user speech. It is determined if the speech signal includes an action to be performed and if the electronic device is the intended recipient of the user speech. If the recognized speech signal include the action and the intended recipient of the user speech is the electronic device, a command is generated for the electronic device to perform the action.

Abstract: A method in a computing device for decoding a weighted finite state transducer (WFST) for automatic speech recognition is described. The method includes sorting a set of one or more WFST arcs based on their arc weight in ascending order. The method further includes iterating through each arc in the sorted set of arcs according to the ascending order until the score of the generated token corresponding to an arc exceeds a score threshold. The method further includes discarding any remaining arcs in the set of arcs that have yet to be considered.

Abstract: Techniques are provided for spoken language understanding based on keyword spotting and speech recognition. A methodology implementing the techniques according to an embodiment includes detecting a user spoken keyword or key-phrase embedded in an initial segment of a received audio signal, which is stored in a buffer. The method further includes triggering an automatic speech recognition (ASR) processor in response to the key-phrase detection. The method further includes performing automatic speech recognition, by the ASR processor, on a combination of the buffered initial segment and one or more additional received segments of the audio signal which include further speech from the user. The method still further includes performing natural language understanding on the recognized speech to determine a user request. The key-phrase is user selectable and serves to wake the ASR processor from a sleeping or idle lower power consumption state, into an active higher power consumption recognition state.

Abstract: Methods, apparatus, systems and articles of manufacture are disclosed for distributed automatic speech recognition. An example apparatus includes a detector to process an input audio signal and identify a portion of the input audio signal including a sound to be evaluated, the sound to be evaluated organized into a plurality of audio features representing the sound. The example apparatus includes a quantizer to process the audio features using a quantization process to reduce the audio features to generate a reduced set of audio features for transmission. The example apparatus includes a transmitter to transmit the reduced set of audio features over a low-energy communication channel for processing.

Abstract: Misspeaking is resolved in a natural language understanding system to interface with a machine. In one example, a user speech utterance is received. A sequence of classifiers is applied to words of the utterance to determine a meaning of the utterance. The meaning in interpreted as a command and the command is applied to a device for execution. The classifiers may include a first classifier to determine an intended function that is a subject of the utterance, a second classifier to determine words with properties that are related to the intended function, and a third classifier to select a property to apply to the function.

Abstract: A processing system includes a processor to execute a neural network application comprising an operation associated with a weight parameter and an input value, and an accelerator circuit, associated with the processor, to perform the operation, the accelerator circuit comprising a weight storage device to store a bit stream encoding the weight parameter, a controller to request a bit from the bit stream, an input data storage to store the input value, and an arithmetic logic unit (ALU) comprising an accumulator circuit to store an accumulation value and an operator circuit to receive the bit and the input value, receive a control signal from the controller, and responsive to determining that the control signal is set to a first value corresponding to a first operation and that that the bit encodes a first status, increase the accumulation value, stored in the accumulation circuit, by the input value.

Abstract: Technologies for improved keyword spotting are disclosed. A compute device may capture speech data from a user of the compute device, and perform automatic speech recognition on the captured speech data. The automatic speech recognition algorithm is configured to both spot keywords as well as provide a full transcription of the captured speech data. The automatic speech recognition algorithm may preferentially match the keywords compared to similar words. The recognized keywords may be used to improve parsing of the transcribed speech data or to improve an assistive agent in holding a dialog with a user of the compute device.

Abstract: A method for context-aware query recognition in an electronic device includes receiving user speech from an input device. A speech signal is generated from the user speech. It is determined if the speech signal includes an action to be performed and if the electronic device is the intended recipient of the user speech. If the recognized speech signal include the action and the intended recipient of the user speech is the electronic device, a command is generated for the electronic device to perform the action.

Abstract: A system, article, and method include techniques of automatic speech recognition using posterior confidence scores.

Abstract: Technologies for identifying sounds are disclosed. A sound identification device may capture sound data, and split the sound data into frames. The sound identification device may then determine an acoustic feature vector for each frame, and determine parameters based on how each acoustic feature varies over the duration of time corresponding to the frames. The sound identification device may then determine if the sound matches a pre-defined sound based on the parameters. In one embodiment, the sound identification device may be a baby monitor, and the pre-defined sound may be a baby crying.

Abstract: A method in a computing device for decoding a weighted finite state transducer (WFST) for automatic speech recognition is described. The method includes sorting a set of one or more WFST arcs based on their arc weight in ascending order. The method further includes iterating through each arc in the sorted set of arcs according to the ascending order until the score of the generated token corresponding to an arc exceeds a score threshold. The method further includes discarding any remaining arcs in the set of arcs that have yet to be considered.

Abstract: A system, article, and method of automatic speech recognition using parallel processing for weighted finite state transducer-based speech decoding.

Abstract: A system, article, and method of automatic speech recognition with dynamic vocabularies is described herein.